{"query": "import torch from torch.distributions import constraints from torch.distributions.categorical import Categorical <del> from torch.distributions.utils import clamp_probs, broadcast_all, log_sum_exp </del> <ins> from torch.distributions.utils import clamp_probs, broadcast_all, _log_sum_exp </ins> from torch.distributions.distribution import Distribution from torch.distributions.transformed_distribution import TransformedDistribution from torch.distributions.transforms import ExpTransform", "pos": ["Leaving this issue here for higher visibility. Originally reported by here: Also mentioned by Piggybacking off of this issue, has strange behavior (reported here: )\nI have fixed this locally. Could I send a PR?\nYes, go ahead!"], "neg": []}
{"query": "begin{array}{ll} i_t = sigma(W_{ii} x_t + b_{ii} + W_{hi} h_{(t-1)} + b_{hi})  f_t = sigma(W_{if} x_t + b_{if} + W_{hf} h_{(t-1)} + b_{hf})  <del> g_t = tanh(W_{ig} x_t + b_{ig} + W_{hc} h_{(t-1)} + b_{hg})  </del> <ins> g_t = tanh(W_{ig} x_t + b_{ig} + W_{hg} h_{(t-1)} + b_{hg})  </ins> o_t = sigma(W_{io} x_t + b_{io} + W_{ho} h_{(t-1)} + b_{ho})  c_t = f_t c_{(t-1)} + i_t g_t  h_t = o_t tanh(c_t)", "pos": ["Hi, Looking at the formulas of LSTM in , in the third equation (gt), the suffix of the second W should be changed from W{hc} to W{hg}. The correct formula: gt=tanh(W{ig}xt+b{ig}+W{hg}h{(t\u22121)}+b_{hg}) Cheers, Navid"], "neg": []}
{"query": "tmp3 = torch.cuda.FloatTensor(t.size()) self.assertEqual(tmp3.data_ptr(), ptr[0], 'allocation not re-used') <ins> def test_noncontiguous_pinned_memory(self): # See issue #3266 x = torch.arange(0, 10).view((2, 5)) self.assertEqual(x.t(), x.t().pin_memory()) </ins> def test_caching_pinned_memory(self): cycles_per_ms = get_cycles_per_ms()", "pos": ["Hello, I was experimenting with a custom which receives a batch of different sized integer sequences and pads them to form 2D tensors. I noticed a weird issue for which I created the following MWE: It's probably because of the different memory ordering on Python's side and CUDA's allocator. Tested on K40 and GTX1080 with driver 375.39, CUDA8.\nPin memory is missing a call inside. Should be an easy fix\nfixed via"], "neg": []}
{"query": "- [Releases and Contributing](#releases-and-contributing) - [The Team](#the-team) <del> | Python |  **`Linux CPU`**   |  **`Linux GPU`** | |--------|--------------------|------------------| | 2.7.8  | [![Build Status](https://travis-ci.com/apaszke/pytorch.svg?token=shqHbUq29zKDxuqzGcjC&branch=master)](https://travis-ci.com/apaszke/pytorch) | | | 2.7    | [![Build Status](https://travis-ci.com/apaszke/pytorch.svg?token=shqHbUq29zKDxuqzGcjC&branch=master)](https://travis-ci.com/apaszke/pytorch) | [![Build Status](http://build.pytorch.org:8080/buildStatus/icon?job=pytorch-master-py2)](https://build.pytorch.org/job/pytorch-master-py2)  | | 3.5    | [![Build Status](https://travis-ci.com/apaszke/pytorch.svg?token=shqHbUq29zKDxuqzGcjC&branch=master)](https://travis-ci.com/apaszke/pytorch) | [![Build Status](http://build.pytorch.org:8080/buildStatus/icon?job=pytorch-master-py3)](https://build.pytorch.org/job/pytorch-master-py3)  | | Nightly| [![Build Status](https://travis-ci.com/apaszke/pytorch.svg?token=shqHbUq29zKDxuqzGcjC&branch=master)](https://travis-ci.com/apaszke/pytorch) | | </del> <ins> | System | Python | Status | | --- | --- | --- | | Linux CPU | 2.7.8, 2.7, 3.5, nightly | [![Build Status](https://travis-ci.org/pytorch/pytorch.svg?branch=master)](https://travis-ci.org/pytorch/pytorch) | | Linux GPU | 2.7 | [![Build Status](http://build.pytorch.org:8080/buildStatus/icon?job=pytorch-master-py2)](https://build.pytorch.org/job/pytorch-master-py2) | | Linux GPU | 3.5 | [![Build Status](http://build.pytorch.org:8080/buildStatus/icon?job=pytorch-master-py3)](https://build.pytorch.org/job/pytorch-master-py3) | </ins> ## More about PyTorch", "pos": ["We should explore using TorchDynamo to capture/optimize optimizer code. I don't anticipate any major issues here, but we should make sure it plays nice with AOT Autograd. cc\nlinking issue related to performance of this:"], "neg": []}
{"query": "]] [[ <del> name: free return: self arguments: - THTensor* self ]] [[ name: retain return: self arguments: - THTensor* self ]] [[ </del> python_name: index name: THPTensor_(getValue)<true> only_register: True", "pos": ["Dataloader call getitem as many times as indices in the current batch. In case datasets support a list of indices in one call, or a native python slice object, add a getbatch (optional) to pytorch dataset class. If dataloader sees the dataset has such a method implemented, make it fetch the batch corresping to the list of indices with one getbatch call rather than many getitem calls. This saves the trouble of calling getitem many times in cases where for example a disk file has to be opened then closed for every call and the underlying data structure supports a slicing or lists of indices for example hdf5 files.\nwill support something similar\nis the latest version of this."], "neg": []}
{"query": "class InstanceNorm2d(_InstanceNorm): r\"\"\"Applies Instance Normalization over a 4d input that is seen as a mini-batch of 3d inputs .. math:: y = frac{x - mean[x]}{ sqrt{Var[x]} + epsilon} * gamma + beta The mean and standard-deviation are calculated per-dimension separately for each object in a mini-batch. Gamma and beta are learnable parameter vectors of size C (where C is the input size).", "pos": ["why 1.21.6? because that's the oldest supported numpy on windows/python3.10: all cc\nHere is the test for pinning numpy to 1.21.6 The reason its was not pinned to this version is that conda can't properly resolve all the dependencies at this step here:\nIt looks like you are correct, I will put up a PR to pin lower then already pinned:\nHere is the PR:"], "neg": []}
{"query": "The upper triangular part of the matrix is defined as the elements on and above the diagonal. <del> The argument :attr:`k` controls which diagonal to consider. </del> <ins> The argument :attr:`diagonal` controls which diagonal to consider. </ins> <del> - :attr:`k` = 0, is the main diagonal. - :attr:`k` > 0, is above the main diagonal. - :attr:`k` < 0, is below the main diagonal. </del> <ins> - :attr:`diagonal` = 0, is the main diagonal. - :attr:`diagonal` > 0, is above the main diagonal. - :attr:`diagonal` < 0, is below the main diagonal. </ins> Args: input (Tensor): the input `Tensor` <del> k (int, optional): the diagonal to consider </del> <ins> diagonal (int, optional): the diagonal to consider </ins> out (Tensor, optional): The result `Tensor` Example::", "pos": ["The states that the argument controls which diagonal to consider. However, this argument is actually named in pytorch 0.2.0."], "neg": []}
{"query": "If :attr:`upper` is ``False``, then lower triangular portion is used. <ins> .. note:: The eigenvalues are returned in ascending order. If :attr:`input` is a batch of matrices, then the eigenvalues of each matrix in the batch is returned in ascending order. </ins> .. note:: Irrespective of the original strides, the returned matrix `V` will be transposed, i.e. with strides `V.contiguous().transpose(-1, -2).stride()`.", "pos": ["Documentation for does not indicate whether or not singular values are returned in descending order. My suspicion is that they are since both LAPACK and MAGMA's ?gesdd return them in descending order. It seems like this should be clarified in the docs to avoid unnecessary sorting, or mistaken assumptions. <!-- A clear and concise description of what content in is an issue. If this has to do with the general website, please file an issue at instead. If this has to do with , please file an issue at --"], "neg": []}
{"query": "embedding_dim (int): the size of each embedding vector max_norm (float, optional): If given, will renormalize the embeddings to always have a norm lesser than this norm_type (float, optional): The p of the p-norm to compute for the max_norm option <del> scale_grad_by_freq (boolean, optional): if given, this will scale gradients by the frequency of </del> <ins> scale_grad_by_freq (bool, optional): if given, this will scale gradients by the frequency of </ins> the words in the dictionary. mode (string, optional): 'sum' | 'mean'. Specifies the way to reduce the bag. Default: 'mean' <del> sparse (boolean, optional): if ``True``, gradient w.r.t. weight matrix will be a sparse tensor. See Notes for </del> <ins> sparse (bool, optional): if ``True``, gradient w.r.t. weight matrix will be a sparse tensor. See Notes for </ins> more details regarding sparse gradients. Attributes:", "pos": ["I have a straightforward use case. Given a token (or word), lookup the embedding of the word, pass that through a linear layer and perform L2 regression. If the token corresponds to the of the embedding layer, throws an error saying OS: macOS High Sierra PyTorch version: 0.3.1 How you installed PyTorch (conda, pip, source): Tried both conda and pip Python version: 3.6.4 CUDA/cuDNN version: CPU GPU models and configuration: NA GCC version (if compiling from source): NA A script to reproduce the bug:\nI got the same trouble, and it throw this error after some epochs of training. Does anyone have some solutions?"], "neg": []}
{"query": "g = tanh(W_{ig} x + b_{ig} + W_{hc} h + b_{hg})  o = mathrm{sigmoid}(W_{io} x + b_{io} + W_{ho} h + b_{ho})  c' = f * c + i * g  <del> h' = o * tanh(c_t)  </del> <ins> h' = o * tanh(c')  </ins> end{array} Args:", "pos": ["When converting a tensor to dlpack, the strides of single-value tensors () should be normalized to . This is the fallout conclusion from . Here is a demo of what is wrong: Collecting environment information... PyTorch version: 1.10.0a0+git2aa19f3 Is debug build: False CUDA used to build PyTorch: Could not collect ROCM used to build PyTorch: N/A OS: Ubuntu 20.04.4 LTS (x8664) GCC version: (GCC) 9.4.0 Clang version: 12.0.1 ( ) CMake version: version 3.21.1 Libc version: glibc-2.31 Python version: 3.9.6 (default, Jul 11 2021, 03:39:48) [GCC 9.3.0] (64-bit runtime) Python platform: Linux-5.4.0-122-generic-x8664-with-glibc2.31 Is CUDA available: False CUDA runtime version: 11.2.152 GPU models and configuration: GPU 0: NVIDIA GeForce GTX 1660 SUPER Nvidia driver version: 470.141.03 cuDNN version: /usr/lib/x8664-linux- HIP runtime version: N/A MIOpen runtime version: N/A Is XNNPACK available: False Versions of relevant libraries: [pip3] mypy==0.812 [pip3] mypy-extensions==0.4.3 [pip3] numpy==1.21.2 [pip3] torch==1.10.0a0+git2aa19f3 [pip3] torchaudio==0.10.0a0+ [conda] magma-cuda112 2.5.2 1 pytorch [conda] mkl 2021.3.0 h726a3e6557 conda-forge [conda] mkl-include 2021.3.0 h726a3e6557 conda-forge [conda] numpy 1.21.2 py39hdbf815f0 conda-forge [conda] torch 1.12.0a0+git1f29b31 dev0 <develop[conda] torchaudio 0.10.0a0+ pypi0 pypi\ndoes this capture the problem?"], "neg": []}
{"query": "def calculate_gain(nonlinearity, param=None): \"\"\"Return the recommended gain value for the given nonlinearity function. The values are as follows: ============ ========================================== nonlinearity gain ============ ==========================================", "pos": ["why 1.21.6? because that's the oldest supported numpy on windows/python3.10: all cc\nHere is the test for pinning numpy to 1.21.6 The reason its was not pinned to this version is that conda can't properly resolve all the dependencies at this step here:\nIt looks like you are correct, I will put up a PR to pin lower then already pinned:\nHere is the PR:"], "neg": []}
{"query": "def index_select(self, dim, index): return IndexSelect(dim)(self, index) <ins> def gather(self, dim, index): return Gather(dim)(self, index) def scatter(self, dim, index, source): return Scatter(dim)(self, index, source) def scatter_(self, dim, index, source): return Scatter(dim, True)(self, index, source) </ins> def masked_copy(self, mask, variable): return MaskedCopy()(self, mask, variable)", "pos": ["<!-- A clear and concise description of what content in is an issue. If this has to do with the general website, please file an issue at instead. If this has to do with , please file an issue at --In the getting started section section the following quote appears: I guess this is from previous versions where CUDA support was not built-in and this can be removed, correct? cc\nThis seems to be easy fixable docs issue\nIf gets merged this is fixed.\nThanks for the PR Just merged it."], "neg": []}
{"query": "# Example of using Container class Net(nn.Container): def __init__(self): <del> super(Net, self).__init__( conv1 = nn.Conv2d(1, 20, 5), relu  = nn.ReLU() ) </del> <ins> super(Net, self).__init__() self.conv1 = nn.Conv2d(1, 20, 5) self.relu  = nn.ReLU() </ins> def forward(self, input): output = self.relu(self.conv1(x)) return output", "pos": ["When I enable the MPS backend on Mac in libtorch, it now crashes with this error when exiting: If I remove: the crash disappears. It only affects exiting/destructing the app. Most likely it is the MPS memory allocator closing down. nightly 1st August 2022. Backtrace (basically it is torch::jit::Module destructor, leading to C10::StorageImpl destructor): cc\n, can you please provide a small test-case to repro the issue ?\nI'me using next C++ code, but can't reproduce this issue void toTest() { torch::Tensor inputTensor = torch::randn({1, 3, 256, 256}); (torch::kMPS); } Tested with the latest mps_graph branch build. Could you please provide a minimal reproducible test?\nThe crash is in torch::jit::Module so you'd likely need to be running a forward. I'll try to get a reproduce. It's not happening in Python which makes it a bit harder.\nI've tried to reproduce with a Module, for that i've created test module using code for resnet18 from this sample And next created next C++ code std::vector&lt;torch::jit::IValueinputs; inputs.pushback(torch::ones({1, 3, 224, 224}).to(torch::kMPS)); at::Tensor output = module.forward(inputs).toTensor().to(torch::kCPU); std::cout &lt;&lt; (1,0, 5) &lt;&lt; 'n'; With this code I see no issues, it runs forward and prints expected output: -1.2350 -0.0649 0.2181 0.0655 -0.2242 [ CPUFloatType{1,5} ]\n, removing the \"High priority\" label, as we are not able to reproduce the issue. Please provide us a failing case and we will take a look.\nWe weren't able to reproduce this issue, so I'll close it. Please reopen the issue if you see again. Thanks."], "neg": []}
{"query": "# (3) initialize mean square values and square gradient storage if not 'm' in state: <del> state['m'] = x.new().resize_as_(dfdx).fill_(1) </del> <ins> state['m'] = x.new().resize_as_(dfdx).zero_() </ins> state['tmp'] = x.new().resize_as_(dfdx)", "pos": ["We should unify this with our normal list creating code for the reasonable cases cc\nSimilar issue to this, I believe:", "The hub tests are broken on macOS. I'm seeing SSL certificate related errors when trying to hit GitHub. Run the macOS tests on CI with TestHub enabled. I tried installing the latest version of from conda-forge (which pulled in ) but this didn't fix the issue. cc\nLooks like the same error I'm getting trying to use Ubuntu 18.04 with fresh install of conda and Pytorch 1.2\nI think this will all be solved with .", "Porting TH operators is essential for code simplicity and performance reasons. Porting guides and Q&A are available in umbrella issue: Feel free to add as a reviewer to get a prioritized review.", "This is of particular use in list comprehensions: This should work:\nI'll need to adapt list comprehensions to use Iterables. I'll take a look!", "I haven't looked into this in depth but this looks fishy: More here: cc\nis this still an issue in 1.3.1?\nI can repro this on master:\nThe problem is: removes the parameters from the module's parameter list and makes them normal class attributes: applies only to parameters, not to the weight attributes.  (which is called from acts on which includes the value of the weight attributes and the bias parameters. So, it contains a mix of CUDA and CPU tensors.\nThis is my workaround for right now: This applies weight norm to an rnn module and overrides the function to deal with the error.\nThis appears to be fixed. do you want to verify?\nI tested this on nightly and it worked for me, so closing.", "As mentioned in title, I want to make sure if pytorch can deal with the situation that each input in a batch has different shapes and thus different computation graphs.\nPlease don't open duplicated issue for discussion. Context will be missed. Try to discuss about it here:", "Be Careful What You Backpropagate: A Case For Linear Output Activations & Gradient Boosting I can work on this if this can be to pytorch. Please let me know. Thanks!\nThis should be fairly straightforward to add in user code using . Eventually it could be to core, but its utility should probably be validated in external repos first."], "neg": []}
{"query": "Example:: >>> tensor = torch.tensor((), dtype=torch.float64) <del> >>> tensor.new_ones((2, 3)) tensor([[ 1.,  1.,  1.], [ 1.,  1.,  1.]], dtype=torch.float64) </del> <ins> >>> tensor.new_zeros((2, 3)) tensor([[ 0.,  0.,  0.], [ 0.,  0.,  0.]], dtype=torch.float64) </ins> \"\"\".format(**new_common_args))", "pos": ["The example given in the docs for newzeros() is actually an example for newones(). Should be a trivial fix. (Sorry, not quite sure what the etiquette/process is for fixing small errors like these). Thanks. The example should read:\nThanks for the catch, If you'd like to take a stab at fixing it, you can fork the pytorch repo and contribute a pull request. Otherwise, I can make the replacement.\nfixed via"], "neg": []}
{"query": "return obj; } <ins> PyObject * THSPTensor_(NewEmpty)() { THSTensorPtr tensor(THSTensor_(new)(LIBRARY_STATE_NOARGS)); THSTensor_(initStorage)(tensor.get()); return THSPTensor_(New)(tensor.release()); } </ins> static void THSPTensor_(dealloc)(THSPTensor* self) { if (self->cdata)", "pos": ["This causes to not work properly on newly created empty sparse tensors. In the example below, both should print 1 instead."], "neg": []}
{"query": "<ins> import os import sys import glob import platform import warnings from itertools import chain from .env import check_env_flag from .cuda import WITH_CUDA, CUDA_HOME def gather_paths(env_vars): return list(chain(*(os.getenv(v, '').split(':') for v in env_vars))) is_conda = 'conda' in sys.version or 'Continuum' in sys.version conda_dir = os.path.join(os.path.dirname(sys.executable), '..') WITH_NCCL = WITH_CUDA and platform.system() != 'Darwin' WITH_SYSTEM_NCCL = False NCCL_LIB_DIR = None NCCL_INCLUDE_DIR = None NCCL_ROOT_DIR = None if WITH_CUDA and not check_env_flag('NO_SYSTEM_NCCL'): ENV_ROOT = os.getenv('NCCL_ROOT_DIR', None) # NCCL_ROOT_DIR takes precedence over NCCL_LIB_DIR lib_paths = list(filter(bool, [ os.path.join(ENV_ROOT, 'lib') if ENV_ROOT is not None else None, os.path.join(ENV_ROOT, 'lib64') if ENV_ROOT is not None else None, os.getenv('NCCL_LIB_DIR'), os.path.join(CUDA_HOME, 'lib'), os.path.join(CUDA_HOME, 'lib64'), '/usr/lib/x86_64-linux-gnu/', '/usr/lib/powerpc64le-linux-gnu/', '/usr/lib/aarch64-linux-gnu/', ] + gather_paths([ 'LIBRARY_PATH', ]) + gather_paths([ 'LD_LIBRARY_PATH', ]))) if os.getenv('NCCL_INCLUDE_DIR') is not None: warnings.warn(\"Ignoring environment variable NCCL_INCLUDE_DIR because \" \"NCCL_INCLUDE_DIR is implicitly assumed as \" \"$NCCL_ROOT_DIR/include or $NCCL_LIB_DIR/../include\") if is_conda: lib_paths.append(os.path.join(conda_dir, 'lib')) for path in lib_paths: if path is None or not os.path.exists(path): continue if glob.glob(os.path.join(path, 'libnccl*')): if os.path.exists((os.path.join(path, '../include/nccl.h'))): NCCL_LIB_DIR = path NCCL_INCLUDE_DIR = path break if NCCL_LIB_DIR is not None: WITH_SYSTEM_NCCL = True NCCL_ROOT_DIR = os.path.abspath(os.path.join(NCCL_LIB_DIR, \"../\")) </ins>", "pos": ["Currently, you can have a libnccl on your system, and pytorch setup favors that instead of compiling one itself from source. The code that takes care of this is: However, this breaks gloo build because gloo gets a path to NCCL via:\nI got the error during the installation: Is it related to this problem?\nthat's a dupe of\nno, that's a consequence of , now you have to do before compiling. I think you don't want to do recursive submodule update, because that will pull in gtest and another copy of nccl.\nAlternatively we can pass new flags to gloo to prevent it from using its own NCCL\nThanks works for me.\nYou can prevent it from using its own nccl, but its cmake still has to be able to find some nccl.\nI think should be sufficient. I'm testing.\nfixes it when we dont have an external NCCL and pytorch compiles it's own NCCL?\nI don't see how NCCLEXTERNAL fixes anything, looks like in gloo NCCLEXTERNAL is used just to set dependency if gloo is compiling its own nccl (which we don't want in any case).\nThe only thing I tested when I was that the vendored nccl was not being built (even if its submodule was checked out); I tested this by adding an to one of the headers. I'll test with the other configurations.\nActually, there are a few bugs here. First, even though GLOOFLAGS looks like it's setting NCCLROOTDIR: it actually isn't because the gloo invocation is which expands to but only is passed to cmake. Suppose you fix this. nccl is still not going to be picked up, because INSTALLDIR is PyTorch's local install dir, but if you disabled nccl builds, there will never actually be a nccl here. So cmake is just going to do it's standard search to find a nccl somewhere. Inversely, if you ever built nccl, and then turn it off, we will keep picking up that locally built version until you delete it. So we shouldn't do that. Suppose you fix this.", "Then I think gloo will (continue to) correctly pick up system nccl and be built with it, and the final thing to fix is that build of PyTorch csrc doesn't know how to search for nccl headers somewhere. This is probably related to . I wonder, maybe we shouldn't do any autodetection at all, and force the user to pass us the include and library paths of nccl if they want to build with system.\nFYI, I committed a change to gloo to stop bundling NCCL itself. It will now happily pick up whatever NCCL you give it through NCCLROOTDIR or by trying to find it in the standard systems paths. See facebookincubator/gloo You do have to set to make Gloo try and find it and compile in support.\nSo to fix this, you should be able to not specify anything aside from if you want to use system NCCL, and add if you have some custom location (either a PyTorch bundled version or a custom install altogether)."], "neg": []}
{"query": "| :attr:`stride` controls the stride for the cross-correlation. | If :attr:`padding` is non-zero, then the input is implicitly zero-padded on both sides <del> for :attr:`padding` number of points | :attr:`dilation` controls the spacing between the kernel points. It is harder to describe, but this `link`_ has a nice visualization of what :attr:`dilation` does. | :attr:`groups` controls the connections between inputs and outputs. </del> <ins> for :attr:`padding` number of points. | :attr:`dilation` controls the spacing between the kernel points; also known as the \u00e0 trous algorithm. It is harder to describe, but this `link`_ has a nice visualization of what :attr:`dilation` does. | :attr:`groups` controls the connections between inputs and outputs. `in_channels` and `out_channels` must both be divisible by `groups`. </ins> |       At groups=1, all inputs are convolved to all outputs. |       At groups=2, the operation becomes equivalent to having two conv layers side by side, each seeing half the input channels, and producing half the output channels, and both subsequently concatenated. <ins> At groups=`in_channels`, each input channel is convolved with its own set of filters (of size `out_channels // in_channels`). </ins> .. note::", "pos": ["Hi, I have installed two versions of CUDA. But I want to use one of it, can I specify it during setup process?\nTry running\nI have tried it, () still returns False. Thx for your reply!\nDo you have the cuda driver installed?\nOf course I do. It raise assertion error which shows \"The NVIDIA driver on your system is too old(found version 8000)\".\nthat's the answer to your problem. upgrade your driver.\nThx!", "When building from source, I got this error. My system is OSX 10.12.4, conda, CUDA 8.0. Any thoughts? Thanks!\ncan you try adding to the file and see if that fixes things? if so i'll commit it.\nactually, that file already has memory included. Are you following instructions from Especially: This part is important:\nworks nicely! Thanks a lot!", "The suggest that takes a as a parameter, but it actually must be wrapped in a Variable first.\nyou are looking at the Tensor docs. When operating on Variable, it is implied from the tutorials that all functions operating on Variables have to use Variables. We didn't write separate docs for Variable itself.", "something wrong with import torchvision Traceback (most recent call last): File \"\", line 3, in <moduleimport torchvision File \"\", line 4, in <moduleimport torchvision.transforms as transforms ImportError: No module named transforms\nyou have a file called for some reason. delete it or change your pythonpath to ignore it.\nHey, I have the same problem persisting when I import torchvision. I have no other file by the same name. ImportError Traceback (most recent call last) <ipython-input-6-3baf6837d8f1in <module() 8 from import Dataset 9 from import DataLoader ---10 from torchvision import transforms ImportError: No module named torchvision Any idea why this is happening?\nDo you have torchvision installed (it is a separate package)?\npip Install torchvision try this\nfrom data import BaseTransform,VOC_CLASSES as labelmap this is not working\npip install torchvision", "Hi, I need to pass input through one nn.Module, then argmax the output, and then pass that output to a second nn.Module. I want to backprop through the argmax back to the weights of the first module. That's impossible right now, though, it seems. When I loop through the gradient of the weights of my first module (with the following code): firstmodel.zerograd() secondmodel.zerograd() loss.backward() # retainvariables=True for param in secondmodel.parameters(): print( 'data: ', ) print( 'grad data: ', ) for param in firstmodel.parameters(): print( 'data: ', ) print( 'grad data: ', ) I get the following error, stating that those gradients can no longer be viewed/exist: -main()\nargmax is not differentiable (and even if you try to think about it as a function differentiable almost everywhere, its derivative is always 0, so you won't get anything meaningful).", "def inithidden(self): h0 = Variable(FloatTensor(self.numlayers, self.batchlen, self.hiddendim).zero(),  requiresgrad=True) c0 = Variable(FloatTensor(self.numlayers, self.batchlen, self.hiddendim).zero(),  requiresgrad=True) return (h0, c0) def forward(self, sequences): lstmout, = ((1, self.batchlen, self.inputdim), ) ... = nn.LSTM(inputdim, hiddendim, numlayers, dropout = dropout, batchfirst=True) = self.inithidden() output = model(traininput) produces the following error although I defined weight and input dimensions as in the documentation: File \"/usr/local/lib/python2.7/dist-\", line 202, in call result = self.forward(input, kwargs) File \"/usr/local/lib/python2.7/dist-\", line 91, in forward output, hidden = func(input, self.allweights, hx) File \"/usr/local/lib/python2.7/dist-\", line 327, in forward return func(input, fargs, fkwargs) File \"/usr/local/lib/python2.7/dist-\", line 227, in forward nexth, output = func(input, hidden, weight) File \"/usr/local/lib/python2.7/dist-\", line 67, in forward hy, output = inner(input, hidden[l], weight[l]) File \"/usr/local/lib/python2.7/dist-\", line 96, in forward hidden = inner(input[i], hidden, *weight) File \"/usr/local/lib/python2.7/dist-\", line 22, in LSTMCell gates = F.linear(input, wih, bih) + F.linear(hx, whh, bhh) File \"/usr/local/lib/python2.7/dist-\", line 744, in add return (other) File \"/usr/local/lib/python2.7/dist-\", line 293, in add return (other, False) File \"/usr/local/lib/python2.7/dist-\", line 287, in add return Add(inplace)(self, other) File \"/usr/local/lib/python2.7/dist-\", line 20, in forward return a.add(b) RuntimeError: inconsistent tensor size at /data/users/soumith/builder/wheel/pytorch-src/torch/lib/TH/generic/THTensorMath.c:827\nI can't reproduce your problem. Could you include a short runnable example that exhibits this behavior?", "You also don't need to manually init hidden anymore, the following would work as well...\nThanks! After removing problem solved.", "Hi, I would like to implement a sharing weight 'alexnet' with two input batches of images, and the fully connected layers are concatenated together like this: However, when I run it, I come across the following error: Does anyone know about it? Thank you very much.\nYou forgot to call on the output of . In the future please remember that we're using the issues for bug reports and feature requests only. If you have questions or need help, please post ."], "neg": []}
{"query": "for module in (nn.RNN, nn.LSTM, nn.GRU): for bias in (True, False): for bidirectional in (False, True): <del> for dropout in (0, 1):  # Because of dropout randomness, can only compare 0 and 1 for batch_first in (False, True): num_directions = 2 if bidirectional else 1 if batch_first: input_val = torch.randn(batch, seq_length, input_size) else: input_val = torch.randn(seq_length, batch, input_size) hx_val = torch.randn(num_layers * num_directions, batch, hidden_size) rnn = module(input_size, </del> <ins> for batch_first in (False, True): num_directions = 2 if bidirectional else 1 if batch_first: input_val = torch.randn(batch, seq_length, input_size) else: input_val = torch.randn(seq_length, batch, input_size) hx_val = torch.randn(num_layers * num_directions, batch, hidden_size) rnn = module(input_size, hidden_size, num_layers, bias=bias, dropout=dropout, bidirectional=bidirectional, batch_first=batch_first) outputs_cpu = forward_backward( False, rnn, input_val, hx_val, rnn.all_weights) rnn_gpu = module(input_size, </ins> hidden_size, num_layers, bias=bias,", "pos": ["cuDNN backend for RNN stabilizes at large mem usage. The following code stabilizes for me at 14+GiB reported by nvidia-smi. The same RNN without cuDNN uses <2.5GiB.\nOops... any idea what's going on with the garbage collector in this case?\nare you running this on a P100 GPU?\nYes, I observe similar behavior on Titan X Pascal\nhmmm, do cudnn RNNs have weird memory-oversubscribing properties? If it wasn't for CuDNN RNN, then it's not unreasonable. we use a custom CUDA allocator that will oversubscribe if needed, but will free the memory blocks it's holding onto if it hits OOM / failed allocations.\nTried the same RNN in torch and am getting ~2.6GiB on nvidia-smi.\nWith cuDNN I'm getting a constant usage of 2241MB on my machine using PyTorch (M40).\nare you using cudnnv5, or v6?\nv5\nno, cudnn does not do memory oversubscribing, and already tried the same network in lua torch with cudnn, where it gets expected memory usage. Also, he is getting OOMs on slightly bigger networks, where apparently caching allocator should have tried to free previous blocks. Interesting that can not repro this.\nJust to make sure did you set when you ran? I left it as False in the repro.\nI can repro in the docker image built from the current version on P100, and I go OOM after 8 batches with the same image on 1080, which generally should not happen - if there is enough memory to go through 1 batch, memory usage should not increase afterwards.\nNo, I just copied your script and ran it. Weird thing is that it goes OOM after 18 batches once I enable . Without it memory remains stable.\nA lot of cudaMalloc allocations go not through THCudaMalloc, and never get freed:\nSo not going through THCudaMalloc is actually expected for storages allocated with THCStorage_(newWithSize), but what's not expected is to every time cudaMalloc bytes of rnn workspace, instead of reusing earlier allocations. (~784 MB allocations are for workspace, ~156MB allocations happen somewhere in criterion).\nMy guess is that the bug is at the Python level.", "Either: Something is holding on to a Tensor longer than it should Something is creating a circular reference which prevents quick clean-up\nThe issue is that in creates cyclic references. It's made worse by the fact that it holds on to more memory than it should (like the workspace). You can see this by adding a as the first line in your for loop. With the explicit call to the GC every iteration, the memory usage doesn't explode.\nhow can these cyclic references be fixed?\nNot sure yet, but I will fix it tomorrow with"], "neg": []}
{"query": "return type(self), (self.tolist(),) def __repr__(self): <del> return repr(str(self)) </del> <ins> return str(self) </ins> def __str__(self): # All strings are unicode in Python 3, while we have to encode unicode", "pos": ["In torch, printing a tensor will print the whole tensor, even if that one is huge. This is sometimes annoying, especially when using a notebook, as it crashes it. Numpy has a very nice way to display big tensors by truncating them. Would it be possible so have something similar for pytorch?\nYeah, we haven't implemented it yet, but that's definitely on our roadmap\nImplemented in .\nI should say that I have some artefacts that appear with while printing a tensor. Here is an example output: gives Curiously, if I simply do It works fine and print nicely, with the truncation for large tensors. I'm using Python 2.7.6 (probably the one that came with my system and not from anaconda). Maybe it's just a bad setup in my machine?\nhmmm, it works fine on python 2.7.12. wonder what changed between .6 and .12\nok i got a repro. I opened my ipython and did this:\nmy example was flawed, sorry. Using works for me, but when only entering it prints the s, as in your example.\nIs it possible to tell the truncation point?\nyes, use the function"], "neg": []}
{"query": "] if WITH_CUDNN: main_libraries += ['cudnn'] <del> library_dirs.append(CUDNN_LIB_DIR) </del> <ins> library_dirs.insert(0, CUDNN_LIB_DIR) </ins> # NOTE: these are at the front, in case there's another cuDNN in CUDA path include_dirs.insert(0, CUDNN_INCLUDE_DIR) if not IS_WINDOWS:", "pos": ["If there was two libcudnn files ( and ) in the folder wouldn't it result in taking the older version ()? Maybe adding a warning or something to insure its the right lib?\nI think and are quite possibly related. EDIT: also", "Compilation goes without a problem but importing torch gives ImportError: /lib/python3.6/site-packages/torch/C.cpython-36m-x8664-linux- undefined symbol: cudnnSetConvolutionGroupCount Compiling from source v0.2.0 works fine. Set up is Python 3.6.2 :: Anaconda custom (64-bit) Cuda compilation tools, release 8.0, V8.0.61 gcc (GCC) 4.8.5 centos-release-scl-rh-2-\nIt's likely a configuration issue of your system. You built with cuDNN7, but cuDNN6 is loaded at runtime. Can you please run and paste the output in a gist?\nOK, here's the .\nThat's weird. It doesn't even seem to be linked. Can you post the output of ? Did you build from source? Can you clean the repo ( ASSUMING YOU DID NO CHANGES YOURSELF - THIS MIGHT DESTROY THEM), build form scratch, and post the build log in a gist as well?\nHey Adam, I made a mistake. This is the correct . And the ldd just in case.\nOk. cuDNN isn't linked. How did you install PyTorch?\nCompiled from source, following the instructions.\nPlease rebuild and post the build log\nDone, here's the .\nThe build log looks correct. I think the problem is that you have multiple Python environments with different PyTorch packages. In your run time log, torch is loaded from , but in your build log it's installed in . Cleaning up the global install should fix it for you.\nOk, so I made sure pytorch wasn't installed in the global python. Checked & removed from the Anaconda . Now the only shows Still, get Again, runtime and the of the from the error\nEh, I'm running out of tricks. Maybe there's sth wrong with the compiler? you can see that there's a flag when building . Try prepending before the build command\nThe sysadmin installed both cuDNN 6.0 and 7.0 and that's causing the problem. I followed but still found it necessary to insert CUDNNLIBDIR at the beginning of the list with .", "Now, after compiling the import doesn't throw any errors, so I'm guessing it's ok.\nhmm yeah, it should also go in the front of . Can you please submit a PR with the fix?\nWill do\ni believe this patch in the second last comment has to be adapted for your refactor."], "neg": []}
{"query": "return output @staticmethod <del> @once_differentiable </del> def backward(ctx, grad_output, grad_indices=None): <del> grad_input = grad_output.new(ctx.input_size).zero_() </del> <ins> grad_input = Variable(grad_output.data.new(ctx.input_size).zero_(), requires_grad=True) </ins> if ctx.return_indices: indices, = ctx.saved_tensors else: indices = ctx.indices dim = ctx.dim if ctx.dim is not None else grad_output.dim() - 1 <del> return (grad_input.scatter_(dim, indices, grad_output),) + (None,) * ctx.num_flags </del> <ins> return (grad_input.scatter(dim, Variable(indices), grad_output),) + (None,) * ctx.num_flags </ins> class Sort(_MultiSelectionFunction):", "pos": ["Just a very small grammatical issue. ! Suggested fix: \"mutates\" to \"mutate\" <!-- A clear and concise description of what content in is an issue. If this has to do with the general website, please file an issue at instead. If this has to do with , please file an issue at --\nThanks for pointing this out. Please feel free to open a pull request for this.\nThanks for adding labels. Forgot to do that. I will make a pull request"], "neg": []}
{"query": "#  binary_linux_libtorch_3.6m_cpu_test: #    environment: #      BUILD_ENVIRONMENT: \"libtorch 3.6m cpu\" <del> #    resource_class: gpu.medium </del> <ins> #    resource_class: gpu.nvidia.small </ins> #    <<: *binary_linux_test # #  binary_linux_libtorch_3.6m_cu90_test: #    environment: #      BUILD_ENVIRONMENT: \"libtorch 3.6m cu90\" <del> #    resource_class: gpu.medium </del> <ins> #    resource_class: gpu.nvidia.small </ins> #    <<: *binary_linux_test #", "pos": ["As it still using deprecated resource class, see See example of the failure in\nFollowup: we need to plan better --we knew it was gonna happen and none of us fixed it early enough. Action item: go through tasks about deprecation in sprint planning"], "neg": []}
{"query": "for weight in self.parameters(): weight.data.uniform_(-stdv, stdv) <ins> def check_forward_args(self, input, hidden, batch_sizes): is_input_packed = batch_sizes is not None expected_input_dim = 2 if is_input_packed else 3 if input.dim() != expected_input_dim: raise RuntimeError( 'input must have {} dimensions, got {}'.format( expected_input_dim, input.dim())) if self.input_size != input.size(-1): raise RuntimeError( 'input.size(-1) must be equal to input_size. Expected {}, got {}'.format( fn.input_size, input.size(-1))) if is_input_packed: mini_batch = batch_sizes[0] else: mini_batch = input.size(0) if self.batch_first else input.size(1) num_directions = 2 if self.bidirectional else 1 expected_hidden_size = (self.num_layers * num_directions, mini_batch, self.hidden_size) def check_hidden_size(hx, expected_hidden_size, msg='Expected hidden size {}, got {}'): if tuple(hx.size()) != expected_hidden_size: raise RuntimeError(msg.format(expected_hidden_size, tuple(hx.size()))) if self.mode == 'LSTM': check_hidden_size(hidden[0], expected_hidden_size, 'Expected hidden[0] size {}, got {}') check_hidden_size(hidden[1], expected_hidden_size, 'Expected hidden[1] size {}, got {}') else: check_hidden_size(hidden, expected_hidden_size) </ins> def forward(self, input, hx=None): is_packed = isinstance(input, PackedSequence) if is_packed:", "pos": ["I run RNN sucessfully with the code. and I run embedding example. Then, got RuntimeError. I set to use nothing seems wrong. And the error info seems not the dim problem? The error is so strange.\nWhat exactly do you mean by \"embedding example\"? Could you provide a code snippet that causes the RuntimeError?\nTo better reproduce the error, I use the official code example of The origin runs well: Due to use , I have to make batch first. And then , it raises error.\nThank you for the example! The say that setting means that the first dim of input is the batch, ie, input has size . however is unchanged, so the following should work: I do agree the error message could be better, though"], "neg": []}
{"query": "} __device__ __forceinline__ void operator()(T* out, T* in) { <del> *out = mask(in) ? *in : ScalarConvert<int, T>::to(0); </del> <ins> *out = mask(out) ? *in : ScalarConvert<int, T>::to(0); </ins> } __device__ __forceinline__ void operator()(T* v) {", "pos": ["On recent PyTorch master (built from source; 0.4.0a0+) on a P100, CUDA 9:\nI think the results on CPU meets the defination of triu (upper triangular part of the matrix). It is really hard to handle with a tensor from expand operation with multi-thread not only on GPU but also on CPU. The unexpected result will also happen on CPU if we apply openmp to it. I think it is necessary to eleminate the difference between the implementation and defination.\nIt doesn\u2019t matter if it\u2019s cpu or not. triu doesn\u2019t have an underscore which means it doesn\u2019t operate in-place and should always give the same result (in new memory)"], "neg": []}
{"query": "} template <typename Dtype, typename Acctype> <ins> __launch_bounds__(CUDA_NUM_THREADS) </ins> __global__ void col2im_kernel(const int n, const Dtype* data_col, const int height, const int width, const int channels, const int kernel_h, const int kernel_w,", "pos": ["I'm trying to run a variant of ERFNet on an NVIDIA TX-2 running Jetpack 3.2 (CUDA 9.0 and CuDNN 7). I get the following error: RuntimeError: cuda runtime error (7) : too many resources requested for launch at ../../../pytorch/torch/lib/THCUNN/im2col.h:120 This error indicates the size of the model + overhead is too large for the GPU? But this is roughly a 700mb model trying to perform inference on a single 640x512 grayscale image on a GPU that has roughly 6.5Gb of free space. I even tried training a new model on images at half that resolution and get the same error. Any tips/feedback is appreciated. PyTorch or Caffe2: PyTorch How you installed PyTorch (conda, pip, source): Source Build command you used (if compiling from source): Followed jetson-reinforcement github OS: Ubuntu 16.04 on Nvidia TX2 PyTorch version: 0.3.0 Python version: 3.6 CUDA/cuDNN version: CUDA 9.0 , CUDNN 7 GPU models and configuration: Tegra X2 GCC version (if compiling from source): 5.4 CMake version: 3.5.1 Versions of any other relevant libraries:\nwe should fix the launch parameters. I presume we cant use as many threads per block on the TX2 as we use on desktop GPUs. can you tell us what the limits of TX2 GPUs are for the fix to im2col\nThreads per block and maximum blocks in the grid are actually the same for TX2 as they are for desktop GPUs , only number of registers is smaller. do you know what might be causing this?\nas said, the number of threads per block is constant across different NVIDIA GPUs, the Tegra series included. Is it possible that it was compiled to require more registers than is available to the TX2, and maybe the kernel invocation of im2col requires some sort of launch_bounds() command?\nTX2 is arm64 platform. I presume pytorch is compiled from source. In this case, there isn't a chance that it was compiled to requre more registers than available on TX2 I think. The build log will be helpful to see.\nthanks for the quick response.", "I will try dig up the build log tomorrow and post it here.\nif there are no launch bounds, it is in fact possible that a kernel is compiled to request more registers than available. At the compile time, compiler does not know how many threads you'll want to launch with, so, potentially, it can use too many registers per thread to later satisfy runtime requirements, and, e.g. trying to launch with 512 or 1024 threads could fail (not even a single block can be put on an SM), whereas launching with 256 would succeed.\nis there a way where we can audit that GPU constraints on our server setting (i.e. not actually sitting and compiling TX2 PyTorch)\nAdding launch_bounds with the max number of threads the kernel is going to be launched with will cause compiler not to overuse registers. We had to do it e.g. for interp kernels when cuda 9 started using more registers\nAhh, that was as I suspected Thank you for clearing this up. do you guys think a fix will be released in the near future (week?). Thanks for all the help\nWorks now, thanks!\nhello, I met the same error on tx2, complied pytorch 0.3.0.. But the error is this: 'cuda runtime error(7): too many resources requested for launch at '. The file is different with '', is that I need add this line 'launch_bounds(1024)' to every function in file ''? Thank you!\nHello I am also facing a similar issue: Here's my setup: How you installed PyTorch (conda, pip, source): Source Build command you used (if compiling from source): Followed jetson-reinforcement github OS: Ubuntu 16.04 on Nvidia TX2 PyTorch version: 0.5.0a0+ (torchvision 0.2.1) Python version: 3.5 CUDA/cuDNN version: CUDA 9.0 , CUDNN 7 GPU models and configuration: Tegra X2 GCC version (if compiling from source): 5.4 CMake version: 3.12.2 Source code: Full log: Do you know what it could be?", "I know that '' without launch_bounds(1024) leads to this error, but I don't know how to fix it... Tried this solution , but still not working (after compilation)\nMy issue was solved by following\nHi, I have the same problem. Can this problem solved by gitting clone the new pytorch, than changing CUDANUMTHREADS =256 in the below two file and compiling it? !\nthat appears to fix some of the functions, but not perhaps all - not sure. For more info, see:"], "neg": []}
{"query": "std::make_shared<autograd::Eval>()->replaceSubgraph(inputs, outputs); } <del> Node* recordTraceHelper(std::string op, // TODO: make this a Symbol at::ArrayRef<Variable> inputs, at::ArrayRef<Variable> outputs) { </del> <ins> Node* recordTrace(std::string op, // TODO: make this a Symbol at::ArrayRef<Variable> inputs, at::ArrayRef<Variable> outputs) { </ins> auto state = getTracingState(inputs); auto& graph = state->graph; // TODO: Technically, we could reduce the scope of the lock, but since we", "pos": ["when executing:\nI'm seeing this as well as of last night. Its preceded by many warnings about : But the first error is the one about . I went back, and I get a good install as late as commit"], "neg": []}
{"query": "add_docstr(torch._C.triu, \"\"\" <del> triu(input, k=0, out=None) -> Tensor </del> <ins> triu(input, diagonal=0, out=None) -> Tensor </ins> Returns the upper triangular part of the matrix (2D Tensor) :attr:`input`, the other elements of the result Tensor :attr:`out` are set to 0.", "pos": ["The states that the argument controls which diagonal to consider. However, this argument is actually named in pytorch 0.2.0."], "neg": []}
{"query": "tensor_view = tensor.view(5, 3) tensor_view.fill_(random.uniform(0, 1)) self.assertEqual((tensor_view-tensor).abs().max(), 0) <ins> self.assertEqual(empty.view_as(empty), empty) self.assertEqual(empty.view(0), empty) </ins> def test_expand(self): result = torch.Tensor()", "pos": ["Please check and fix all the PyTorch web links format from to , like in the official site.\nclosed via"], "neg": []}
{"query": "// most fields THPFunction* self = (THPFunction*)obj; new (&self->cdata) torch::autograd::PyFunction(obj); <ins> new (&self->cdata_ptr) std::weak_ptr<torch::autograd::PyFunction>(); </ins> self->cdata.num_inputs = -1; self->cdata.is_stochastic = PyObject_IsInstance(obj, THPStochasticFunctionClass); return obj;", "pos": ["there is a bug. the code is from The code for installing cuda10.2 should be corrected to The code for installing cuda10.2 should be corrected to cc\nThank you for reporting the typo", "This is how I implement the decoder of a sequence to sequence model I'm not sure about the new autograd mechanics but this worked in the previous version. If I didn't make the unrolling codes a function it will work. It will also work with CPU. I compiled from source with Python 2.7, CUDA 8.0 and Cudnn 6\nThanks for the minimal reproduction! It broke OpenNMT and a big internal model but we weren't looking forward to isolating what parts of those had the issue.\nThat's great, I'll take a look today. Thanks!\nI've pushed a fix to .\nHow does the inference work in this case? I mean, once you train it how do you test it in the test phase?"], "neg": []}
{"query": "self.assertEqual(input_var.requires_grad, broadcasted_var.requires_grad) @unittest.skipIf(not TEST_MULTIGPU, \"multi-GPU not supported\") <ins> def test_broadcast_no_grad(self): x = torch.randn(1, 2, dtype=torch.cuda.float32, requires_grad=True) with torch.no_grad(): broadcasted = Broadcast.apply((0, 1), x) self.assertTrue(x.requires_grad) for output in broadcasted: self.assertFalse(output.requires_grad) @unittest.skipIf(not TEST_MULTIGPU, \"multi-GPU not supported\") </ins> def test_replicate(self): module = nn.Linear(10, 5).float().cuda() input = Variable(torch.randn(2, 10).float().cuda())", "pos": ["OS: ubuntu 16.04 PyTorch version: 0.4.0a0+ How you installed PyTorch (conda, pip, source): source Python version: 3.6.3 CUDA/cuDNN version: v9.1, cudnn-9.1-linux-x64-v7 GPU models and configuration: 4x Tesla M60 GCC version (if compiling from source): 5.4.0 When using context () on a model with dataparallel operation (same with ) on 1 GPU the flag requires_grad does not come back to True. If only one GPU is used then everything works fine. Script: On 2 GPUs, this outputs which is wrong in the last line. On 1 GPU (set ) the output is correct.\nis looking into it"], "neg": []}
{"query": "int batch_size = THTensor_(size)(input, 0); THTensor_(resize1d)(output, batch_size); <ins> int invalid_target = -1;  // We cannot throw an exception inside omp parallel </ins> int i; #pragma omp parallel for private(i) for (i = 0; i < batch_size; i++) { int cur_target = THTensor_fastGet1d(target, i) - TH_INDEX_BASE; <del> if (cur_target == ignore_index) { THTensor_fastSet1d(output, i, 0.0f); continue; </del> <ins> if (cur_target >= 0 && cur_target < n_classes) { if (cur_target == ignore_index) { THTensor_fastSet1d(output, i, 0.0f); continue; } real cur_weight = weights ? THTensor_fastGet1d(weights, cur_target) : 1.0f; THTensor_fastSet1d(output, i, -THTensor_fastGet2d(input, i, cur_target) * cur_weight); } else { THAtomicCompareAndSwap(&invalid_target, -1, cur_target); </ins> } <del> real cur_weight = weights ? THTensor_fastGet1d(weights, cur_target) : 1.0f; THTensor_fastSet1d(output, i, -THTensor_fastGet2d(input, i, cur_target) * cur_weight); </del> <ins> } if (invalid_target >= 0) { THError(\"Target %d out of bounds\", invalid_target); </ins> } return;", "pos": ["Claim: When I use an invalid label with F.crossentropy with reduce=False, I don't receive an error message. I do receive an error message when reduce=True. Code: Result: No error message received. Expected error message: Context: I was using , accidentally used an invalid label, and got extremely large losses and extremely small losses, and inconsistently so. Environment: OS: Ubuntu 16.04 on Docker and host PyTorch version: 0.3.0.post4 How you installed PyTorch: conda Python version: 3.6.3"], "neg": []}
{"query": "limits = new_limits # Find the first and the last bin defining the support of the histogram: <del> cum_counts = np.cumsum(np.greater(counts, 0, dtype=np.int32)) </del> <ins> cum_counts = np.cumsum(np.greater(counts, 0)) </ins> start, end = np.searchsorted(cum_counts, [0, cum_counts[-1] - 1], side=\"right\") start = int(start) end = int(end) + 1", "pos": ["The use of in is deprecated a while ago, and the deprecation expired in Numpy 1.24+: Source: Minimal example: Error message with Numpy 1.24+: This error does not occur in Numpy 1.23.5. Numpy should have print a deprecation warning about this: but I think PyTorch has suppressed the warning.\n(related: there are a few more numpy1.24 compatibility issues at )\nso how can use add_histogram function in numpy 1.24.1?\nSince this issue has been resolved, I believe you can use  as before (i.e., as in NumPy < 1.24).\nI was using Python 3.9, torch 1.9.1 (with CUDA 11.1) and numpy 1.24.2 to run something today and still ran into this exact problem, don't really know about pytorch 2.0+ though. Thank you for the post though, otherwise might have cost me several more hours to figure this out.\nI have this issue during yolo v7 training - cum_counts = (np.greater(counts, 0, dtype=np.int32)) TypeError: No loop matching the specified signature and casting was found for ufunc greater.\nHi, I'm still getting this error with numpy 1.24 and tensorboard 2.15.\nsame problem with numpy==1.24.3 torch==1.13.1+cu117 tensorboard==2.13.0\nSame problem with numpy == 1.24.0 tensorboard == 2.14.0 torch == 1.13.1\nSame problem with numpy == 1.26.4 tensorboard == 2.16.2 torch == 1.13.1+cu116\nUsing can solve this issue\nsame problem when I used with numpy==1.26.4 torch==1.13.0 tensorboard==2.16.2\ngo to the path of the in your or your main env like this -C:Userslegionanaconda3envsganLibsite-packagestorchutilstensorboard open and change following line to"], "neg": []}
{"query": "if (threadIdx.x == 0) { // Make sure the sum of our distribution didn't overflow assert(!isinf(sum)); <ins> assert(THCNumerics<AccT>::gt(sum, accZero)); </ins> asmem[0] = sum; smem[0] = sampled[curDist];", "pos": ["Looking at the values returned from tn.multinomial(1) and counting them, it seems like it returns the values of the multinomial distribution taken from t, i.e. that it has ignored the signs of the values. However, it seem like always returns a tensor filled with zeros.\nlooking into this"], "neg": []}
{"query": "<del> FROM nvidia/cuda:9.1-cudnn7-devel-ubuntu16.04 RUN echo \"deb http://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1604/x86_64 /\" > /etc/apt/sources.list.d/nvidia-ml.list </del> <ins> FROM nvidia/cuda:9.0-cudnn7-devel-ubuntu16.04 </ins> RUN apt-get update && apt-get install -y --no-install-recommends  build-essential ", "pos": ["OS: MacOS Sierra PyTorch version: How you installed PyTorch (conda, pip, source): Just pulled the dockerfile I pulled the dockerfile just now and tried to build it using a tool that we plan on open-sourcing at some point. The command translates to a simple .\nCC\nIt seems to work now."], "neg": []}
{"query": "# for-loop bodies don't define scopes, so we have to save the variables # we want to close over in some way def do_test(self, name=name, self_size=self_size, args=new_args, test_name=test_name, <del> test_output_indices=test_output_indices): </del> <ins> output_process_fn=output_process_fn): </ins> def check(name): is_magic_method = name[:2] == '__' and name[-2:] == '__' is_inplace = name[-1] == \"_\" and not is_magic_method", "pos": ["I am having problem in running the backward propagation with svd. In particular, it seems working when the first dimension of the matrix x (below) is greater than the second. But I get the following error if the first is smaller than the second. PyTorch version: 2018.04.20 Is debug build: No CUDA used to build PyTorch: 8.0.61 OS: Ubuntu 16.04.4 LTS GCC version: (Ubuntu 5.4.0-6ubuntu1~16.04.9) 5.4.0 CMake version: Could not collect Python version: 3.5 Is CUDA available: No CUDA runtime version: No CUDA GPU models and configuration: No CUDA Nvidia driver version: No CUDA cuDNN version: No CUDA Versions of relevant libraries: [pip3] numpy (1.13.3) [pip3] torch (2018.4.20) [pip3] torchvision (0.2.0) [conda] pytorch 0.3.1 py35cuda8.0.61cudnn7.0.52 pytorch [conda] pytorch-nightly 2018.04.20 py35cuda8.0.61cudnn7.1.21 pytorch [conda] torchvision 0.2.0 py35heaa392f_1 pytorch\nRepro'ed on 0.4 branch. works btw.\nWhat does \"Repro'ed\" means? Should I install from that branch? I need to compute the kernel of a matrix, therefore I need the last columns of the matrix V (which is why I have to pass ).\nIt's a shorthand for \"reproduced\", which means that the bug still hasn't been fixed. Sorry, we'll try to get to it soon."], "neg": []}
{"query": "but they must be :ref:`broadcastable <broadcasting-semantics>`. .. math:: <del> text{out}_i = min(text{tensor}_i, text{other}_i) \"\"\" + r\"\"\" </del> <ins> text{{out}}_i = min(text{{tensor}}_i, text{{other}}_i) </ins> .. note:: When the shapes do not match, the shape of the returned output tensor follows the :ref:`broadcasting rules <broadcasting-semantics>`.", "pos": ["with , both max and min are broken <img width=\"880\" alt=\"Screen Shot 2020-03-23 at 1 10 07 PM\" src=\"https://user-\"on , min is broken <img width=\"875\" alt=\"Screen Shot 2020-03-23 at 1 13 23 PM\" src=\"https://user-\"\nFixed by and"], "neg": []}
{"query": "endforeach(flag_var) endif() <del> set (CMAKE_CXX_FLAGS_DEBUG \"${CMAKE_CXX_FLAGS_DEBUG} -fno-omit-frame-pointer\") set (CMAKE_LINKER_FLAGS_DEBUG \"${CMAKE_STATIC_LINKER_FLAGS_DEBUG} -fno-omit-frame-pointer\") </del> <ins> set (CMAKE_CXX_FLAGS_DEBUG \"${CMAKE_CXX_FLAGS_DEBUG} -fno-omit-frame-pointer -O0\") set (CMAKE_LINKER_FLAGS_DEBUG \"${CMAKE_STATIC_LINKER_FLAGS_DEBUG} -fno-omit-frame-pointer -O0\") </ins> if (USE_ASAN) set (CMAKE_CXX_FLAGS_DEBUG \"${CMAKE_CXX_FLAGS_DEBUG} -fsanitize=address\") set (CMAKE_LINKER_FLAGS_DEBUG \"${CMAKE_STATIC_LINKER_FLAGS_DEBUG} -fsanitize=address\")", "pos": ["I think this may have been intentional, so I want to float this first before I apply the change. Historically, PyTorch did DEBUG builds with , because it prevents inlining, which means that you can inspect variables in all stack frames. However, in master, DEBUG mode appears to be compiled with optimizations, which means that any inlined functions are non-inspectable. The following patch will turn off optimizations when compiled in DEBUG mode. Is there any reason not to apply this?"], "neg": []}
{"query": "// See Note [Numpy memory management] &THNumpyArrayAllocator, new NumpyArrayAllocator(numpy_array))); <ins> THStorage_(clearFlag)(storage.get(), TH_STORAGE_RESIZABLE); </ins> result = THTensor_(newWithStorage)(LIBRARY_STATE storage, 0, sizes, strides); } else", "pos": ["When both PyTorch and Pandas are imported, calling results in a segfault. The issue was originally reported by It looks like calling returns a null pointer here: Examining in GDB: I suspect it's both PyTorch and Pandas doing threading magic and somehow getting in each other's way. I haven't yet been able to provide a minimal example that triggers this behaviour and does not use my package; it's likely that I'm doing something odd that causes this. More documentation here: Possibly related issues: and Python version: 3.6 PyTorch version: v0.2.0 Pandas version: 0.20.1"], "neg": []}
{"query": "def parameters(self, memo=None): if memo is None: memo = set() <del> super(Container, self).parameters(memo) </del> <ins> for p in super(Container, self).parameters(memo): yield p </ins> for module in self.children(): for p in module.parameters(memo): yield p", "pos": ["While using torch on a Macbook pro with M1 ship, using python 3.8.10 and torch 1.12.1, I came across a weird bug that happened because I forgot to put a tensor on MPS. The reproducible code is the following: As you can see, if is on MPS and on CPU, the line wrongly replaces the index in place of the index . I would expect to have an error saying both tensors should be on the same device instead. I get an error: cc\nThe error seems to happen somewhere within Also I'm getting a warning \"The operator 'aten::nonzero' is not currently supported on the MPS backend and will fall back to run on the CPU. This may have performance implications\" when printing tensors on MPS.\nHi - thanks for reporting this issue. This is actually an allowed usecase (copying CPU tensors into MPS tensors / MPS -CPU / MPS -MPS): The issue is on MPS side when copying a CPU tensor into a strided MPS tensor. I'll take a look."], "neg": []}
{"query": "HingeEmbeddingLoss, MarginRankingLoss backend.register_function('Linear', Linear) <del> backend.register_function('BatchNorm', BatchNorm) </del> backend.register_function('ConvNd', ConvNd) backend.register_function('RNN', RNN) backend.register_function('RNNTanhCell', RNNTanhCell)", "pos": ["As Arthur Szlam reports, fb-internal cudnn is still lagging behind, and giving batch-size 1024 with batchnorm is raising an error from cudnn. Need to check for compile-time version and disable this codepath\nThe 1024 limitation was removed in 5.1.10"], "neg": []}
{"query": "import torch from torch.distributions import constraints from torch.distributions.distribution import Distribution <del> from torch.distributions.utils import probs_to_logits, logits_to_probs, log_sum_exp, lazy_property, broadcast_all </del> <ins> from torch.distributions.utils import probs_to_logits, logits_to_probs, _log_sum_exp, lazy_property, broadcast_all </ins> class Categorical(Distribution):", "pos": ["Leaving this issue here for higher visibility. Originally reported by here: Also mentioned by Piggybacking off of this issue, has strange behavior (reported here: )\nI have fixed this locally. Could I send a PR?\nYes, go ahead!"], "neg": []}
{"query": "def fn(*inputs): output = getattr(inputs[0], name)(*inputs[1:]) <del> if test_output_indices is None: return output else: return tuple(output[i] for i in test_output_indices) </del> <ins> return output_process_fn(output) </ins> if not is_inplace and name not in EXCLUDE_GRADCHECK: run_grad_and_gradgrad_checks(self, name, test_name, fn,", "pos": ["I am having problem in running the backward propagation with svd. In particular, it seems working when the first dimension of the matrix x (below) is greater than the second. But I get the following error if the first is smaller than the second. PyTorch version: 2018.04.20 Is debug build: No CUDA used to build PyTorch: 8.0.61 OS: Ubuntu 16.04.4 LTS GCC version: (Ubuntu 5.4.0-6ubuntu1~16.04.9) 5.4.0 CMake version: Could not collect Python version: 3.5 Is CUDA available: No CUDA runtime version: No CUDA GPU models and configuration: No CUDA Nvidia driver version: No CUDA cuDNN version: No CUDA Versions of relevant libraries: [pip3] numpy (1.13.3) [pip3] torch (2018.4.20) [pip3] torchvision (0.2.0) [conda] pytorch 0.3.1 py35cuda8.0.61cudnn7.0.52 pytorch [conda] pytorch-nightly 2018.04.20 py35cuda8.0.61cudnn7.1.21 pytorch [conda] torchvision 0.2.0 py35heaa392f_1 pytorch\nRepro'ed on 0.4 branch. works btw.\nWhat does \"Repro'ed\" means? Should I install from that branch? I need to compute the kernel of a matrix, therefore I need the last columns of the matrix V (which is why I have to pass ).\nIt's a shorthand for \"reproduced\", which means that the bug still hasn't been fixed. Sorry, we'll try to get to it soon."], "neg": []}
{"query": "-DCMAKE_SHARED_LINKER_FLAGS=\"$LDFLAGS\"  -DCMAKE_INSTALL_LIBDIR=\"$INSTALL_DIR/lib\"  -DCUDA_NVCC_FLAGS=\"$C_FLAGS\"  <del> -DCWRAP_FILES_BASE=\"$BASE_DIR/torch\"  </del> <ins> -Dcwrap_files=\"$CWRAP_FILES\"  </ins> -DTH_INCLUDE_PATH=\"$INSTALL_DIR/include\"  -DTH_LIB_PATH=\"$INSTALL_DIR/lib\"  -DTH_LIBRARIES=\"$INSTALL_DIR/lib/libTH$LD_POSTFIX\" ", "pos": ["There are previous issues addressing this, but I want to write this up with a plan in a new issue. We should make differentiable instead of having a wrapper class. The Variable class adds conceptual overhead without sufficient benefit. It makes it harder for new users to adopt PyTorch by adding a second API which is similar but not exactly the same as Tensor. There are few prerequisites: on s need to be at least as fast as operations on () need scalar support at the Python level () The constructor will be deprecated and become a no-op. The accessor will be deprecated and return a view that has requiresgrad=False. The constructors will take an optional keyword argument There will be a number of internal changes, but we shouldn't break most existing code: performance reasons, we will no longer build up the graph of operations by default when unless the operation uses stochastic variables. We'll use a context manager or global variable to enable graph creation. new differentiable Tensor will need support all the operations currently on Tensor, including in-place and operations. -place operations on leaf Variables which requiregrad are not currently supported. In the new API, the tensor's value will be updated and it will remain a leaf node. (This is important for optimizers, which update leaf Variables in-place). There are probably a few other subtle differences between Tensor and Variable APIs. I'll update this issue as we address them. See also: 23/738 # Greg will be fixing the tests, Tensor vs. Variable requiresgrad 2/1837 # Just testbroadcast from TestTorch OK OK # error at top level OK OK OK OK OK OK OK OK 10/120 # missing sspaddmm, dsmm/mm, addmm 1/257 TestTorch.testbroadcast 2OK\nProperly handling in-place might not be as easy as it seems. Sometimes you actually need to move the Variable in the graph. Consider this case:\nOne of the cases of differences between variable and tensor is computation of var/std:\nIMO Variable gives user a clear difference between graph node and data. .data attribute and Variable constructor provide convenient and explicit ways to take control of the graph (or graphs).\nThis is done. (Still some clean-up to do)\nIs the conda PyTorch out of date? I'm getting the following error: .\nYes, we haven't released 0.4, which has the Variable Tensor merge, yet"], "neg": []}
{"query": "z3-solver==4.12.2.0 tensorboard==2.13.0 optree==0.9.1 <ins> # NB: test_hparams_* from test_tensorboard is failing with protobuf 5.26.0 in # which the stringify metadata is wrong when escaping double quote protobuf==3.20.2 </ins>", "pos": ["Platforms: macos, mac This test was disabled because it is failing on main branch (). cc\n<!-- validation-comment-start --<bodyHello there! From the DISABLED prefix in this issue title, it looks like you are attempting to disable a test in PyTorch CI. The information I have parsed is below: Test name: Platforms for which to skip the test: mac, macos Disabled by Within ~15 minutes, will be disabled in PyTorch CI for these platforms: mac, macos. Please verify that your test name looks correct, e.g., . To modify the platforms list, please include a line in the issue body, like below. The default action will disable the test for all platforms if no platforms list is specified. We currently support the following platforms: asan, dynamo, inductor, linux, mac, macos, rocm, slow, win, windows.</body<!-- validation-comment-end --"], "neg": []}
{"query": "return PyLong_FromLong((int64_t) driverVersion); } <ins> PyObject * THCPModule_getCompiledVersion(PyObject *self) { return PyLong_FromLong((long) CUDA_VERSION); } </ins> PyObject * THCPModule_getRNGState(PyObject *_unused) { HANDLE_TH_ERRORS", "pos": ["There's quite a few user-reports of installing the CUDA 7.5 binary of pytorch and then running on GTX 1080 / 1070, and there will be reports of installing 7.5 / 8.0 binaries and running on V100. Raising a user-warning in such situations and letting them know will fix the problem.\nThe version of CUDA used to compile the binary should be available thanks to this PR:\nI think is a more robust way of getting the CUDA used to compile PyTorch, as currently if there are multiple CUDA installed in the machine, the heuristics used in might not match those used by CMake FindCUDA. I think that we should modify the to return the cuda version used during compilation. I have the impression that it will not require changing much because doesn't require cuda to execute. What do you think?"], "neg": []}
{"query": "If :attr:`compute_uv` is ``False``, the returned `U` and `V` matrices will be zero matrices of shape :math:`(m times m)` and :math:`(n times n)` respectively. :attr:`some` will be ignored here. <ins> .. note:: The singular values are returned in descending order. If :attr:`input` is a batch of matrices, then the singular values of each matrix in the batch is returned in descending order. </ins> .. note:: The implementation of SVD on CPU uses the LAPACK routine `?gesdd` (a divide-and-conquer algorithm) instead of `?gesvd` for speed. Analogously, the SVD on GPU uses the MAGMA routine `gesdd` as well.", "pos": ["Documentation for does not indicate whether or not singular values are returned in descending order. My suspicion is that they are since both LAPACK and MAGMA's ?gesdd return them in descending order. It seems like this should be clarified in the docs to avoid unnecessary sorting, or mistaken assumptions. <!-- A clear and concise description of what content in is an issue. If this has to do with the general website, please file an issue at instead. If this has to do with , please file an issue at --"], "neg": []}
{"query": "auto set_history = [&](Variable& var, uint32_t output_nr, bool is_input, bool is_modified, bool is_differentiable) { if (!is_differentiable) { <del> if (!var.requires_grad()) return; </del> <ins> if (!var.requires_grad()) { return; } </ins> // NB: we don't support returning non-differentiable views that could require grad <del> // (this could happen if someone were to return an input to the function). </del> if (var.is_view()) { throw std::runtime_error(\"Returning Variables sharing storage with other Variables \" \"that require grad is not supported in Python functions. \" \"Please submit a feature request if you hit this error.\"); } <del> var.detach_(); </del> <ins> // Return detached aliases of inputs, instead of changing their requires_grad // property. if (is_input) { var = var.detach(); } else { var.detach_(); } </ins> } else if (is_modified) { if (var.is_leaf() && var.requires_grad()) { throw std::runtime_error(\"a leaf Variable that requires grad has been used in an in-place operation.\");", "pos": ["OS: ubuntu 16.04 PyTorch version: 0.4.0a0+ How you installed PyTorch (conda, pip, source): source Python version: 3.6.3 CUDA/cuDNN version: v9.1, cudnn-9.1-linux-x64-v7 GPU models and configuration: 4x Tesla M60 GCC version (if compiling from source): 5.4.0 When using context () on a model with dataparallel operation (same with ) on 1 GPU the flag requires_grad does not come back to True. If only one GPU is used then everything works fine. Script: On 2 GPUs, this outputs which is wrong in the last line. On 1 GPU (set ) the output is correct.\nis looking into it"], "neg": []}
{"query": "Otherwise, :attr:`dim` is squeezed (see :func:`torch.squeeze`), resulting in the output tensor having 1 fewer dimension than :attr:`input`.\"\"\"}) factory_common_args = merge_dicts(common_args, parse_kwargs(\"\"\" dtype (:class:`torch.dtype`, optional): the desired data type of returned tensor. Default: if ``None``, uses a global default (see :func:`torch.set_default_tensor_type`).", "pos": ["with , both max and min are broken <img width=\"880\" alt=\"Screen Shot 2020-03-23 at 1 10 07 PM\" src=\"https://user-\"on , min is broken <img width=\"875\" alt=\"Screen Shot 2020-03-23 at 1 13 23 PM\" src=\"https://user-\"\nFixed by and"], "neg": []}
{"query": "\"\"\"Fills the input Tensor or Variable with values according to the method described in \"Understanding the difficulty of training deep feedforward neural networks\" - Glorot, X. & Bengio, Y. (2010), using a normal distribution. The resulting tensor will have values sampled from :math:`N(0, std)` where <del> :math:`std = gain times sqrt{2 / (fan_in + fan_out)}`. Also known as Glorot initialisation. </del> <ins> :math:`std = gain times sqrt{2 / (fan_in + fan_out)}`. Also known as Glorot initialisation. </ins> Args: tensor: an n-dimensional torch.Tensor or autograd.Variable", "pos": ["why 1.21.6? because that's the oldest supported numpy on windows/python3.10: all cc\nHere is the test for pinning numpy to 1.21.6 The reason its was not pinned to this version is that conda can't properly resolve all the dependencies at this step here:\nIt looks like you are correct, I will put up a PR to pin lower then already pinned:\nHere is the PR:"], "neg": []}
{"query": "#include \"torch/csrc/autograd/engine.h\" #include <set> <ins> #include <string> </ins> #include <THPP/THPP.h> using thpp::Tensor;", "pos": ["Not sure what is the reason for these errors, any suggestions? I suspect the clang version is not supported? Here is the\nclang have more strict checking , pull request will fixed it.\nfixed, thanks to"], "neg": []}
{"query": "\"\"\" return self._apply(lambda t: t.cuda(device_id)) <del> def cpu(self, device_id=None): </del> <ins> def cpu(self): </ins> \"\"\"Moves all model parameters and buffers to the CPU.\"\"\" return self._apply(lambda t: t.cpu())", "pos": ["as shown in source code, this parameters is not used. Any reason that we keep it?\nthis can be cleaned up."], "neg": []}
{"query": "uniforms = clamp_probs(self.logits.new(self._extended_shape(sample_shape)).uniform_()) gumbels = -((-(uniforms.log())).log()) scores = (self.logits + gumbels) / self.temperature <del> return scores - log_sum_exp(scores) </del> <ins> return scores - _log_sum_exp(scores) </ins> def log_prob(self, value): K = self._categorical._num_events", "pos": ["Leaving this issue here for higher visibility. Originally reported by here: Also mentioned by Piggybacking off of this issue, has strange behavior (reported here: )\nI have fixed this locally. Could I send a PR?\nYes, go ahead!"], "neg": []}
{"query": "def make_grad_input(fn): def call_grad_input(self, bufs, input, weight, grad_output): backend = type2backend[type(input)] <del> grad_input = input.new().resize_as_(input).zero_() </del> <ins> grad_input = input.new().resize_as_(input) </ins> kernel_size = weight.size()[2:] args = parse_arguments(self, fn.arguments[5:], bufs, kernel_size) getattr(backend, fn.name)(backend.library_state, input, grad_output,", "pos": ["THNN convolutions are slower due to unnecessary .zero() calls on gradients: is not needed same for bias, prob worth adding a constant in C-side that controls whether to override grad or not\nSame for batch_norm and auto thnn functions, not sure if all of them though. Prob the speedups will be quite significant.\nI think they are needed. THNN functions accumulate the gradient instead of overwriting the buffer, don't they?\nonly grad wrt weight, inputs are overriden\nWe should reopen this, still many functions have unnecessary .zero() calls on gradinput and gradparam, e.g.: : : most of the functions don't need it\nI think the ones are currently necessary.\nfor nn yes, but cudnn has: and on cudnnConvolutionBackwardFilter and on cudnnBatchNormalizationBackward which control addition of gradients and should be set to 1 and 0 in pytorch cudnn (it is and 1 in )\nYeah, we can remove it on BatchNorm. If you look at Conv, it already only zeros out grad_weight for cunn.\nthis is now fixed in master:"], "neg": []}
{"query": "} auto SavedVariable::unpack() -> std::shared_ptr<Variable> { <del> if (!data) return nullptr; </del> <ins> if (!data) { if (version) { throw std::runtime_error(ERR_BACKWARD_TWICE); } return nullptr; } </ins> int current_version = **version; if (expected_version != current_version) {", "pos": ["Usually calling backwards twice gives this message: But conv2d just segfaults. $ pt A B Segmentation fault (core dumped)"], "neg": []}
{"query": "('trace', (M, M), ()), ('cross', (S, 3), ((S, 3),)), ('cross', (S, 3, S), ((S, 3, S), 1), 'dim'), <ins> ('inverse', (S, S), (), '', (), [skipIfNoLapack]), </ins> ('clone', (S, M, S), ()), ('eq', (S, S, S), ((S, S, S),)), ('ne', (S, S, S), ((S, S, S),)),", "pos": ["Command to install the latest nightly binary: installs : However, PyTorch itself reports : I guess this version tag might be defined during the build process from e.g. and then forwarded from , but also source builds should use based on the tag. Unsure, if this is just a reporting issue or if we are building the nightly binaries with and executing them with , so creating this issue to dig a bit further into it. CC cc\nThis is really weird, I've put up measure in place to validate against it... Will investigate today (and moving to pytorch repo)\ncan you please check that 2.2.0 reports (and is build) against the correct version of nccl library?\nConfirmed for nightlies : Same issue for RC:\nLooking at the build log Looks like we are compiling with: USESYSTEMNCCL\nOn Windows I see following exception: We compiling with should we return None rather then exception ?\nIn this case we are indeed using 2.18.1: Do you know why we are using the system-wide NCCL installation instead of the one? Should we delete the system NCCL and use by default?\nLooks like its set here: Looks like it was by: Let me propose the PR to unset this flag : USESYSTEM_NCCL\nI guess we are using the system NCCL installation to be able to during the runtime. If we would use the module I assume we would link NCCL into the PyTorch binaries. If so, we should make sure to update the NCCL version whenever is updated. This could break our dependency on the NCCL pip package.\nYes. Also when we update this nccl here:\nWe used following path  What would be good path for nccl 2.19.3 and cuda 12.1 ? As per this issue: Will be using 12.0 ncck 2.19.3\nInstalling system NCCL into the nightly build docker container feels great, but we need to figure out, how to keep it in sync with PyTorch. I wonder if one can, for example, just check nccl version from this requirements file and build from source?\nyes probably we should build and upload nccl for our supported cuda versions, same way as we build docker containers, by following details here:\nVerified with final rc"], "neg": []}
{"query": "case NPY_DOUBLE: return kDouble; case NPY_FLOAT: return kFloat; case NPY_HALF: return kHalf; <ins> case NPY_LONGLONG: </ins> case NPY_INT64: return kLong; case NPY_INT32: return kInt; case NPY_INT16: return kShort;", "pos": ["Minimal example: Torch initializes the tensor with values which are different from the numpy array. I've seen issues for similar kinds of problems (see ). Maybe the easiest solution to all these initialization problems might be, to check for the same values of the original and initialized array/tensor instead of initializing tensors silently with wrong numbers. What do you think? Furthermore, I've opened an issue which is also related this one in terms of initializing a numpy array with a number of type and then initializing a torch tensor from that array, see Feel free to merge them if you think that both issues can be solved together.\nCould you please provide some information about what version of pytorch and numpy you're using? ! It works as expected for me on master\nI'm using torch 0.3.0.post4 (latest one on website of PyTorch) and numpy 1.13.3.\nI'm running the same thing, this is strange. In the first line , I assumed that came from . Is this the case?\nNope, it's indeed the python built-in type. Actually this issue came up when I tried to get the shape of (i.e. , , returns ).\nThanks for the clarification, I'm now seeing your results. Looks like the tensor constructor doesn't work well with python2.7 long types in this case\nSure, no problem. Thanks for looking into it!\nThe problem seems to be that np.longlong and np.int64 are the same thing (at least on a 64bit system, I think), but the NPYLONGLONG and NPYINT64 macros are different... I'll put up a pr for this soon.", "A minimal example: Although has and this type is listed as supported, the array can't be converted. Consider that is initialized with an upper bound of type . I'm not completely sure, whether this issue is related to torch or numpy, but numpy initializes the array with values as expected, whereas torch throws an inconsistent error message.\nEdit: nevermind, I didn't run the same exact thing ! Works for me on master and 0.3. Seems like it could be related to"], "neg": []}
{"query": "#endif /* !IS_CUDA */ #undef BUILD_REAL_FMT", "pos": ["On Python 3.5.2 errors out with . map and map2 seem to be work just fine."], "neg": []}
{"query": "<ins> [[ name: THPTensor_(elementSize) python_name: element_size cpu_half: True auto_gpu: False only_register: True ]] [[ name: THPTensor_(storage) python_name: storage cpu_half: True auto_gpu: False only_register: True ]] [[ name: storageOffset python_name: storage_offset cpu_half: True auto_gpu: False return: long arguments: - THTensor* self ]] [[ name: nDimension python_name: ndimension cpu_half: True auto_gpu: False return: long arguments: - THTensor* self ]] [[ name: THPTensor_(nDimension) python_name: dim cpu_half: True auto_gpu: False only_register: True method_flags: METH_KEYWORDS ]] [[ python_name: index name: THPTensor_(getValue)<true> only_register: True override_method_flags: METH_O ]] [[ python_name: _set_index name: THPTensor_(setIndex) only_register: True ]] [[ python_name: _check_advanced_indexing name: THPTensor_(checkAdvancedIndexing) cpu_half: False only_register: True override_method_flags: METH_O ]] [[ python_name: _advanced_index_add name: THPTensor_(advancedIndexAdd) cpu_half: False only_register: True ]] [[ python_name: _advanced_index_select name: THPTensor_(advancedIndexSelect) cpu_half: False only_register: True ]] [[ name: resize_ return: self cname: resize cpu_half: True before_call: THPUtils_assert(arg_self->storage->flag & TH_STORAGE_RESIZABLE, \"calling resize_ on a tensor that has non-resizable storage. Clone it first \" \"or create a new tensor instead.\"); arguments: - THTensor* self - arg: THSize* size long_args: True - CONSTANT NULL ]] [[ name: zeros variants: - function auto_gpu: False return: argument 0 arguments: - arg: THTensor* result output: True - arg: THSize* size long_args: True ]] [[ name: zeros_like cname: zerosLike variants: - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* input ]] [[ name: ones variants: - function auto_gpu: False return: argument 0 arguments: - arg: THTensor* result output: True - arg: THSize* size long_args: True ]] [[ name: ones_like cname: onesLike variants: - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* input ]] [[ name: numel return: long cname: nElement cpu_half: True auto_gpu: False variants: - method - function arguments: - THTensor* self ]] [[ name: THPTensor_(numel) python_name: nelement cpu_half: True auto_gpu: False only_register: True method_flags: METH_KEYWORDS ]] [[ name: set_ cname: set cpu_half: True auto_gpu: False return: argument 0 options: - cname: set arguments: - THTensor* self - THTensor* source - cname: setStorage arguments: - THTensor* self - CONSTANT NULL, 0, NULL, NULL - cname: setStorage before_call: THLongStoragePtr __storage_size(THLongStorage_newWithSize1(THStorage_(size)(LIBRARY_STATE arg_storage))); arguments: - THTensor* self - THStorage* storage - CONSTANT 0 - CONSTANT __storage_size.get() - CONSTANT NULL - cname: setStorage arguments: - THTensor* self - THStorage* sourceStorage - long storage_offset - THSize* size - arg: THStride* stride default: NULL ]] [[ name: THPTensor_(select) python_name: select cpu_half: True auto_gpu: False only_register: True ]] [[ name: THPTensor_(size) python_name: size cpu_half: True auto_gpu: False method_flags: METH_KEYWORDS only_register: True ]] [[ name: THPTensor_(stride) python_name: stride cpu_half: True auto_gpu: False method_flags: METH_KEYWORDS only_register: True ]] [[ name: fill_ cname: fill return: self arguments: - THTensor* self - real value ]] [[ name: isSameSizeAs python_name: is_same_size cpu_half: True auto_gpu: False return: bool arguments: - THTensor* self - THTensor* other ]] [[ name: isContiguous python_name: is_contiguous cpu_half: True auto_gpu: False return: bool arguments: - THTensor* self ]] [[ name: isSetTo python_name: is_set_to cpu_half: True auto_gpu: False return: bool arguments: - THTensor* self - THTensor* tensor ]] [[ name: maskedFill_ cname: maskedFill python_name: masked_fill_ return: self arguments: - arg: THTensor* self broadcast: mask inplace fallback types:Byte - THBoolTensor* mask - real value ]] [[ name: maskedCopy_ cname: maskedCopy python_name: masked_scatter_ return: self arguments: - arg: THTensor* self broadcast: mask inplace fallback types:Byte - THBoolTensor* mask - THTensor* source ]] [[ name: maskedSelect python_name: masked_select variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - arg: THTensor* self broadcast: mask fallback types:Byte - THBoolTensor* mask ]] [[ name: transpose variants: - method - function cname: newTranspose cpu_half: True auto_gpu: False return: THTensor* arguments: - THTensor* self - arg: long dim0 wrap_dim: self - arg: long dim1 wrap_dim: self ]] [[ name: transpose_ cname: transpose cpu_half: True auto_gpu: False return: self arguments: - THTensor* self - THTensor* self - arg: long dim0 wrap_dim: self - arg: long dim1 wrap_dim: self ]] [[ name: t variants: - method - function auto_gpu: False cname: newTranspose return: THTensor* before_call: | long ndim = arg_self->nDimension; THPUtils_assert(ndim == 2, \"t() expects a 2D tensor, but self is %ldD\", ndim); arguments: - THTensor* self - CONSTANT 0 - CONSTANT 1 ]] [[ name: t_ cname: transpose auto_gpu: False return: self before_call: | long ndim = arg_self->nDimension; THPUtils_assert(ndim == 2, \"t_() expects a 2D tensor, but self is %ldD\", ndim); arguments: - THTensor* self - THTensor* self - CONSTANT 0 - CONSTANT 1 ]] [[ name: squeeze cpu_half: True variants: - method - function return: argument 0 options: - arguments: - arg: THTensor* result output: True - THTensor* self - cname: squeeze1d arguments: - arg: THTensor* result output: True - THTensor* self - arg: long dim wrap_dim: self ]] [[ name: squeeze_ cpu_half: True return: self options: - cname: squeeze arguments: - THTensor* self - THTensor* self - cname: squeeze1d arguments: - THTensor* self - THTensor* self - arg: long dim wrap_dim: self ]] [[ name: unsqueeze variants: - method - function cpu_half: True auto_gpu: False return: argument 0 cname: unsqueeze1d arguments: - arg: THTensor* result output: True - THTensor* self - arg: long dim wrap_dim: self+1 ]] [[ name: unsqueeze_ cpu_half: True auto_gpu: False return: self cname: unsqueeze1d arguments: - THTensor* self - THTensor* self - arg: long dim wrap_dim: self+1 ]] [[ name: nonzero variants: - method - function return: argument 0 arguments: - arg: THIndexTensor* result output: True - THTensor* self ]] [[ name: contiguous cname: newContiguous return: THTensor* arguments: - THTensor* self ]] [[ name: clone cname: newClone return: THTensor* aten_sparse: True arguments: - THTensor* self ]] [[ name: view cname: newView auto_gpu: False return: THTensor* arguments: - THTensor* self - arg: THSize* size long_args: True ]] [[ name: expand cname: newExpand return: THTensor* arguments: - THTensor* self - arg: THSize* size long_args: True ]] [[ name: resizeAs_ python_name: resize_as_ cname: resizeAs return: self arguments: - THTensor* self - THTensor* the_template ]] [[ name: indexSelect python_name: index_select variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self - arg: long dim wrap_dim: self - THIndexTensor* index ]] [[ name: indexCopy_ python_name: index_copy_ cname: indexCopy return: argument 0 arguments: - THTensor* self - arg: long dim wrap_dim: self - THIndexTensor* index - THTensor* source ]] [[ name: indexAdd_ python_name: index_add_ cname: indexAdd return: argument 0 arguments: - THTensor* self - arg: long dim wrap_dim: self - THIndexTensor* index - THTensor* source ]] [[ name: indexFill_ python_name: index_fill_ cname: indexFill return: argument 0 arguments: - THTensor* self - arg: long dim wrap_dim: self - THIndexTensor* index - real value ]] [[ name: narrow cpu_half: True auto_gpu: False return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self - arg: long dimension wrap_dim: self - long start - long length ]] [[ name: unfold cpu_half: True auto_gpu: False return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self - arg: long dimension wrap_dim: self - long size - long step ]] [[ name: range variants: - function backends: - CPU - CUDA return: argument 0 before_arg_assign: | PyErr_WarnEx(PyExc_UserWarning, \"torch.range is deprecated in favor of torch.arange \" \"and will be removed in 0.3. Note that arange generates values in [start; end), \" \"not [start; end].\", 1); arguments: - arg: THTensor* result output: True - accreal start - accreal end - arg: accreal step default: 1 ]] [[ name: arange variants: - function backends: - CPU - CUDA return: argument 0 options: - arguments: - arg: THTensor* result output: True - accreal start - accreal end - accreal step - arguments: - arg: THTensor* result output: True - accreal start - accreal end - CONSTANT 1 ]] [[ name: scatter_ return: argument 0 options: - cname: scatter arguments: - THTensor* self - arg: long dim wrap_dim: self - THIndexTensor* index - THTensor* src - cname: scatterFill arguments: - THTensor* self - arg: long dim wrap_dim: self - THIndexTensor* index - real value ]] [[ name: scatter_add_ return: argument 0 options: - cname: scatterAdd arguments: - THTensor* self - arg: long dim wrap_dim: self - THIndexTensor* index - THTensor* src ]] [[ name: gather variants: - method - function return: argument 0 before_call: | THLongStoragePtr _size(THIndexTensor_(newSizeOf)(LIBRARY_STATE arg_index)); THTensor_(resize)(LIBRARY_STATE arg_result, _size, NULL); arguments: - arg: THTensor* result output: True resize: index - THTensor* self - arg: long dim wrap_dim: self - THIndexTensor* index ]] [[ name: THPTensor_stateless_(cat) python_name: cat method_flags: METH_KEYWORDS only_register: True variants: - function ]] [[ name: data_ptr defined_if: \"!IS_DISTRIBUTED\" with_gil: True auto_gpu: False return: void* cpu_half: True cname: data arguments: - THTensor* self ]] [[ name: equal variants: - method - function return: bool arguments: - THTensor* self - THTensor* other ]] [[ python_name: copy_ name: THPTensor_(copy_) cpu_half: True method_flags: METH_KEYWORDS only_register: True ]] [[ name: __and__ variants: - method - function return: argument 0 options: - cname: bitand arguments: - arg: THTensor* result output: True - THTensor* self - real value - cname: cbitand arguments: - arg: THTensor* result output: True - arg: THTensor* self broadcast: other fallback - THTensor* other ]] [[ name: __iand__ variants: - method - function return: argument 0 options: - cname: bitand arguments: - THTensor* self - THTensor* self - real value - cname: cbitand arguments: - THTensor* self - arg: THTensor* self broadcast: other inplace fallback - THTensor* other ]] [[ name: __or__ variants: - method - function return: argument 0 options: - cname: bitor arguments: - arg: THTensor* result output: True - THTensor* self - real value - cname: cbitor arguments: - arg: THTensor* result output: True - arg: THTensor* self broadcast: other fallback - THTensor* other ]] [[ name: __ior__ variants: - method - function return: argument 0 options: - cname: bitor arguments: - THTensor* self - THTensor* self - real value - cname: cbitor arguments: - THTensor* self - arg: THTensor* self broadcast: other inplace fallback - THTensor* other ]] [[ name: __xor__ variants: - method - function return: argument 0 options: - cname: bitxor arguments: - arg: THTensor* result output: True - THTensor* self - real value - cname: cbitxor arguments: - arg: THTensor* result output: True - arg: THTensor* self broadcast: other fallback - THTensor* other ]] [[ name: __ixor__ variants: - method - function return: argument 0 options: - cname: bitxor arguments: - THTensor* self - THTensor* self - real value - cname: cbitxor arguments: - THTensor* self - arg: THTensor* self broadcast: other inplace fallback - THTensor* other ]] [[ name: __lshift__ variants: - method - function return: argument 0 options: - cname: lshift arguments: - arg: THTensor* result output: True - THTensor* self - real value - cname: clshift arguments: - arg: THTensor* result output: True - arg: THTensor* self broadcast: other fallback - THTensor* other ]] [[ name: __ilshift__ variants: - method - function return: argument 0 options: - cname: lshift arguments: - THTensor* self - THTensor* self - real value - cname: clshift arguments: - THTensor* self - arg: THTensor* self broadcast: other inplace fallback - THTensor* other ]] [[ name: __rshift__ variants: - method - function return: argument 0 options: - cname: rshift arguments: - arg: THTensor* result output: True - THTensor* self - real value - cname: crshift arguments: - arg: THTensor* result output: True - arg: THTensor* self broadcast: other fallback - THTensor* other ]] [[ name: __irshift__ variants: - method - function return: argument 0 options: - cname: rshift arguments: - THTensor* self - THTensor* self - real value - cname: crshift arguments: - THTensor* self - arg: THTensor* self broadcast: other inplace fallback - THTensor* other ]] [[ name: THPTensor_(apply) python_name: apply_ defined_if: \"!IS_DISTRIBUTED\" backends: - CPU cpu_half: True only_register: True override_method_flags: METH_O ]] [[ name: THPTensor_(map) python_name: map_ defined_if: \"!IS_DISTRIBUTED\" backends: - CPU cpu_half: True only_register: True ]] [[ name: THPTensor_(map2) python_name: map2_ defined_if: \"!IS_DISTRIBUTED\" backends: - CPU cpu_half: True only_register: True ]] [[ name: lt variants: - method return: argument 0 options: - cname: ltValue arguments: - arg: THBoolTensor* result output: True - THTensor* self - real value - cname: ltTensor arguments: - arg: THBoolTensor* result output: True - arg: THTensor* self broadcast: other fallback - THTensor* other ]] [[ name: lt_ return: self options: - cname: ltValueT arguments: - THTensor* self - THTensor* self - real value - cname: ltTensorT arguments: - THTensor* self - arg: THTensor* self broadcast: other inplace fallback - arg: THTensor* other ]] [[ name: lt variants: - function return: argument 0 options: - cname: ltValue arguments: - arg: THBoolTensor* result output: True - THTensor* tensor - real value - cname: ltTensor arguments: - arg: THBoolTensor* result output: True - arg: THTensor* tensor broadcast: other fallback - THTensor* other - cname: ltValueT arguments: - arg: THTensor* result output: True - THTensor* tensor - real value - cname: ltTensorT arguments: - arg: THTensor* result output: True - arg: THTensor* tensor broadcast: other fallback - THTensor* other ]] [[ name: gt variants: - method return: argument 0 options: - cname: gtValue arguments: - arg: THBoolTensor* result output: True - THTensor* self - real value - cname: gtTensor arguments: - arg: THBoolTensor* result output: True - arg: THTensor* self broadcast: other fallback - THTensor* other ]] [[ name: gt_ return: self options: - cname: gtValueT arguments: - THTensor* self - THTensor* self - real value - cname: gtTensorT arguments: - THTensor* self - arg: THTensor* self broadcast: other inplace fallback - THTensor* other ]] [[ name: gt variants: - function return: argument 0 options: - cname: gtValue arguments: - arg: THBoolTensor* result output: True - THTensor* tensor - real value - cname: gtTensor arguments: - arg: THBoolTensor* result output: True - arg: THTensor* tensor broadcast: other fallback - THTensor* other - cname: gtValueT arguments: - arg: THTensor* result output: True - THTensor* tensor - real value - cname: gtTensorT arguments: - arg: THTensor* result output: True - arg: THTensor* tensor broadcast: other fallback - THTensor* other ]] [[ name: le variants: - method return: argument 0 options: - cname: leValue arguments: - arg: THBoolTensor* result output: True - THTensor* self - real value - cname: leTensor arguments: - arg: THBoolTensor* result output: True - arg: THTensor* self broadcast: other fallback - THTensor* other ]] [[ name: le_ return: self options: - cname: leValueT arguments: - THTensor* self - THTensor* self - real value - cname: leTensorT arguments: - THTensor* self - arg: THTensor* self broadcast: other inplace fallback - THTensor* other ]] [[ name: le variants: - function return: argument 0 options: - cname: leValue arguments: - arg: THBoolTensor* result output: True - THTensor* tensor - real value - cname: leTensor arguments: - arg: THBoolTensor* result output: True - arg: THTensor* tensor broadcast: other fallback - THTensor* other - cname: leValueT arguments: - arg: THTensor* result output: True - THTensor* tensor - real value - cname: leTensorT arguments: - arg: THTensor* result output: True - arg: THTensor* tensor broadcast: other fallback - THTensor* other ]] [[ name: ge variants: - method return: argument 0 options: - cname: geValue arguments: - arg: THBoolTensor* result output: True - THTensor* self - real value - cname: geTensor arguments: - arg: THBoolTensor* result output: True - arg: THTensor* self broadcast: other fallback - THTensor* other ]] [[ name: ge_ return: self options: - cname: geValueT arguments: - THTensor* self - THTensor* self - real value - cname: geTensorT arguments: - THTensor* self - arg: THTensor* self broadcast: other inplace fallback - THTensor* other ]] [[ name: ge variants: - function return: argument 0 options: - cname: geValue arguments: - arg: THBoolTensor* result output: True - THTensor* tensor - real value - cname: geTensor arguments: - arg: THBoolTensor* result output: True - arg: THTensor* tensor broadcast: other fallback - THTensor* other - cname: geValueT arguments: - arg: THTensor* result output: True - THTensor* tensor - real value - cname: geTensorT arguments: - arg: THTensor* result output: True - arg: THTensor* tensor broadcast: other fallback - THTensor* other ]] [[ name: eq variants: - method return: argument 0 options: - cname: eqValue arguments: - arg: THBoolTensor* result output: True - THTensor* self - real value - cname: eqTensor arguments: - arg: THBoolTensor* result output: True - arg: THTensor* self broadcast: other fallback - THTensor* other ]] [[ name: eq_ return: self options: - cname: eqValueT arguments: - THTensor* self - THTensor* self - real value - cname: eqTensorT arguments: - THTensor* self - arg: THTensor* self broadcast: other inplace fallback - THTensor* other ]] [[ name: eq variants: - function return: argument 0 options: - cname: eqValue arguments: - arg: THBoolTensor* result output: True - THTensor* tensor - real value - cname: eqTensor arguments: - arg: THBoolTensor* result output: True - arg: THTensor* tensor broadcast: other fallback - THTensor* other - cname: eqValueT arguments: - arg: THTensor* result output: True - THTensor* tensor - real value - cname: eqTensorT arguments: - arg: THTensor* result output: True - arg: THTensor* tensor broadcast: other fallback - THTensor* other ]] [[ name: ne variants: - method return: argument 0 options: - cname: neValue arguments: - arg: THBoolTensor* result output: True - THTensor* self - real value - cname: neTensor arguments: - arg: THBoolTensor* result output: True - arg: THTensor* self broadcast: other fallback - THTensor* other ]] [[ name: ne_ return: self options: - cname: neValueT arguments: - THTensor* self - THTensor* self - real value - cname: neTensorT arguments: - THTensor* self - arg: THTensor* self broadcast: other inplace fallback - THTensor* other ]] [[ name: ne variants: - function return: argument 0 options: - cname: neValue arguments: - arg: THBoolTensor* result output: True - THTensor* tensor - real value - cname: neTensor arguments: - arg: THBoolTensor* result output: True - arg: THTensor* tensor broadcast: other fallback - THTensor* other - cname: neValueT arguments: - arg: THTensor* result output: True - THTensor* tensor - real value - cname: neTensorT arguments: - arg: THTensor* result output: True - arg: THTensor* tensor broadcast: other fallback - THTensor* other ]] [[ name: min variants: - method - function options: - cname: minall return: real arguments: - THTensor* self - cname: cmin return: argument 0 arguments: - arg: THTensor* result output: True - arg: THTensor* self broadcast: other fallback - THTensor* other - cname: min return: argument 0,1 arguments: - arg: THTensor* min output: True - arg: THIndexTensor* min_indices output: True - THTensor* self - arg: long dim wrap_dim: self - bool keepdim - cname: min return: argument 0,1 before_call: maybeThrowBackCompatKeepdimWarn(\"min\"); arguments: - arg: THTensor* min output: True - arg: THIndexTensor* min_indices output: True - THTensor* self - arg: long dim wrap_dim: self - CONSTANT false ]] [[ name: max variants: - method - function options: - cname: maxall return: real arguments: - THTensor* self - cname: cmax return: argument 0 arguments: - arg: THTensor* result output: True - arg: THTensor* self broadcast: other fallback - THTensor* other - cname: max return: argument 0,1 arguments: - arg: THTensor* max output: True - arg: THIndexTensor* max_indices output: True - THTensor* self - arg: long dim wrap_dim: self - bool keepdim - cname: max return: argument 0,1 before_call: maybeThrowBackCompatKeepdimWarn(\"max\"); arguments: - arg: THTensor* max output: True - arg: THIndexTensor* max_indices output: True - THTensor* self - arg: long dim wrap_dim: self - CONSTANT false ]] [[ name: kthvalue backends: - CPU variants: - method - function return: argument 0,1 options: - before_call: long __last_dim = THTensor_(nDimension)(LIBRARY_STATE ((THPTensor*)$arg2)->cdata)-1; arguments: - arg: THTensor* values output: True - arg: THIndexTensor* indices output: True - THTensor* self - long k - CONSTANT __last_dim - bool keepdim - before_call: | long __last_dim = THTensor_(nDimension)(LIBRARY_STATE ((THPTensor*)$arg2)->cdata)-1; maybeThrowBackCompatKeepdimWarn(\"kthvalue\"); arguments: - arg: THTensor* values output: True - arg: THIndexTensor* indices output: True - THTensor* self - long k - CONSTANT __last_dim - CONSTANT false - arguments: - arg: THTensor* values output: True - arg: THIndexTensor* indices output: True - THTensor* self - long k - arg: long dim wrap_dim: self - bool keepdim - before_call: maybeThrowBackCompatKeepdimWarn(\"kthvalue\"); arguments: - arg: THTensor* values output: True - arg: THIndexTensor* indices output: True - THTensor* self - long k - arg: long dim wrap_dim: self - CONSTANT false ]] [[ name: mode variants: - method - function return: argument 0,1 options: - before_call: long __last_dim = THTensor_(nDimension)(LIBRARY_STATE ((THPTensor*)$arg2)->cdata)-1; arguments: - arg: THTensor* values output: True - arg: THIndexTensor* indices output: True - THTensor* self - CONSTANT __last_dim - bool keepdim - before_call: | long __last_dim = THTensor_(nDimension)(LIBRARY_STATE ((THPTensor*)$arg2)->cdata)-1; maybeThrowBackCompatKeepdimWarn(\"mode\"); arguments: - arg: THTensor* values output: True - arg: THIndexTensor* indices output: True - THTensor* self - CONSTANT __last_dim - CONSTANT false - arguments: - arg: THTensor* values output: True - arg: THIndexTensor* indices output: True - THTensor* self - arg: long dim wrap_dim: self - bool keepdim - before_call: maybeThrowBackCompatKeepdimWarn(\"mode\"); arguments: - arg: THTensor* values output: True - arg: THIndexTensor* indices output: True - THTensor* self - arg: long dim wrap_dim: self - CONSTANT false ]] [[ name: median variants: - method - function return: argument 0,1 options: - cname: medianall return: real arguments: - THTensor* self - cname: median before_call: long __last_dim = THTensor_(nDimension)(LIBRARY_STATE ((THPTensor*)$arg2)->cdata)-1; arguments: - arg: THTensor* values output: True - arg: THIndexTensor* indices output: True - THTensor* self - CONSTANT __last_dim - bool keepdim - before_call: maybeThrowBackCompatKeepdimWarn(\"median\"); arguments: - arg: THTensor* values output: True - arg: THIndexTensor* indices output: True - THTensor* self - arg: long dim wrap_dim: self - CONSTANT false - arguments: - arg: THTensor* values output: True - arg: THIndexTensor* indices output: True - THTensor* self - arg: long dim wrap_dim: self - bool keepdim ]] [[ name: sort variants: - method - function return: argument 0,1 options: - before_call: long __last_dim = THTensor_(nDimension)(LIBRARY_STATE ((THPTensor*)$arg2)->cdata)-1; arguments: - arg: THTensor* values output: True - arg: THIndexTensor* indices output: True - THTensor* self - CONSTANT __last_dim - arg: bool descending default: \"false\" kwarg_only: True - arguments: - arg: THTensor* values output: True - arg: THIndexTensor* indices output: True - THTensor* self - arg: long dim wrap_dim: self - arg: bool descending default: \"false\" ]] [[ name: topk variants: - method - function return: argument 0,1 options: - before_call: long __last_dim = THTensor_(nDimension)(LIBRARY_STATE ((THPTensor*)$arg2)->cdata)-1; arguments: - arg: THTensor* values output: True - arg: THIndexTensor* indices output: True - THTensor* self - long k - CONSTANT __last_dim - CONSTANT true - CONSTANT true - arguments: - arg: THTensor* values output: True - arg: THIndexTensor* indices output: True - THTensor* self - long k - arg: long dim wrap_dim: self - arg: bool largest default: \"true\" - arg: bool sorted default: \"true\" ]] [[ name: all types: - Byte backends: - CPU - CUDA cname: logicalall return: bool arguments: - THTensor* self ]] [[ name: any types: - Byte backends: - CPU - CUDA cname: logicalany return: bool arguments: - THTensor* self ]] [[ name: getDevice python_name: get_device backends: - CUDA return: long arguments: - THTensor* self ]] [[ name: THPTensor_(new) python_name: new method_flags: METH_KEYWORDS backends: - CUDA only_register: True ]] [[ name: THPTensor_(recordStream) python_name: record_stream override_method_flags: METH_O backends: - CUDA only_register: True ]] [[ name: abs return: argument 0 types: - floating_point - Long - Int - Short backends: - CPU - CUDA variants: - method - function arguments: - arg: THTensor* destination output: True - THTensor* self ]] [[ name: abs_ cname: abs return: self types: - floating_point - Long - Int - Short backends: - CPU - CUDA arguments: - THTensor* self - THTensor* self ]] [[ name: sigmoid_ types: - floating_point backends: - CPU - CUDA cname: sigmoid return: self arguments: - THTensor* self - THTensor* self ]] [[ name: sigmoid types: - floating_point backends: - CPU - CUDA cname: sigmoid variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: log_ types: - floating_point backends: - CPU - CUDA cname: log return: self arguments: - THTensor* self - THTensor* self ]] [[ name: log types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: log1p_ types: - floating_point backends: - CPU - CUDA cname: log1p return: self arguments: - THTensor* self - THTensor* self ]] [[ name: log1p types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: lgamma types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: lgamma_ types: - floating_point backends: - CPU - CUDA cname: lgamma return: self arguments: - THTensor* self - THTensor* self ]] [[ name: exp_ types: - floating_point backends: - CPU - CUDA cname: exp return: self arguments: - THTensor* self - THTensor* self ]] [[ name: exp types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: cos_ types: - floating_point backends: - CPU - CUDA cname: cos return: self arguments: - THTensor* self - THTensor* self ]] [[ name: cos types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: acos_ types: - floating_point backends: - CPU - CUDA cname: acos return: self arguments: - THTensor* self - THTensor* self ]] [[ name: acos types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: cosh_ types: - floating_point backends: - CPU - CUDA cname: cosh return: self arguments: - THTensor* self - THTensor* self ]] [[ name: cosh types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: sin_ types: - floating_point backends: - CPU - CUDA cname: sin return: self arguments: - THTensor* self - THTensor* self ]] [[ name: sin types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: asin_ types: - floating_point backends: - CPU - CUDA cname: asin return: self arguments: - THTensor* self - THTensor* self ]] [[ name: asin types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: sinh_ types: - floating_point backends: - CPU - CUDA cname: sinh return: self arguments: - THTensor* self - THTensor* self ]] [[ name: sinh types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: tan_ types: - floating_point backends: - CPU - CUDA cname: tan return: self arguments: - THTensor* self - THTensor* self ]] [[ name: tan types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: atan_ types: - floating_point backends: - CPU - CUDA cname: atan return: self arguments: - THTensor* self - THTensor* self ]] [[ name: atan types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: tanh_ types: - floating_point backends: - CPU - CUDA cname: tanh return: self arguments: - THTensor* self - THTensor* self ]] [[ name: tanh types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: sqrt_ types: - floating_point backends: - CPU - CUDA cname: sqrt return: self arguments: - THTensor* self - THTensor* self ]] [[ name: sqrt types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: rsqrt_ types: - floating_point backends: - CPU - CUDA cname: rsqrt return: self arguments: - THTensor* self - THTensor* self ]] [[ name: rsqrt types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: ceil_ types: - floating_point backends: - CPU - CUDA cname: ceil return: self arguments: - THTensor* self - THTensor* self ]] [[ name: ceil types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: floor_ types: - floating_point backends: - CPU - CUDA cname: floor return: self arguments: - THTensor* self - THTensor* self ]] [[ name: floor types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: round_ types: - floating_point backends: - CPU - CUDA cname: round return: self arguments: - THTensor* self - THTensor* self ]] [[ name: round types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: trunc_ types: - floating_point backends: - CPU - CUDA cname: trunc return: self arguments: - THTensor* self - THTensor* self ]] [[ name: trunc types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: frac_ types: - floating_point backends: - CPU - CUDA cname: frac return: self arguments: - THTensor* self - THTensor* self ]] [[ name: frac types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: mean types: - floating_point backends: - CPU - CUDA variants: - method - function options: - cname: meanall return: accreal arguments: - THTensor* self - cname: mean return: argument 0 arguments: - arg: THTensor* destination output: True - THTensor* self - arg: long dim wrap_dim: self - bool keepdim - cname: mean return: argument 0 before_call: maybeThrowBackCompatKeepdimWarn(\"mean\"); arguments: - arg: THTensor* destination output: True - THTensor* self - arg: long dim wrap_dim: self - CONSTANT false ]] [[ name: var types: - floating_point backends: - CPU - CUDA variants: - method - function options: - cname: varall return: accreal arguments: - THTensor* self - arg: bool unbiased if_true: 0 if_false: 1 default: 0 - cname: var return: argument 0 arguments: - arg: THTensor* destination output: True - THTensor* self - arg: long dim wrap_dim: self - arg: bool unbiased if_true: 0 if_false: 1 default: 0 - bool keepdim - cname: var return: argument 0 before_call: maybeThrowBackCompatKeepdimWarn(\"var\"); arguments: - arg: THTensor* destination output: True - THTensor* self - arg: long dim wrap_dim: self - arg: bool unbiased if_true: 0 if_false: 1 default: 0 - CONSTANT false ]] [[ name: std types: - floating_point backends: - CPU - CUDA variants: - method - function options: - cname: stdall return: accreal arguments: - THTensor* self - arg: bool unbiased if_true: 0 if_false: 1 default: 0 - cname: std return: argument 0 arguments: - arg: THTensor* destination output: True - THTensor* self - arg: long dim wrap_dim: self - arg: bool unbiased if_true: 0 if_false: 1 default: 0 - bool keepdim - cname: std return: argument 0 before_call: maybeThrowBackCompatKeepdimWarn(\"std\"); arguments: - arg: THTensor* destination output: True - THTensor* self - arg: long dim wrap_dim: self - arg: bool unbiased if_true: 0 if_false: 1 default: 0 - CONSTANT false ]] [[ name: norm types: - floating_point backends: - CPU - CUDA variants: - method - function options: - cname: normall return: accreal arguments: - THTensor* self - arg: real p default: AS_REAL(2) - cname: norm return: argument 0 arguments: - arg: THTensor* destination output: True - THTensor* self - real p - arg: long dim wrap_dim: self - bool keepdim - cname: norm return: argument 0 before_call: maybeThrowBackCompatKeepdimWarn(\"norm\"); arguments: - arg: THTensor* destination output: True - THTensor* self - real p - arg: long dim wrap_dim: self - CONSTANT false ]] [[ name: renorm types: - floating_point backends: - CPU - CUDA variants: - method - function options: - cname: renorm return: argument 0 arguments: - arg: THTensor* destination output: True - THTensor* self - real p - arg: long dim wrap_dim: self - real maxnorm ]] [[ name: renorm_ types: - floating_point backends: - CPU - CUDA options: - cname: renorm return: self arguments: - THTensor* self - THTensor* self - real p - arg: long dim wrap_dim: self - real maxnorm ]] [[ name: dist types: - floating_point backends: - CPU - CUDA variants: - method - function options: - cname: dist return: accreal arguments: - arg: THTensor* self broadcast: other fallback - THTensor* other - arg: real p default: AS_REAL(2) ]] [[ name: reciprocal types: - floating_point backends: - CPU - CUDA variants: - method - function options: - cname: cinv return: argument 0 arguments: - arg: THTensor* destination output: True - THTensor* self ]] [[ name: reciprocal_ types: - floating_point backends: - CPU - CUDA options: - cname: cinv return: self arguments: - THTensor* self - THTensor* self ]] [[ name: neg types: - floating_point - Long - Int - Short backends: - CPU - CUDA variants: - method - function options: - cname: neg return: argument 0 arguments: - arg: THTensor* destination output: True - THTensor* self ]] [[ name: neg_ types: - floating_point - Long - Int - Short backends: - CPU - CUDA options: - cname: neg return: self arguments: - THTensor* self - THTensor* self ]] [[ name: atan2 types: - Float - Double backends: - CPU backend_type_pairs: [[CPU,Float],[CPU,Double],[CUDA,Float]] variants: - method - function cname: atan2 return: argument 0 arguments: - arg: THTensor* destination output: True - arg: THTensor* self broadcast: other fallback - THTensor* other ]] [[ name: atan2_ backend_type_pairs: [[CPU,Float],[CPU,Double],[CUDA,Float]] cname: atan2 return: argument 0 arguments: - THTensor* self - arg: THTensor* self broadcast: other fallback inplace - THTensor* other ]] [[ name: pow types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 options: - cname: pow arguments: - arg: THTensor* destination output: True - THTensor* self - real exponent - cname: cpow arguments: - arg: THTensor* destination output: True - arg: THTensor* self broadcast: exponent fallback - THTensor* exponent - cname: tpow arguments: - arg: THTensor* destination output: True - real base - THTensor* self ]] [[ name: pow_ types: - floating_point backends: - CPU - CUDA return: argument 0 cname: pow options: - cname: pow arguments: - THTensor* self - THTensor* self - real exponent - cname: cpow arguments: - THTensor* self - arg: THTensor* self broadcast: exponent inplace fallback - THTensor* exponent ]] [[ name: lerp types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 cname: lerp arguments: - arg: THTensor* destination output: True - arg: THTensor* self broadcast: end fallback - THTensor* end - real weight ]] [[ name: lerp_ types: - floating_point backends: - CPU - CUDA return: self cname: lerp arguments: - THTensor* self - arg: THTensor* self broadcast: end fallback inplace - THTensor* end - real weight ]] [[ name: linspace types: - Float - Double backends: - CPU variants: - function return: argument 0 arguments: - arg: THTensor* result output: True - real start - real end - arg: long steps default: 100 ]] [[ name: logspace types: - Float - Double backends: - CPU variants: - function return: argument 0 arguments: - arg: THTensor* result output: True - real start - real end - arg: long steps default: 100 ]] [[ name: histc types: - Float - Double backends: - CPU variants: - method - function return: argument 0 options: - arguments: - arg: THTensor* destination output: True - THTensor* self - CONSTANT 100 - CONSTANT 0 - CONSTANT 0 - arguments: - arg: THTensor* destination output: True - THTensor* self - long bins - CONSTANT 0 - CONSTANT 0 - arguments: - arg: THTensor* destination output: True - THTensor* self - long bins - real min - CONSTANT 0 - arguments: - arg: THTensor* destination output: True - THTensor* self - long bins - real min - real max ]] [[ name: zero_ cname: zero return: self arguments: - THTensor* self ]] [[ name: sum variants: - method - function options: - cname: sumall return: accreal arguments: - THTensor* self - cname: sum return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self - arg: long dim wrap_dim: self - bool keepdim - cname: sum return: argument 0 before_call: maybeThrowBackCompatKeepdimWarn(\"sum\"); arguments: - arg: THTensor* result output: True - THTensor* self - arg: long dim wrap_dim: self - CONSTANT false ]] [[ name: prod variants: - method - function options: - cname: prodall return: accreal arguments: - THTensor* self - cname: prod return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self - arg: long dim wrap_dim: self - bool keepdim - cname: prod return: argument 0 before_call: maybeThrowBackCompatKeepdimWarn(\"prod\"); arguments: - arg: THTensor* result output: True - THTensor* self - arg: long dim wrap_dim: self - CONSTANT false ]] [[ name: cumsum variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self - arg: long dim wrap_dim: self ]] [[ name: cumprod variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self - arg: long dim wrap_dim: self ]] [[ name: sign variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self ]] [[ name: sign_ cname: sign return: self arguments: - THTensor* self - THTensor* self ]] [[ name: trace variants: - method - function return: accreal arguments: - THTensor* self ]] [[ name: add variants: - method - function return: argument 0 options: - cname: add arguments: - arg: THTensor* result output: True - THTensor* self - real value - cname: cadd aten_sparse: True arguments: - arg: THTensor* result output: True - arg: THTensor* self broadcast: other fallback - arg: real value default: AS_REAL(1) - THTensor* other - sparse: True cname: spcadd aten_dense_sparse: True arguments: - arg: THTensor* result output: True - THTensor* self - arg: real value default: AS_REAL(1) - THSTensor* other ]] [[ name: add_ return: argument 0 options: - cname: add arguments: - THTensor* self - THTensor* self - real value - cname: cadd aten_sparse: True arguments: - THTensor* self - arg: THTensor* self broadcast: other inplace fallback - arg: real value default: AS_REAL(1) - THTensor* other - sparse: True cname: spcadd aten_dense_sparse: True arguments: - THTensor* self - THTensor* self - arg: real value default: AS_REAL(1) - THSTensor* other ]] [[ name: sub variants: - method - function return: argument 0 options: - cname: sub arguments: - arg: THTensor* result output: True - THTensor* self - real value - cname: csub arguments: - arg: THTensor* result output: True - arg: THTensor* self broadcast: other fallback - arg: real value default: AS_REAL(1) - THTensor* other ]] [[ name: sub_ return: argument 0 options: - cname: sub arguments: - THTensor* self - THTensor* self - real value - cname: csub arguments: - THTensor* self - arg: THTensor* self broadcast: other inplace fallback - arg: real value default: AS_REAL(1) - THTensor* other ]] [[ name: mul variants: - method - function return: argument 0 options: - cname: mul arguments: - arg: THTensor* result output: True - THTensor* self - real value - cname: cmul arguments: - arg: THTensor* result output: True - arg: THTensor* self broadcast: other fallback - arg: THTensor* other ]] [[ name: mul_ return: argument 0 options: - cname: mul arguments: - THTensor* self - THTensor* self - real value - cname: cmul arguments: - THTensor* self - arg: THTensor* self broadcast: other inplace fallback - THTensor* other ]] [[ name: div variants: - method - function return: argument 0 options: - cname: div arguments: - arg: THTensor* result output: True - THTensor* self - real value - cname: cdiv arguments: - arg: THTensor* result output: True - arg: THTensor* self broadcast: other fallback - THTensor* other ]] [[ name: div_ return: argument 0 options: - cname: div arguments: - THTensor* self - THTensor* self - real value - cname: cdiv arguments: - THTensor* self - arg: THTensor* self broadcast: other inplace fallback - THTensor* other ]] [[ name: fmod return: argument 0 variants: - method - function options: - cname: fmod arguments: - arg: THTensor* result output: True - THTensor* self - real value - cname: cfmod arguments: - arg: THTensor* result output: True - arg: THTensor* self broadcast: other fallback - THTensor* other ]] [[ name: fmod_ return: argument 0 options: - cname: fmod arguments: - THTensor* self - THTensor* self - real value - cname: cfmod arguments: - THTensor* self - arg: THTensor* self broadcast: other inplace fallback - THTensor* other ]] [[ name: remainder return: argument 0 variants: - method - function options: - cname: remainder arguments: - arg: THTensor* result output: True - THTensor* self - real value - cname: cremainder arguments: - arg: THTensor* result output: True - arg: THTensor* self broadcast: other fallback - arg: THTensor* other ]] [[ name: remainder_ return: argument 0 options: - cname: remainder arguments: - THTensor* self - THTensor* self - real value - cname: cremainder arguments: - THTensor* self - arg: THTensor* self broadcast: other inplace fallback - THTensor* other ]] [[ name: clamp variants: - method - function return: argument 0 options: - cname: clamp arguments: - arg: THTensor* destination output: True - THTensor* self - real min - real max - cname: cmaxValue arguments: - arg: THTensor* result output: True - THTensor* self - arg: real min kwarg_only: True - cname: cminValue arguments: - arg: THTensor* result output: True - THTensor* self - arg: real max kwarg_only: True ]] [[ name: clamp_ cname: clamp return: self options: - cname: clamp arguments: - THTensor* self - THTensor* self - real min - real max - cname: cmaxValue arguments: - THTensor* self - THTensor* self - arg: real min kwarg_only: True - cname: cminValue arguments: - THTensor* self - THTensor* self - arg: real max kwarg_only: True ]] [[ name: dot backend_type_pairs: [[CUDA,floating_point], [CPU,all]] variants: - method - function return: accreal arguments: - arg: THTensor* self assert_ndim: 1 - arg: THTensor* tensor assert_ndim: 1 ]] [[ name: tril variants: - method - function return: argument 0 arguments: - arg: THTensor* destination output: True - THTensor* self - arg: long diagonal default: 0 ]] [[ name: tril_ cname: tril return: self arguments: - THTensor* self - THTensor* self - arg: long diagonal default: 0 ]] [[ name: triu variants: - method - function return: argument 0 arguments: - arg: THTensor* destination output: True - THTensor* self - arg: long diagonal default: 0 ]] [[ name: triu_ cname: triu return: self arguments: - THTensor* self - THTensor* self - arg: long diagonal default: 0 ]] [[ name: cross variants: - method - function return: argument 0 arguments: - arg: THTensor* destination output: True - THTensor* self - THTensor* other - arg: long dim default: -1 ]] [[ name: eye backends: - CPU - CUDA variants: - function return: argument 0 options: - arguments: - arg: THTensor* result output: True - long n - argument 1 - arguments: - arg: THTensor* result output: True - long n - long m ]] [[ name: diag variants: - method - function return: argument 0 options: - arguments: - arg: THTensor* result output: True - THTensor* self - arg: long diagonal default: 0 ]] [[ name: addmm variants: - method - function return: argument 0 options: - arguments: - arg: THTensor* result output: True - arg: real beta default: AS_REAL(1) - arg: THTensor* self broadcast: mat1,mat2 dims:mat1.dim0,mat2.dim1 - arg: real alpha default: AS_REAL(1) - THTensor* mat1 - THTensor* mat2 - cname: spaddmm sparse: yes arguments: - arg: THTensor* result output: True - arg: real beta default: AS_REAL(1) - THTensor* self - arg: real alpha default: AS_REAL(1) - THSTensor* mat1 - THTensor* mat2 ]] [[ name: addmm_ return: self options: - cname: addmm arguments: - THTensor* self - arg: real beta default: AS_REAL(1) - THTensor* self - arg: real alpha default: AS_REAL(1) - THTensor* mat1 - THTensor* mat2 - cname: spaddmm sparse: yes arguments: - arg: THTensor* self - arg: real beta default: AS_REAL(1) - THTensor* self - arg: real alpha default: AS_REAL(1) - THSTensor* mat1 - THTensor* mat2 ]] [[ name: addmv variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - arg: real beta default: AS_REAL(1) - arg: THTensor* self broadcast: mat,vec dims:mat.dim0 - arg: real alpha default: AS_REAL(1) - THTensor* mat - THTensor* vec ]] [[ name: addmv_ cname: addmv return: self arguments: - THTensor* self - arg: real beta default: AS_REAL(1) - THTensor* self - arg: real alpha default: AS_REAL(1) - THTensor* mat - THTensor* vec ]] [[ name: addr variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - arg: real beta default: AS_REAL(1) - arg: THTensor* self broadcast: vec1,vec2 dims:vec1.dim0,vec2.dim0 - arg: real alpha default: AS_REAL(1) - THTensor* vec1 - THTensor* vec2 ]] [[ name: addr_ cname: addr return: self arguments: - THTensor* self - arg: real beta default: AS_REAL(1) - THTensor* self - arg: real alpha default: AS_REAL(1) - THTensor* vec1 - THTensor* vec2 ]] [[ name: ger cname: addr variants: - method - function return: argument 0 before_call: | long s1 = THTensor_(size)(LIBRARY_STATE ((THPTensor*)$arg4)->cdata, 0); long s2 = THTensor_(size)(LIBRARY_STATE ((THPTensor*)$arg5)->cdata, 0); THTensor_(resize2d)(LIBRARY_STATE ((THPTensor*)$arg0)->cdata, s1, s2); arguments: - arg: THTensor* result output: True resize: [ [self,0], [vec2,0] ] - CONSTANT AS_REAL(0) - argument 0 - CONSTANT AS_REAL(1) - THTensor* self - THTensor* vec2 ]] [[ name: mv cname: addmv variants: - method - function return: argument 0 before_call: | long s = THTensor_(size)(LIBRARY_STATE ((THPTensor*)$arg4)->cdata, 0); THTensor_(resize1d)(LIBRARY_STATE ((THPTensor*)$arg0)->cdata, s); #if !IS_CUDA THTensor_(zero)(LIBRARY_STATE ((THPTensor*)$arg0)->cdata); #endif arguments: - arg: THTensor* result output: True resize: [ [self, 0] ] cpu_zero: True - CONSTANT AS_REAL(0) - argument 0 - CONSTANT AS_REAL(1) - THTensor* self - THTensor* vec ]] [[ name: mm variants: - method - function return: argument 0 options: - cname: addmm before_call: | long s1 = THTensor_(size)(LIBRARY_STATE ((THPTensor*)$arg4)->cdata, 0); long s2 = THTensor_(size)(LIBRARY_STATE ((THPTensor*)$arg5)->cdata, 1); THTensor_(resize2d)(LIBRARY_STATE ((THPTensor*)$arg0)->cdata, s1, s2); #if !IS_CUDA THTensor_(zero)(LIBRARY_STATE ((THPTensor*)$arg0)->cdata); #endif arguments: - arg: THTensor* result output: True resize: [ [self, 0], [mat2,1] ] cpu_zero: True - CONSTANT AS_REAL(0) - argument 0 - CONSTANT AS_REAL(1) - THTensor* self - THTensor* mat2 - cname: spaddmm sparse: True arguments: - arg: THTensor* result output: True - CONSTANT AS_REAL(0) - argument 0 - CONSTANT AS_REAL(1) - THSTensor* self - THTensor* mat2 ]] [[ name: bmm cname: baddbmm variants: - method - function return: argument 0 before_call: | long s1 = THTensor_(size)(LIBRARY_STATE ((THPTensor*)$arg4)->cdata, 0); long s2 = THTensor_(size)(LIBRARY_STATE ((THPTensor*)$arg4)->cdata, 1); long s3 = THTensor_(size)(LIBRARY_STATE ((THPTensor*)$arg5)->cdata, 2); THTensor_(resize3d)(LIBRARY_STATE ((THPTensor*)$arg0)->cdata, s1, s2, s3); #if !IS_CUDA THTensor_(zero)(LIBRARY_STATE ((THPTensor*)$arg0)->cdata); #endif arguments: - arg: THTensor* result output: True resize: [ [self,0], [self,1], [mat2,2] ] cpu_zero: True - CONSTANT AS_REAL(0) - argument 0 - CONSTANT AS_REAL(1) - THTensor* self - THTensor* mat2 ]] [[ name: addbmm variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - arg: real beta default: AS_REAL(1) - arg: THTensor* self broadcast: batch1,batch2 dims:batch1.dim1,batch2.dim2 - arg: real alpha default: AS_REAL(1) - THTensor* batch1 - THTensor* batch2 ]] [[ name: addbmm_ cname: addbmm return: self arguments: - THTensor* self - arg: real beta default: AS_REAL(1) - THTensor* self - arg: real alpha default: AS_REAL(1) - THTensor* batch1 - THTensor* batch2 ]] [[ name: baddbmm variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - arg: real beta default: AS_REAL(1) - arg: THTensor* self broadcast: batch1,batch2 dims:batch1.dim0,batch1.dim1,batch2.dim2 - arg: real alpha default: AS_REAL(1) - THTensor* batch1 - THTensor* batch2 ]] [[ name: baddbmm_ cname: baddbmm return: argument 0 arguments: - THTensor* self - arg: real beta default: AS_REAL(1) - THTensor* self - arg: real alpha default: AS_REAL(1) - THTensor* batch1 - THTensor* batch2 ]] [[ name: addcmul variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - arg: THTensor* self broadcast: tensor1,tensor2 fallback - arg: real value default: AS_REAL(1) - THTensor* tensor1 - THTensor* tensor2 ]] [[ name: addcmul_ options: - cname: addcmul return: argument 0 arguments: - THTensor* self - arg: THTensor* self broadcast: tensor1,tensor2 inplace fallback - arg: real value default: AS_REAL(1) - THTensor* tensor1 - THTensor* tensor2 - cname: spaddcmul defined_if: \"!IS_DISTRIBUTED\" return: argument 0 arguments: - THTensor* self - THTensor* self - arg: real value default: AS_REAL(1) - THSTensor* tensor1 - THSTensor* tensor2 ]] [[ name: addcdiv variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - arg: THTensor* self broadcast: tensor1,tensor2 fallback - arg: real value default: AS_REAL(1) - THTensor* tensor1 - THTensor* tensor2 ]] [[ name: addcdiv_ cname: addcdiv return: argument 0 arguments: - THTensor* self - arg: THTensor* self broadcast: tensor1,tensor2 inplace fallback - arg: real value default: AS_REAL(1) - THTensor* tensor1 - THTensor* tensor2 ]] [[ name: gesv types: - Float - Double backends: - CPU - CUDA variants: - method - function return: argument 0,1 arguments: - arg: THTensor* solution output: True - arg: THTensor* lu output: True - THTensor* self - THTensor* A ]] [[ name: gels types: - Float - Double backends: - CPU - CUDA variants: - method - function return: argument 0,1 arguments: - arg: THTensor* res1 output: True - arg: THTensor* res2 output: True - THTensor* self - THTensor* A ]] [[ name: trtrs types: - Float - Double backends: - CPU variants: - method - function return: argument 0,1 arguments: - arg: THTensor* res1 output: True - arg: THTensor* res2 output: True - THTensor* self - THTensor* A - arg: bool upper if_true: U if_false: L default: U - arg: bool transpose if_true: T if_false: N default: N - arg: bool unitriangular if_true: U if_false: N default: N ]] [[ name: symeig cname: syev types: - Float - Double backends: - CPU - CUDA variants: - method - function return: argument 0,1 arguments: - arg: THTensor* res1 output: True - arg: THTensor* res2 output: True - THTensor* self - arg: bool eigenvectors if_true: V if_false: N default: N - arg: bool upper if_true: U if_false: L default: U ]] [[ name: eig cname: geev types: - Float - Double backends: - CPU - CUDA variants: - method - function return: argument 0,1 arguments: - arg: THTensor* res1 output: True - arg: THTensor* res2 output: True - THTensor* self - arg: bool eigenvectors if_true: V if_false: N default: N ]] [[ name: svd cname: gesvd types: - Float - Double backends: - CPU - CUDA variants: - method - function return: argument 0,1,2 arguments: - arg: THTensor* res1 output: True - arg: THTensor* res2 output: True - arg: THTensor* res3 output: True - THTensor* self - arg: bool some if_true: S if_false: A default: S ]] [[ name: inverse cname: getri types: - Float - Double backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* output output: True - THTensor* self ]] [[ name: potrf types: - Float - Double backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* output output: True - THTensor* self - arg: bool upper if_true: U if_false: L default: U ]] [[ name: potrs types: - Float - Double backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self - THTensor* input2 - arg: bool upper if_true: U if_false: L default: U ]] [[ name: potri types: - Float - Double backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* output output: True - THTensor* self - arg: bool upper if_true: U if_false: L default: U ]] [[ name: pstrf types: - Float - Double backends: - CPU variants: - method - function return: argument 0,1 after_call: THIntTensor_sub(((THPIntTensor*)$arg1)->cdata, ((THPIntTensor*)$arg1)->cdata, 1); arguments: - arg: THTensor* res1 output: True - arg: THIntegerTensor* res2 output: True - THTensor* self - arg: bool upper if_true: U if_false: L default: U - arg: real tol default: -1 ]] [[ name: qr types: - Float - Double backends: - CPU - CUDA variants: - method - function return: argument 0,1 arguments: - arg: THTensor* res1 output: True - arg: THTensor* res2 output: True - THTensor* self ]] [[ name: geqrf types: - Float - Double backends: - CPU variants: - method - function return: argument 0,1 arguments: - arg: THTensor* res1 output: True - arg: THTensor* res2 output: True - THTensor* self ]] [[ name: orgqr types: - Float - Double backends: - CPU variants: - method - function return: argument 0,1 arguments: - arg: THTensor* result output: True - THTensor* self - THTensor* input2 ]] [[ name: ormqr types: - Float - Double backends: - CPU variants: - method - function return: argument 0,1 arguments: - arg: THTensor* result output: True - THTensor* self - THTensor* input2 - THTensor* input3 - arg: bool left if_true: L if_false: R default: L - arg: bool transpose if_true: T if_false: N default: N ]] [[ name: btrifact cname: btrifact types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0,1 arguments: - arg: THTensor* result output: True - arg: THIntegerTensor* pivots output: True - arg: THIntegerTensor* info kwarg_only: True default: NULL - arg: bool pivot kwarg_only: True default: \"true\" - THTensor* self ]] [[ name: btrisolve cname: btrisolve types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THTensor* result output: True - THTensor* self - THTensor* LU_data - THIntegerTensor* LU_pivots ]] [[ name: randperm defined_if: \"!IS_DISTRIBUTED\" backends: - CPU variants: - function return: argument 0 arguments: - arg: THTensor* result output: True - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - long n ]] [[ name: random_ defined_if: \"!IS_DISTRIBUTED\" backends: - CPU return: self options: - cname: random arguments: - THTensor* self - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - cname: cappedRandom arguments: - THTensor* self - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - long to - cname: clampedRandom arguments: - THTensor* self - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - long from - long to ]] [[ name: multinomial types: - floating_point backends: - CPU - CUDA variants: - method - function return: argument 0 arguments: - arg: THIndexTensor* result output: True - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - THTensor* self - long num_samples - arg: bool replacement default: \"false\" ]] [[ name: uniform_ types: - floating_point backends: - CPU - CUDA cname: uniform return: self arguments: - THTensor* self - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - arg: double from default: 0 - arg: double to default: 1 ]] [[ name: normal types: - floating_point backends: - CPU - CUDA return: argument 0 variants: - function options: - cname: normal_means arguments: - arg: THTensor* output output: True - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - THTensor* means - arg: double std default: 1 - cname: normal_stddevs arguments: - arg: THTensor* output output: True - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - arg: double mean default: 0 - THTensor* std - cname: normal_means_stddevs arguments: - arg: THTensor* output output: True - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - THTensor* means - THTensor* std ]] [[ name: normal_ types: - floating_point backends: - CPU - CUDA cname: normal return: self arguments: - THTensor* self - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - arg: double mean default: 0 - arg: double std default: 1 ]] [[ name: cauchy_ types: - floating_point backends: - CPU - CUDA cname: cauchy return: self arguments: - THTensor* self - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - arg: double median default: 0 - arg: double sigma default: 1 ]] [[ name: logNormal_ cname: logNormal python_name: log_normal_ types: - floating_point backends: - CPU - CUDA return: self arguments: - THTensor* self - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - arg: double mean default: 1 - arg: double std default: 2 ]] [[ name: exponential_ types: - floating_point backends: - CPU - CUDA cname: exponential return: self arguments: - THTensor* self - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - arg: double lambd default: 1 ]] [[ name: rand types: - floating_point backends: - CPU - CUDA variants: - function return: argument 0 arguments: - arg: THTensor* result output: True - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - arg: THSize* size long_args: True ]] [[ name: randn types: - floating_point backends: - CPU - CUDA variants: - function return: argument 0 arguments: - arg: THTensor* result output: True - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - arg: THSize* size long_args: True ]] [[ name: geometric_ defined_if: \"!IS_DISTRIBUTED\" backends: - CPU - CUDA cname: geometric return: self arguments: - THTensor* self - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - double p ]] [[ name: bernoulli types: - Float - Double backends: - CPU return: argument 0 variants: - method - function before_call: THTensor_(resizeAs)(LIBRARY_STATE ((THPTensor*)$arg0)->cdata, ((THPTensor*)$arg2)->cdata); cname: BERNOULLI_TENSOR arguments: - arg: THTensor* output output: True resize: self - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - THTensor* self ]] [[ name: bernoulli_ defined_if: \"!IS_DISTRIBUTED\" backends: - CPU - CUDA return: self options: - cname: bernoulli arguments: - THTensor* self - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - arg: double p default: 0.5 - cname: bernoulli_FloatTensor arguments: - THTensor* self - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - BackendFloatTensor* float_p - cname: bernoulli_DoubleTensor arguments: - THTensor* self - arg: THGenerator* generator default: THPDefaultGenerator->cdata kwarg_only: True - BackendDoubleTensor* float_p ]] </ins>", "pos": ["There are previous issues addressing this, but I want to write this up with a plan in a new issue. We should make differentiable instead of having a wrapper class. The Variable class adds conceptual overhead without sufficient benefit. It makes it harder for new users to adopt PyTorch by adding a second API which is similar but not exactly the same as Tensor. There are few prerequisites: on s need to be at least as fast as operations on () need scalar support at the Python level () The constructor will be deprecated and become a no-op. The accessor will be deprecated and return a view that has requiresgrad=False. The constructors will take an optional keyword argument There will be a number of internal changes, but we shouldn't break most existing code: performance reasons, we will no longer build up the graph of operations by default when unless the operation uses stochastic variables. We'll use a context manager or global variable to enable graph creation. new differentiable Tensor will need support all the operations currently on Tensor, including in-place and operations. -place operations on leaf Variables which requiregrad are not currently supported. In the new API, the tensor's value will be updated and it will remain a leaf node. (This is important for optimizers, which update leaf Variables in-place). There are probably a few other subtle differences between Tensor and Variable APIs. I'll update this issue as we address them. See also: 23/738 # Greg will be fixing the tests, Tensor vs. Variable requiresgrad 2/1837 # Just testbroadcast from TestTorch OK OK # error at top level OK OK OK OK OK OK OK OK 10/120 # missing sspaddmm, dsmm/mm, addmm 1/257 TestTorch.testbroadcast 2OK\nProperly handling in-place might not be as easy as it seems. Sometimes you actually need to move the Variable in the graph. Consider this case:\nOne of the cases of differences between variable and tensor is computation of var/std:\nIMO Variable gives user a clear difference between graph node and data. .data attribute and Variable constructor provide convenient and explicit ways to take control of the graph (or graphs).\nThis is done. (Still some clean-up to do)\nIs the conda PyTorch out of date? I'm getting the following error: .\nYes, we haven't released 0.4, which has the Variable Tensor merge, yet"], "neg": []}
{"query": "torch_version=version_str, is_debug_build=debug_mode_str, python_version='{}.{} ({}-bit runtime)'.format(sys.version_info[0], sys.version_info[1], sys.maxsize.bit_length() + 1), <ins> python_platform=get_python_platform(), </ins> is_cuda_available=cuda_available_str, cuda_compiled_version=cuda_version_str, cuda_runtime_version=get_running_cuda_version(run_lambda),", "pos": ["TensorPipe uses some advanced features of the Linux kernel. It does so on a \"best effort\" basis (i.e., use them if available, fall back on something else if not). This logic is unfortunately hard to test because it would require specific kernel versions on our CI machines. So it occasionally happens that we find out about issues when users report them. To help investigate these reports we've found it useful to know the kernel version and the glibc version used by the user (both when building and when running, if they're different). It would be great if these could be to the report produced by collect_env.\nCC"], "neg": []}
{"query": "self.assertEqual(v[mask].shape, (3, 7, 3)) self.assertEqual(v[mask], torch.stack([v[0], v[2], v[3]])) <ins> v = Variable(torch.Tensor([1])) self.assertEqual(v[v == 0], Variable(torch.Tensor())) </ins> def test_multiple_byte_mask(self): v = Variable(torch.randn(5, 7, 3)) # note: these broadcast together and are transposed to the first dim", "pos": ["In PyTorch master 0.4 , variables cannot be indexed. This used to work in PyTorch 0.2 :\nThis is breaking some of Pyro's inference algorithms on PyTorch master:"], "neg": []}
{"query": "lib_paths = list(filter(bool, [ os.getenv('CUDNN_LIB_DIR'), os.path.join(CUDA_HOME, 'lib'), <del> os.path.join(CUDA_HOME, 'lib64') </del> <ins> os.path.join(CUDA_HOME, 'lib64'), '/usr/lib/x86_64-linux-gnu/', </ins> ])) include_paths = list(filter(bool, [ os.getenv('CUDNN_INCLUDE_DIR'), os.path.join(CUDA_HOME, 'include'), <ins> '/usr/include/' </ins> ])) for path in lib_paths: if path is None or not os.path.exists(path):", "pos": ["Kind of similar to Sometimes our Eigen build fails this way: It doesn't seem to reliably repro. cc\nI got some sort of similar-ish error persistently on a test branch:\nNot that persistently; a PR stacked on the failing one succeeded. So there is something nondeterministic going on here.\nanother instance: CI on\nLooks like an NVCC bug. The workground is protecting eigen headers with Can we extract some basic code example and compile it against NVCC?\nAnother one in\nCan you try with 10.2?\nMingbo, can you put general information on the issue you are seeing here in this issue?\nlatest word on this is from mingbo: \"from what I can tell, it's a flaky one. might depend on which VS version. I hit the problem before and problem fixed (at least seem so) after upgrading VS 2019 latest version\" and \"the same error showed up once last week, and re-run the job finished without problem.\"\nHi. Did anyone have this issue persistently, ie that a rebuild did not \"fix\" it? We've had this issue in another (closed source) project. Exposing cuda to less includes (such as boost) seems to help but is not predictable.\nIt's never been persistent for us, but this is mostly from our Windows CI builds where we blow everything away and rebuild from scratch, which may help.\nYet another occurrence in the forum: It seems to be fixed by , which implies there may be a race condition in reading the include files or generating the intermediate source files.\nI was able to reproduce the issue with a more verbose output with And then I inspected the variables carefully and found out that the VC env is activated twice. We should really avoid that.\nYou are SO COOL!!!\nI tried to build from scratch several times with this PR and could not reproduce the issue anymore. Let's assume it's fixed.\nLooks like it was not resolved.\nAnother occurrence: This time it seems to claim the error is around the usage of .\nThis sounds like a good old multiple-instances-of-nvcc-write-to-the-same-temp-file One can invoke to see all temporary filenames it creates. I guess on windows they are more likely to collide with each other than on linux. One solution to the problem would be to invoke nvcc sequentially.", "Another is to tweak environment (there must be something similar to TMPDIR on Windows, mustn't it?)\nYes, they are called TMP and TEMP.\nIs there any example of the latter method on Unix systems?\nI've written a utility . Hopefully it can be used to resolve this problem if that's the case.\nNew occurrence after the previous fix:\nseems related to\nIf this is the case, maybe we could apply to fix that.\nunfortunately, this introduces another thrust problem (e.g. ): D:/programs64/cuda101/includethrust/detail/complex/catrig.h(607): error: no instance of function template \"thrust::atanh\" matches the argument list argument types are: (double) with problematic include chain -thrust/complex.h -thrust/detail/complex/catrig.h (has cplusplus conditions) -thrust/detail/complex/c99math.h where \"using ::atanh\" is omitted for MSVC. Patching c99math.h (or even thrustdetailconfigcppdialect.h without using /Zc:_cplusplus) seems to work for me, but I'm not sure what's the best way to fix this within pytorch tree and whether c++ dialect is the only problem here.\nThanks for the investigation. BTW, could you please share with us how did you patch c99math.h?\nDisregard that, still got same unreproducible C2993 error with /Zc:_cplusplus. Earlier, I had \"build\" directory reproducibly failing (at ), and hardcoding c++17 in thrustdetailconfigcppdialect.h fixed that. But looks like it is just a symptom of some inconsistence in environment/flags (across nvcc stages maybe?) just \"using ::atanh\"\nYeah, maybe. I am considering to add a wrapper to retry the build a few times.\nI can now stably reproduce a related error by passing in .\nBy commenting out some part of the code, I find the code that causes the previous error. I don't know why it is executed in the device code.\nWell, I figured out the reason. It seems that and are implicitly marked as and . This behavior can be disabled by adding .\nNice work!\nSeems to be a nvcc bug\nThanks for reminding. I have commented there.\nThis is indeed an nvcc bug.", "There is no known workaround at the moment, but the next release of the CUDA toolkit will contain a fix. Ref thrust/thrust.\nMet this problem:\nJust retry a few times.\nThe CUDA TK 11.0 RC is now available: This issue has been fixed in that version of nvcc.\nThanks for the info. Actually, our release binaries rely on the cudatoolkit package in anaconda cloud. So we won't be able to publish the binaries with CUDA 11 now. I'm just curious whether that is maintained by Nvidia? Another question, any plan to backport the changes to exisiting CUDA versions (e.g. cuda 10.1, cuda 10.2)?\nI don't have that information unfortunately -- it'd be best to contact the maintainer of that package and see when they plan to update. I'm not aware of any plans to backport this fix.\nI have also obtained this error with CUDA 10.1/cuDNN 7.6.4 on Win Server 2019.\nthis is a cuda bug. Please see above discussion and upgrade to cuda 11.\nI fixed the error by installing sscache for use with cuda 10.1 and cuda 10.2, as done here.\nit's an intermittent error and the probability of occurrence varies depending on your system. So it's not really fixed, but you may have found a way to reduce the occurrence in your system, which may be sufficient in your case :)\nAnother instance, this time inside protobuf: Running with I found the syntax error is introduced by which is the cuda compiler front end. operates on the output from the compiler's pre-processor and that output barely even changed to trigger the build error, only line number metadata and some extra whitespace was changed in the pre-processed output. Here is the diff of the two inputs, one of which failed and the other didn't: And the corresponding change in the output: This failure reproduced reliably on the CI machine. Running nvcc manually multiple times, it always fails or passes given the same inputs. I think it's a deterministic failure, but is annoyingly sensitive to any changes in the input file whatsoever.\nclosing due to age\nWait, do we think this is fixed?", "It's not clear to me it is...\nI think this was related to CUDA 10.2 builds on windows which we don't actually support anymore so it is probably safe to keep closed."], "neg": []}
{"query": "if self.is_cuda: raise TypeError(\"cannot pin '{0}' only CPU memory can be pinned\" .format(self.type())) <del> storage = self.storage() </del> <ins> storage = self.contiguous().storage() </ins> if storage is None: storage = (self.storage_type())() return type(self)().set_(storage.pin_memory()).view_as(self)", "pos": ["Hello, I was experimenting with a custom which receives a batch of different sized integer sequences and pads them to form 2D tensors. I noticed a weird issue for which I created the following MWE: It's probably because of the different memory ordering on Python's side and CUDA's allocator. Tested on K40 and GTX1080 with driver 375.39, CUDA8.\nPin memory is missing a call inside. Should be an easy fix\nfixed via"], "neg": []}
{"query": "build_libs_cmd = ['bash', 'torch/lib/build_libs.sh'] my_env = os.environ.copy() my_env[\"PYTORCH_PYTHON\"] = sys.executable <ins> if WITH_SYSTEM_NCCL: my_env[\"NCCL_ROOT_DIR\"] = NCCL_ROOT_DIR </ins> if WITH_CUDA: build_libs_cmd += ['--with-cuda'] if subprocess.call(build_libs_cmd + libs, env=my_env) != 0:", "pos": ["Currently, you can have a libnccl on your system, and pytorch setup favors that instead of compiling one itself from source. The code that takes care of this is: However, this breaks gloo build because gloo gets a path to NCCL via:\nI got the error during the installation: Is it related to this problem?\nthat's a dupe of\nno, that's a consequence of , now you have to do before compiling. I think you don't want to do recursive submodule update, because that will pull in gtest and another copy of nccl.\nAlternatively we can pass new flags to gloo to prevent it from using its own NCCL\nThanks works for me.\nYou can prevent it from using its own nccl, but its cmake still has to be able to find some nccl.\nI think should be sufficient. I'm testing.\nfixes it when we dont have an external NCCL and pytorch compiles it's own NCCL?\nI don't see how NCCLEXTERNAL fixes anything, looks like in gloo NCCLEXTERNAL is used just to set dependency if gloo is compiling its own nccl (which we don't want in any case).\nThe only thing I tested when I was that the vendored nccl was not being built (even if its submodule was checked out); I tested this by adding an to one of the headers. I'll test with the other configurations.\nActually, there are a few bugs here. First, even though GLOOFLAGS looks like it's setting NCCLROOTDIR: it actually isn't because the gloo invocation is which expands to but only is passed to cmake. Suppose you fix this. nccl is still not going to be picked up, because INSTALLDIR is PyTorch's local install dir, but if you disabled nccl builds, there will never actually be a nccl here. So cmake is just going to do it's standard search to find a nccl somewhere. Inversely, if you ever built nccl, and then turn it off, we will keep picking up that locally built version until you delete it. So we shouldn't do that. Suppose you fix this.", "Then I think gloo will (continue to) correctly pick up system nccl and be built with it, and the final thing to fix is that build of PyTorch csrc doesn't know how to search for nccl headers somewhere. This is probably related to . I wonder, maybe we shouldn't do any autodetection at all, and force the user to pass us the include and library paths of nccl if they want to build with system.\nFYI, I committed a change to gloo to stop bundling NCCL itself. It will now happily pick up whatever NCCL you give it through NCCLROOTDIR or by trying to find it in the standard systems paths. See facebookincubator/gloo You do have to set to make Gloo try and find it and compile in support.\nSo to fix this, you should be able to not specify anything aside from if you want to use system NCCL, and add if you have some custom location (either a PyTorch bundled version or a custom install altogether)."], "neg": []}
{"query": "} void THTensor_(sparseMask)(THSTensor *r_, THTensor *t, THSTensor *mask) { <ins> THArgCheck(mask->coalesced, 2, \"mask is uncoalesced\"); </ins> THSTensor_(resizeAs)(r_, mask); if (mask->nnz == 0) { THSTensor_(zero)(r_);", "pos": ["and I talked about the intended semantics of sparsemask when given a sparse tensor as an argument, and we came to the following conclusions: with an uncoalesced sparse tensor does not make sense, full stop. Unlike values, where your update might be linear and work correctly on an uncoalesced sparse tensor, there isn't any good reason why you want to pluck out values from the dense tensor multiple times, it is basically always a bug. the (eventual) documentation for sparsemask, we will recommend calling coalesce before doing any operations involving sparsemask (or nonlinear updates to indices/values). should add a test to sparsemask which checks if the flag of the input tensor is true, and raise an error if it is not. This is a cheap safety check which will prevent people from forgetting to call coalesce (as in (2)). There is a performance consequence to calling coalesce every iteration of training. If the gradient is producing coalesced vectors, we can alleviate this problem by adding a preprocessing step to coalesce() which tests if the input is already coalesced, and bugs out in that case. However, I think this is out of scope for this particular ticket for now."], "neg": []}
{"query": "auto num_threads = leaked->get_thread_count(); // NOLINTNEXTLINE(modernize-make-unique) threadpool.reset(new PThreadPool(num_threads)); <del> TORCH_WARN(\"Leaking Caffe2 thread-pool after fork.\"); </del> } } return threadpool.get();", "pos": ["When numworker 0 (in DataLoader), I call setnum_threads function and get some warnings. Execute the following script: got log: there should be no any warnings. PyTorch Version (e.g., 1.0): torch1.9 OS (e.g., Linux): Centos7 How you installed PyTorch (, , source): pip Build command you used (if compiling from source): no Python version: 3.8 CUDA/cuDNN version: 11.1/ GPU models and configuration: TeslaV100 Any other relevant information:\nDuplicate of .\nclosing as dupplicate"], "neg": []}
{"query": "from numbers import Integral import warnings import math <ins> from operator import mul from functools import reduce </ins> import torch from torch._C import _infer_size", "pos": ["Right now it makes everything .\nTo be clear, for a tensor, the problem is only when all of N, H, and W are 1. (So BatchNorm2d on batch size of 1 is OK as long as you don't have a 1x1 image). What's the desired behavior? The only reasonable behavior I can think of is: Raise an exception when the dimensions of which you are normalizing are one or Output zero (+ optional affine transform) I'm not sure the outputing zero is a good idea. I can't think of a case where that's what you want."], "neg": []}
{"query": "#include <TH/TH.h> #include <libshm.h> #include \"THP.h\" <ins> #include \"byte_order.h\" </ins> #include \"generic/Storage.cpp\" #include <TH/THGenerateAllTypes.h> #include \"generic/StorageCopy.cpp\" #include <TH/THGenerateAllTypes.h>", "pos": ["<!-- A clear and concise description of what the bug is. --The target name is used by a test executable in mkldnn. Steps to reproduce the behavior: BUILDTEST=1 build <!-- If you have a code sample, error messages, stack traces, please provide it here as well --<!-- A clear and concise description of what you expected to happen. --CMake configure passes. Please copy and paste the output from our (or fill out the checklist below manually). You can get the script and run it with: PyTorch Version (e.g., 1.0): v1.5.0-rc2 OS (e.g., Linux): Windows How you installed PyTorch (, , source): source Build command you used (if compiling from source): python build Python version: 3.x CUDA/cuDNN version: 9.2 / 10.1 / 10.2 / NA GPU models and configuration: Tesla M60 Any other relevant information: <!-- Add any other context about the problem here. --We should not build tests for mkldnn.\nI have encountered the same error, how do I resolve it?\nIt is actually mkldnn that is out of sync. So running the following script should help.\nThanks a lot for this I posted this on the PyTorch Discussions but to no avail, and now I've finally understood the reason of this error. Thanks a lot. Surprisingly, for me the build works out if I build from the master repo, but changing to a different version always fails."], "neg": []}
{"query": "upstream=\"$1\" pr=\"$2\" git diff --name-only \"$upstream\" \"$pr\" <del> git diff --name-only \"$upstream\" \"$pr\" | grep -Eq '^(CMakeLists.txt|Makefile|.gitmodules|.jenkins/caffe2|binaries|caffe|caffe2|cmake|conda|docker|docs/caffe2|modules|scripts|third_party)' </del> <ins> # For safety, unconditionally trigger for any changes. #git diff --name-only \"$upstream\" \"$pr\" | grep -Eq '^(CMakeLists.txt|Makefile|.gitmodules|.jenkins/caffe2|binaries|caffe|caffe2|cmake|conda|docker|docs/caffe2|modules|scripts|third_party)' </ins>", "pos": ["This is most obvious by observing binary sizes, for example on pypi: (6.0 MB ) (33.8 MB ) Also, from the build log it's clear that 3.8 for some reason build with flag: CI cc\nWhich is a good old", "Einsum currently modifies variables in-place (without an explicit indication that it does so), which prevents pytorch from automatically backpropagating. Results in the following runtime error: Demonstration and discussion can also be found . PyTorch or Caffe2: PyTorch How you installed PyTorch (conda, pip, source): pip OS: PyTorch version: 0.4.0 Python version: 3.6.1\nOops. Thank you for reporting and the minimal example. I'll see to get that fixed.\nSaw the same issue over here. Thank you !\nIs cloning the tensor before passing to einsum a valid workaround?\nYes, but a somewhat expensive one. If you have the ability to recompile, you could also apply the PR fixing this. Somehow it seems to be stuck in the review queue because it hit an unrelated CI problem...\nOkay so how to use it then?\nJust get a recent enough master and it will work."], "neg": []}
{"query": "void evaluate_function(FunctionTask& task); ReadyQueue& ready_queue(int device); void start_threads(); <del> virtual void thread_main(std::shared_ptr<ReadyQueue> queue); </del> <ins> virtual void thread_main(std::shared_ptr<ReadyQueue> queue, int device); </ins> virtual void thread_on_exception(FunctionTask& task, std::exception& e); std::once_flag start_threads_flag;", "pos": ["Consider following code, and run it with multi gpus (e.g. 4): It will output: That is, the default device would always be zero even in DataParallel. My pytorch version is 0.1.122. I think this is not the desired behavior. It would cause some troubles. I tried to insert my own cuda kernel into backward to calculate the gradients, it become very slow, and I fixed it by (grad.getdevice()). Anyway, I think currentdevice in forward and backward should be the same, or could anyone explain to me why they are different?\nthis seems like something we should fix: ?\n<!-- drci-comment-start --:pagefacingup: Preview :pagefacingup: Preview :question: Need help or want to give feedback on the CI? Visit our Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. <!-- drci-comment-end --"], "neg": []}
{"query": "GCC version: {gcc_version} Clang version: {clang_version} CMake version: {cmake_version} <ins> Libc version: {libc_version} </ins> Python version: {python_version} <ins> Python platform: {python_platform} </ins> Is CUDA available: {is_cuda_available} CUDA runtime version: {cuda_runtime_version} GPU models and configuration: {nvidia_gpu_models}", "pos": ["TensorPipe uses some advanced features of the Linux kernel. It does so on a \"best effort\" basis (i.e., use them if available, fall back on something else if not). This logic is unfortunately hard to test because it would require specific kernel versions on our CI machines. So it occasionally happens that we find out about issues when users report them. To help investigate these reports we've found it useful to know the kernel version and the glibc version used by the user (both when building and when running, if they're different). It would be great if these could be to the report produced by collect_env.\nCC"], "neg": []}
{"query": "DESTINATION \"${TH_INSTALL_BIN_SUBDIR}\") ENDMACRO(Install_Required_Library libname) <del> IF (BLAS_FOUND AND BLAS_INSTALL_LIBRARIES) IF (BLAS_goto2_LIBRARY) Install_Required_Library(${BLAS_goto2_LIBRARY}) Install_Required_Library(\"${libpath}/libgfortran\") Install_Required_Library(\"${libpath}/libquadmath\") Install_Required_Library(\"${libpath}/libgcc\") </del> <ins> FIND_PACKAGE(BLAS) IF(BLAS_FOUND) SET(USE_BLAS 1) IF(BLAS_INFO STREQUAL \"mkl\") ADD_DEFINITIONS(-DTH_BLAS_MKL) </ins> ENDIF() <del> IF (BLAS_openblas_LIBRARY) Install_Required_Library(${BLAS_openblas_LIBRARY}) Install_Required_Library(\"${libpath}/libquadmath\") Install_Required_Library(\"${libpath}/libgfortran\") Install_Required_Library(\"${libpath}/libgcc\") ENDIF() ENDIF() </del> <ins> ENDIF(BLAS_FOUND) </ins> FIND_PACKAGE(LAPACK) IF(LAPACK_FOUND)", "pos": ["currently compiled master does not use any blas for any ops and uses our lame fallback C code. I'm looking into it. Basically USE_BLAS is not being set in"], "neg": []}
{"query": "else: self.gradInput = self.gradInput if not wasTable else input.clone() for i, module in enumerate(self.modules): <del> currentGradInput = getattr(module, method)(input, gradOutput[i], scale) </del> <ins> if method == 'updateGradInput': currentGradInput = module.updateGradInput(input, gradOutput[i]) elif method == 'backward': currentGradInput = module.backward(input, gradOutput[i], scale) </ins> if i == 0: self.gradInput.resize_as_(currentGradInput).copy_(currentGradInput) else:", "pos": ["Subclassing torch.Tensor with fails due to a on the latest nightly builds, whereas it previously worked without issue. Run this snippet of code: And you'll recieve an error message like this: Changing the to results in the exact same error message. There should be no error message as it should be working like it was previously in Torch 1.8.1 and below. I think that the issue started on Torch version: , and I can confirm that it still exists as of , . The issue started on the nightly builds at least 2 days ago, if that helps resolve things. It's been causing some of Captum's CircleCI tests to fail, and the tests show the first occurrence of the error on . cc\nThis is due to either or\nI think the problem is I'm not supposed to delete because that means the super() call doesn't work. Instead, I should just verify that the passed in class is a subclass, then this should work."], "neg": []}
{"query": "from contextlib import contextmanager import torch <ins> import torch.cuda </ins> import torch.distributed as dist from common import TestCase", "pos": ["is causing a crash when we try to build pytorch. Crash looks like this: (all times in US/West) Our 7AM, 8AM builds and earlier were all fine but 9AM, 10AM, 11AM builds all crashed with that same traceback. The new import was in which I think went in between 8AM and 9AM today See command below: See here if needed for : <!-- A clear and concise description of what you expected to happen. --Please copy and paste the output from our (or fill out the checklist below manually). You can get the script and run it with: PyTorch Version (e.g., 1.0): building from master OS (e.g., Linux): linux How you installed PyTorch (, , source): source Build command you used (if compiling from source): see above under 'reproduce' Python version: 3.6\nmaybe easier to read for build command used:\nplease try in the CircleCi environment?\nin our build script seems to have fixed this. Thanks !"], "neg": []}
{"query": "[13, 14, 15, 16], [17, 18, 19, 20], ]) <del> exp_v = self.ValueTensor([7, 14, 14, 3, 20]) </del> <ins> exp_v = self.ValueTensor([7, 14, 3, 20]) </ins> res = dense._sparse_mask(x) expected = self.SparseTensor(i, exp_v, torch.Size([5, 4])) self.assertEqual(res, expected)", "pos": ["and I talked about the intended semantics of sparsemask when given a sparse tensor as an argument, and we came to the following conclusions: with an uncoalesced sparse tensor does not make sense, full stop. Unlike values, where your update might be linear and work correctly on an uncoalesced sparse tensor, there isn't any good reason why you want to pluck out values from the dense tensor multiple times, it is basically always a bug. the (eventual) documentation for sparsemask, we will recommend calling coalesce before doing any operations involving sparsemask (or nonlinear updates to indices/values). should add a test to sparsemask which checks if the flag of the input tensor is true, and raise an error if it is not. This is a cheap safety check which will prevent people from forgetting to call coalesce (as in (2)). There is a performance consequence to calling coalesce every iteration of training. If the gradient is producing coalesced vectors, we can alleviate this problem by adding a preprocessing step to coalesce() which tests if the input is already coalesced, and bugs out in that case. However, I think this is out of scope for this particular ticket for now."], "neg": []}
{"query": "return new_var; } <ins> const char* ERR_BACKWARD_TWICE = \"Trying to backward through the graph a second time, but the buffers have \" \"already been freed. Specify retain_variables=True when calling backward \" \"the first time.\"; </ins> }} // namespace torch::autograd", "pos": ["Usually calling backwards twice gives this message: But conv2d just segfaults. $ pt A B Segmentation fault (core dumped)"], "neg": []}
{"query": "} // Replace with nonzeros auto nonzero = index.nonzero(); <del> for (int64_t j = 0; j < nonzero.size(1); j++) { result.emplace_back(nonzero.select(1, j)); </del> <ins> auto is_empty = nonzero.numel() == 0; for (int64_t j = 0; j < index.dim(); j++) { if (is_empty) { // We can't call select on an empty tensor so we just create an empty // tensor. result.emplace_back(nonzero.type().tensor()); } else { result.emplace_back(nonzero.select(1, j)); } </ins> } } else { result.emplace_back(index);", "pos": ["In PyTorch master 0.4 , variables cannot be indexed. This used to work in PyTorch 0.2 :\nThis is breaking some of Pyro's inference algorithms on PyTorch master:"], "neg": []}
{"query": "idx = Variable(torch.LongTensor()) self.assertEqual(x[idx].numel(), 0) <ins> # empty assignment should have no effect but not throw an exception y = x.clone() y[idx] = -1 self.assertEqual(x, y) mask = torch.zeros(4, 3).byte() y[mask] = -1 self.assertEqual(x, y) </ins> def test_basic_advanced_combined(self): # From the NumPy indexing example x = Variable(torch.arange(0, 12).view(4, 3))", "pos": ["odd difference between cuda Tensor and Variable: encountered in torch.distributions.Categorical, built yesterday from master\nrelated to hmm, here's the issue is for variables, so probably not (immediately) related\ndifferent code path but similar issue\nThis is a regression from the latest 0.3 binaries, which don't have this error."], "neg": []}
{"query": "assert_strict_equal(zc, z) <del> def index_variable(num_indices, max_indices): index = torch.randperm(max_indices)[:num_indices].long() </del> <ins> def index_variable(shape, max_indices): if not isinstance(shape, tuple): shape = (shape,) index = torch.rand(*shape).mul_(max_indices).floor_().long() return Variable(index, requires_grad=False) def gather_variable(shape, index_dim, max_indices): assert len(shape) == 2 assert index_dim < 2 batch_dim = 1 - index_dim index = torch.LongTensor(*shape) for i in range(shape[index_dim]): index.select(index_dim, i).copy_( torch.randperm(max_indices)[:shape[batch_dim]]) </ins> return Variable(index, requires_grad=False)", "pos": ["<!-- A clear and concise description of what content in is an issue. If this has to do with the general website, please file an issue at instead. If this has to do with , please file an issue at --In the getting started section section the following quote appears: I guess this is from previous versions where CUDA support was not built-in and this can be removed, correct? cc\nThis seems to be easy fixable docs issue\nIf gets merged this is fixed.\nThanks for the PR Just merged it."], "neg": []}
{"query": "result = x[rows[:, None], columns] self.assertEqual(result.data.tolist(), [[0, 2], [9, 11]]) <ins> def test_empty_index(self): x = Variable(torch.arange(0, 12).view(4, 3)) idx = Variable(torch.LongTensor()) self.assertEqual(x[idx].numel(), 0) </ins> def test_basic_advanced_combined(self): # From the NumPy indexing example x = Variable(torch.arange(0, 12).view(4, 3))", "pos": ["In PyTorch master 0.4 , variables cannot be indexed. This used to work in PyTorch 0.2 :\nThis is breaking some of Pyro's inference algorithms on PyTorch master:"], "neg": []}
{"query": "def forward(ctx, input, grid, padding_mode='zeros'): ctx.save_for_backward(input, grid) <ins> if input.device != grid.device: raise RuntimeError((\"input (device {}) and grid (device {}) must be on the same device\" + \"for grid_sampler\").format(input.device, grid.device)) </ins> if padding_mode == 'zeros': ctx.padding_mode = MODE_ZEROS elif padding_mode == 'border':", "pos": ["yields irrelevant errors when passing wrong arguments First one: Second one: PyTorch or Caffe2: pytorch 0.4 How you installed PyTorch (conda, pip, source): pip Build command you used (if compiling from source): OS: ubuntu16.04 PyTorch version: 0.4 Python version: 3.6 CUDA/cuDNN version: GPU models and configuration: 1080 GCC version (if compiling from source): CMake version: Versions of any other relevant libraries:\nThank you for the report! I'll move them to the new check for same GPU (which includes reasonable CPU vs. GPU reporting).\nIn PyTorch master, the error message has already improved for the second case:\nI am getting exactly the first error reported: and I don't understand what is the problem. Can someone help on this?\nif you use PyTorch v0.4.1, you should get a better error message helping you on what's wrong.\nThanks for your reply I found out that the problem was the hidden state in my recurrent net, I did not use the .cuda()"], "neg": []}
{"query": "IF(MSVC) # we want to respect the standard, and we are bored of those **** . ADD_DEFINITIONS(-D_CRT_SECURE_NO_DEPRECATE=1) <del> LIST(APPEND CUDA_NVCC_FLAGS \"-Xcompiler /wd4819 -Xcompiler /wd4503\") </del> <ins> LIST(APPEND CUDA_NVCC_FLAGS \"-Xcompiler /wd4819 -Xcompiler /wd4503 -Xcompiler /wd4190 -Xcompiler /wd4244 -Xcompiler /wd4251 -Xcompiler /wd4275 -Xcompiler /wd4522\") </ins> ADD_DEFINITIONS(-DTH_EXPORTS) IF (NOT NO_CUDA) ADD_DEFINITIONS(-DTHC_EXPORTS)", "pos": ["That's pretty big! In contrast, an OS X build log is only 1.3M. Is there a reason they have to be this big? CC\nMy guess is that it's because of warnings. I remember having to get through thousands of them when looking for the error that failed the build.\nYes, the major part of them are warnings. And some of them is not fixable. We can hide them using some flags on the compilers.\nAnd this build configuration actually contains both the CPU and GPU build, which may also be a possible reason.\nThe reduced log size after PR is about 1.2M."], "neg": []}
{"query": "} if (k == 3) { <del> if (should_compute_output(0)) { grad_input = view3d(*grad_input); </del> <ins> if (grad_input) { grad_input = view3d(*grad_input); } if (grad_weight) { grad_weight = view3d(*grad_weight); </ins> } <del> grad_weight = view3d(*grad_weight); </del> } auto outputs =  as_tensor_list(std::move(grad_input),", "pos": ["I have a network defined as: --. In this particular case, I want to keep the two conv layer weights fixed, and only adapt the middle layer weights. In the code below I have three ways of defining the conv layers parameters, one with buffers, one with parameters that have , and one with regular parameters (which doesn't keep the weights fixed, but I'm using it as a reference). If I use the first two definitions (commented out in the code below), when I call it results in a kernel crash. If I use the third definition, it works fine albeit without maintaining fixed weights. In order to apply the linear layer I reshape the output and inputs, and I believe the crash stems from somewhere there (on a side note: Is this the right way to apply a linear layer on a time sequence?)\nLooks like the problem is with . If I reimplement this network using 2D convolutions with unit widths it works fine."], "neg": []}
{"query": "# static library only NANOPB_STATIC_LIB = os.path.join(lib_path, 'libprotobuf-nanopb.a') <del> if WITH_NCCL and (subprocess.call('ldconfig -p | grep libnccl >/dev/null', shell=True) == 0 or subprocess.call('/sbin/ldconfig -p | grep libnccl >/dev/null', shell=True) == 0): SYSTEM_NCCL = True </del> main_compile_args = ['-D_THP_CORE'] main_libraries = ['shm'] main_link_args = [TH_LIB, THS_LIB, THNN_LIB, ATEN_LIB, NANOPB_STATIC_LIB]", "pos": ["Currently, you can have a libnccl on your system, and pytorch setup favors that instead of compiling one itself from source. The code that takes care of this is: However, this breaks gloo build because gloo gets a path to NCCL via:\nI got the error during the installation: Is it related to this problem?\nthat's a dupe of\nno, that's a consequence of , now you have to do before compiling. I think you don't want to do recursive submodule update, because that will pull in gtest and another copy of nccl.\nAlternatively we can pass new flags to gloo to prevent it from using its own NCCL\nThanks works for me.\nYou can prevent it from using its own nccl, but its cmake still has to be able to find some nccl.\nI think should be sufficient. I'm testing.\nfixes it when we dont have an external NCCL and pytorch compiles it's own NCCL?\nI don't see how NCCLEXTERNAL fixes anything, looks like in gloo NCCLEXTERNAL is used just to set dependency if gloo is compiling its own nccl (which we don't want in any case).\nThe only thing I tested when I was that the vendored nccl was not being built (even if its submodule was checked out); I tested this by adding an to one of the headers. I'll test with the other configurations.\nActually, there are a few bugs here. First, even though GLOOFLAGS looks like it's setting NCCLROOTDIR: it actually isn't because the gloo invocation is which expands to but only is passed to cmake. Suppose you fix this. nccl is still not going to be picked up, because INSTALLDIR is PyTorch's local install dir, but if you disabled nccl builds, there will never actually be a nccl here. So cmake is just going to do it's standard search to find a nccl somewhere. Inversely, if you ever built nccl, and then turn it off, we will keep picking up that locally built version until you delete it. So we shouldn't do that. Suppose you fix this.", "Then I think gloo will (continue to) correctly pick up system nccl and be built with it, and the final thing to fix is that build of PyTorch csrc doesn't know how to search for nccl headers somewhere. This is probably related to . I wonder, maybe we shouldn't do any autodetection at all, and force the user to pass us the include and library paths of nccl if they want to build with system.\nFYI, I committed a change to gloo to stop bundling NCCL itself. It will now happily pick up whatever NCCL you give it through NCCLROOTDIR or by trying to find it in the standard systems paths. See facebookincubator/gloo You do have to set to make Gloo try and find it and compile in support.\nSo to fix this, you should be able to not specify anything aside from if you want to use system NCCL, and add if you have some custom location (either a PyTorch bundled version or a custom install altogether)."], "neg": []}
{"query": "CUDA_INC ?= $(CUDA_HOME)/include NVCC ?= $(CUDA_HOME)/bin/nvcc <del> NVCC_GENCODE ?= -gencode=arch=compute_35,code=sm_35  </del> <ins> NVCC_GENCODE ?= -gencode=arch=compute_30,code=sm_30  -gencode=arch=compute_35,code=sm_35  </ins> -gencode=arch=compute_50,code=sm_50  -gencode=arch=compute_52,code=sm_52  -gencode=arch=compute_52,code=compute_52", "pos": ["I'm running PyTorch on a system with 4 Nvidia K10 GPUs and can't get the Multi-GPU code to work. I attempted to modify the first neural net in the 60-minute tutorial to run on all my GPUs, by adding this line to the cell defining the network: If I then attempt to use the net as follows , Jupyter waits a few seconds and reports the kernel has died. Following along in I see all GPUs encountering an at the same moment. Am I using the API wrong, could there be a configuration issue in my system, or is this a bug in the DataParallel implementation? I'm using CUDA 8 and PyTorch 0.1.6. Edit: my Nvidia driver version is 367.48 I also forgot to add, executing the above code on CUDA without parallelization (i.e. ) works fine.\nIs there a stack-trace that you see appear on the command-line of the Jupyter process? Is this a cloud provider box? (I ask to see if I can reproduce it myself)\nThis is the only command-line output Jupyter gives related to the error: I'm running this on an in-house compute node.\nthanks. this is helpful, I'll look further. does nccl have cuda code that we need to build for specific archs? K10 is 3.0 and i build 3.0 code for thc and thcunn\nYes, that's the problem: The README says \"Kepler or newer\". does NCCL work on K10s?\nIt should (once you add the necessary compilation flags)."], "neg": []}
{"query": "'clang_version', 'cmake_version', 'os', <ins> 'libc_version', </ins> 'python_version', <ins> 'python_platform', </ins> 'is_cuda_available', 'cuda_runtime_version', 'nvidia_driver_version',", "pos": ["TensorPipe uses some advanced features of the Linux kernel. It does so on a \"best effort\" basis (i.e., use them if available, fall back on something else if not). This logic is unfortunately hard to test because it would require specific kernel versions on our CI machines. So it occasionally happens that we find out about issues when users report them. To help investigate these reports we've found it useful to know the kernel version and the glibc version used by the user (both when building and when running, if they're different). It would be great if these could be to the report produced by collect_env.\nCC"], "neg": []}
{"query": "namespace torch { namespace autograd { <ins> struct FunctionTask { BackwardTask* base; std::shared_ptr<Function> fn; GradBuffer grad; FunctionTask(BackwardTask* base, std::shared_ptr<Function> fn, GradBuffer grad) : base(base) , fn(fn) , grad(std::move(grad)) {} }; </ins> struct ReadyQueue { std::deque<FunctionTask> queue; std::condition_variable not_empty;", "pos": ["Version include <stringin some cpp files: define order in The pre-defined struct is not passed by clang in macOS, so we should reorder them.\nSimple fix"], "neg": []}
{"query": "from .module import Module from .utils import _quadruple, _ntuple <del> from .._functions.padding import ConstantPad2d as F_ConstantPad2d </del> <ins> from .. import functional as F </ins> # TODO: grad_output size asserts in THNN class ReflectionPad2d(Module): <ins> r\"\"\"Pads the input tensor using the reflection of the input boundary. Args: padding (int, tuple): the size of the padding. If is int, uses the same padding in all boundaries. If a 4-tuple, uses (paddingLeft, paddingRight, paddingTop, paddingBottom) Shape: - Input: :math:`(N, C, H_{in}, W_{in})` - Output: :math:`(N, C, H_{out}, W_{out})` where :math:`H_{out} = H_{in} + paddingTop + paddingBottom` :math:`W_{out} = W_{in} + paddingLeft + paddingRight` Examples:: >>> m = nn.ReflectionPad2d(3) >>> input = autograd.Variable(torch.randn(16, 3, 320, 480)) >>> output = m(input) >>> # using different paddings >>> m = nn.ReflectionPad2d((3, 3, 6, 6)) >>> output = m(input) \"\"\" </ins> def __init__(self, padding): super(ReflectionPad2d, self).__init__() self.padding = _quadruple(padding) def forward(self, input): <del> return self._backend.ReflectionPad2d(*self.padding)(input) </del> <ins> return F.pad(input, self.padding, 'reflect') </ins> def __repr__(self): return self.__class__.__name__ + ' ' + str(self.padding) class ReplicationPad2d(Module): <ins> r\"\"\"Pads the input tensor using replication of the input boundary. Args: padding (int, tuple): the size of the padding. If is int, uses the same padding in all boundaries. If a 4-tuple, uses (paddingLeft, paddingRight, paddingTop, paddingBottom) Shape: - Input: :math:`(N, C, H_{in}, W_{in})` - Output: :math:`(N, C, H_{out}, W_{out})` where :math:`H_{out} = H_{in} + paddingTop + paddingBottom` :math:`W_{out} = W_{in} + paddingLeft + paddingRight` Examples:: >>> m = nn.ReplicationPad2d(3) >>> input = autograd.Variable(torch.randn(16, 3, 320, 480)) >>> output = m(input) >>> # using different paddings >>> m = nn.ReplicationPad2d((3, 3, 6, 6)) >>> output = m(input) \"\"\" </ins> def __init__(self, padding): super(ReplicationPad2d, self).__init__() self.padding = _quadruple(padding) def forward(self, input): <del> return self._backend.ReplicationPad2d(*self.padding)(input) </del> <ins> return F.pad(input, self.padding, 'replicate') </ins> def __repr__(self): return self.__class__.__name__ + ' ' + str(self.padding) class ReplicationPad3d(Module): <ins> r\"\"\"Pads the input tensor using replication of the input boundary. Args: padding (int, tuple): the size of the padding. If is int, uses the same padding in all boundaries. If a 6-tuple, uses (paddingLeft, paddingRight, paddingTop, paddingBottom, paddingFront, paddingBack) Shape: - Input: :math:`(N, C, D_{in}, H_{in}, W_{in})` - Output: :math:`(N, C, D_{out}, H_{out}, W_{out})` where :math:`D_{out} = D_{in} + paddingFront + paddingBack` :math:`H_{out} = H_{in} + paddingTop + paddingBottom` :math:`W_{out} = W_{in} + paddingLeft + paddingRight` Examples:: >>> m = nn.ReplicationPad3d(3) >>> input = autograd.Variable(torch.randn(16, 3, 8, 320, 480)) >>> output = m(input) >>> # using different paddings >>> m = nn.ReplicationPad3d((3, 3, 6, 6, 1, 1)) >>> output = m(input) \"\"\" </ins> def __init__(self, padding): super(ReplicationPad3d, self).__init__() self.padding = _ntuple(6)(padding) def forward(self, input): <del> return self._backend.ReplicationPad3d(*self.padding)(input) </del> <ins> return F.pad(input, self.padding, 'replicate') </ins> def __repr__(self): return self.__class__.__name__ + ' ' + str(self.padding) class ZeroPad2d(Module): <ins> r\"\"\"Pads the input tensor boundaries with zero. Args: padding (int, tuple): the size of the padding. If is int, uses the same padding in all boundaries. If a 4-tuple, uses (paddingLeft, paddingRight, paddingTop, paddingBottom) Shape: - Input: :math:`(N, C, H_{in}, W_{in})` - Output: :math:`(N, C, H_{out}, W_{out})` where :math:`H_{out} = H_{in} + paddingTop + paddingBottom` :math:`W_{out} = W_{in} + paddingLeft + paddingRight` Examples:: >>> m = nn.ZeroPad2d(3) >>> input = autograd.Variable(torch.randn(16, 3, 320, 480)) >>> output = m(input) >>> # using different paddings >>> m = nn.ZeroPad2d((3, 3, 6, 6)) >>> output = m(input) \"\"\" </ins> def __init__(self, padding): super(ZeroPad2d, self).__init__() self.padding = _quadruple(padding) def forward(self, input): <del> return F_ConstantPad2d(pad=self.padding, value=0)(input) </del> <ins> return F.pad(input, self.padding, 'constant', 0) </ins> def __repr__(self): return self.__class__.__name__ + ' ' + str(self.padding) class ConstantPad2d(Module): <ins> r\"\"\"Pads the input tensor boundaries with a constant value. Args: padding (int, tuple): the size of the padding. If is int, uses the same padding in all boundaries. If a 4-tuple, uses (paddingLeft, paddingRight, paddingTop, paddingBottom) Shape: - Input: :math:`(N, C, H_{in}, W_{in})` - Output: :math:`(N, C, H_{out}, W_{out})` where :math:`H_{out} = H_{in} + paddingTop + paddingBottom` :math:`W_{out} = W_{in} + paddingLeft + paddingRight` Examples:: >>> m = nn.ConstantPad2d(3, 3.5) >>> input = autograd.Variable(torch.randn(16, 3, 320, 480)) >>> output = m(input) >>> # using different paddings >>> m = nn.ConstantPad2d((3, 3, 6, 6), 3.5) >>> output = m(input) \"\"\" </ins> def __init__(self, padding, value): super(ConstantPad2d, self).__init__()", "pos": ["After merging cuda 12.4 for nightly. I observe following failures: libtorch-cuda124-shared-with-deps-cxx11-abi-build: Build Official Docker Images 124: We need to resolve these before continuing adding CUDA 12.4 CI workflows 0 cc\nattempt 3 worked for libtorch build issue:\nFor Docker image, related discussion:\nI will instance a new bug for the docker job failure.", "Is there a way to use reflection padding instead of zero padding with nn.Conv2d ? I see there is a but the option is not in nn.Conv2d. Thanks,\nare in , but they're not exposed in the docs.\nThanks!"], "neg": []}
{"query": "if fn.batch_first and not is_input_packed: input = input.transpose(0, 1) <del> if (not is_input_packed and input.dim() != 3) or (is_input_packed and input.dim() != 2): raise RuntimeError( 'input must have 3 dimensions, got {}'.format(input.dim())) if fn.input_size != input.size(-1): raise RuntimeError('input.size(-1) must be equal to input_size. Expected {}, got {}'.format( fn.input_size, input.size(-1) )) </del> if fn.dropout != 0 and cudnn.version() < 5103: raise RuntimeError('dropout supported only in cudnn v5.1 and above')", "pos": ["I run RNN sucessfully with the code. and I run embedding example. Then, got RuntimeError. I set to use nothing seems wrong. And the error info seems not the dim problem? The error is so strange.\nWhat exactly do you mean by \"embedding example\"? Could you provide a code snippet that causes the RuntimeError?\nTo better reproduce the error, I use the official code example of The origin runs well: Due to use , I have to make batch first. And then , it raises error.\nThank you for the example! The say that setting means that the first dim of input is the batch, ie, input has size . however is unchanged, so the following should work: I do agree the error message could be better, though"], "neg": []}
{"query": "// Kernel for fast unfold+copy // (borrowed from Caffe: https://github.com/BVLC/caffe/blob/master/src/caffe/layers/conv_layer.cu) template <typename Dtype> <ins> __launch_bounds__(CUDA_NUM_THREADS) </ins> __global__ void im2col_kernel(const int n, const Dtype* data_im, const int height, const int width, const int ksize_h, const int ksize_w,", "pos": ["I'm trying to run a variant of ERFNet on an NVIDIA TX-2 running Jetpack 3.2 (CUDA 9.0 and CuDNN 7). I get the following error: RuntimeError: cuda runtime error (7) : too many resources requested for launch at ../../../pytorch/torch/lib/THCUNN/im2col.h:120 This error indicates the size of the model + overhead is too large for the GPU? But this is roughly a 700mb model trying to perform inference on a single 640x512 grayscale image on a GPU that has roughly 6.5Gb of free space. I even tried training a new model on images at half that resolution and get the same error. Any tips/feedback is appreciated. PyTorch or Caffe2: PyTorch How you installed PyTorch (conda, pip, source): Source Build command you used (if compiling from source): Followed jetson-reinforcement github OS: Ubuntu 16.04 on Nvidia TX2 PyTorch version: 0.3.0 Python version: 3.6 CUDA/cuDNN version: CUDA 9.0 , CUDNN 7 GPU models and configuration: Tegra X2 GCC version (if compiling from source): 5.4 CMake version: 3.5.1 Versions of any other relevant libraries:\nwe should fix the launch parameters. I presume we cant use as many threads per block on the TX2 as we use on desktop GPUs. can you tell us what the limits of TX2 GPUs are for the fix to im2col\nThreads per block and maximum blocks in the grid are actually the same for TX2 as they are for desktop GPUs , only number of registers is smaller. do you know what might be causing this?\nas said, the number of threads per block is constant across different NVIDIA GPUs, the Tegra series included. Is it possible that it was compiled to require more registers than is available to the TX2, and maybe the kernel invocation of im2col requires some sort of launch_bounds() command?\nTX2 is arm64 platform. I presume pytorch is compiled from source. In this case, there isn't a chance that it was compiled to requre more registers than available on TX2 I think. The build log will be helpful to see.\nthanks for the quick response.", "I will try dig up the build log tomorrow and post it here.\nif there are no launch bounds, it is in fact possible that a kernel is compiled to request more registers than available. At the compile time, compiler does not know how many threads you'll want to launch with, so, potentially, it can use too many registers per thread to later satisfy runtime requirements, and, e.g. trying to launch with 512 or 1024 threads could fail (not even a single block can be put on an SM), whereas launching with 256 would succeed.\nis there a way where we can audit that GPU constraints on our server setting (i.e. not actually sitting and compiling TX2 PyTorch)\nAdding launch_bounds with the max number of threads the kernel is going to be launched with will cause compiler not to overuse registers. We had to do it e.g. for interp kernels when cuda 9 started using more registers\nAhh, that was as I suspected Thank you for clearing this up. do you guys think a fix will be released in the near future (week?). Thanks for all the help\nWorks now, thanks!\nhello, I met the same error on tx2, complied pytorch 0.3.0.. But the error is this: 'cuda runtime error(7): too many resources requested for launch at '. The file is different with '', is that I need add this line 'launch_bounds(1024)' to every function in file ''? Thank you!\nHello I am also facing a similar issue: Here's my setup: How you installed PyTorch (conda, pip, source): Source Build command you used (if compiling from source): Followed jetson-reinforcement github OS: Ubuntu 16.04 on Nvidia TX2 PyTorch version: 0.5.0a0+ (torchvision 0.2.1) Python version: 3.5 CUDA/cuDNN version: CUDA 9.0 , CUDNN 7 GPU models and configuration: Tegra X2 GCC version (if compiling from source): 5.4 CMake version: 3.12.2 Source code: Full log: Do you know what it could be?", "I know that '' without launch_bounds(1024) leads to this error, but I don't know how to fix it... Tried this solution , but still not working (after compilation)\nMy issue was solved by following\nHi, I have the same problem. Can this problem solved by gitting clone the new pytorch, than changing CUDANUMTHREADS =256 in the below two file and compiling it? !\nthat appears to fix some of the functions, but not perhaps all - not sure. For more info, see:"], "neg": []}
{"query": "- THTensor* output - std::vector<int> pad - std::vector<int> stride <del> - std::vector<int> dilation </del> <ins> - std::vector<int> dilation </ins> - int groups - bool benchmark ]]", "pos": ["We should check that the version from cudnn.h matches the version from the loaded shared library. Really weird things happen if they don't match."], "neg": []}
{"query": "dim_args_idx = test[4] if len(test) == 5 else [] <ins> skipTestIf = test[5] if len(test) == 6 else [] </ins> for dim_perm in product([-1, 1], repeat=len(dim_args_idx)): test_name = basic_test_name new_args = [arg * dim_perm[dim_args_idx.index(i)] if i in dim_args_idx else arg for i, arg in enumerate(args)]", "pos": ["Command to install the latest nightly binary: installs : However, PyTorch itself reports : I guess this version tag might be defined during the build process from e.g. and then forwarded from , but also source builds should use based on the tag. Unsure, if this is just a reporting issue or if we are building the nightly binaries with and executing them with , so creating this issue to dig a bit further into it. CC cc\nThis is really weird, I've put up measure in place to validate against it... Will investigate today (and moving to pytorch repo)\ncan you please check that 2.2.0 reports (and is build) against the correct version of nccl library?\nConfirmed for nightlies : Same issue for RC:\nLooking at the build log Looks like we are compiling with: USESYSTEMNCCL\nOn Windows I see following exception: We compiling with should we return None rather then exception ?\nIn this case we are indeed using 2.18.1: Do you know why we are using the system-wide NCCL installation instead of the one? Should we delete the system NCCL and use by default?\nLooks like its set here: Looks like it was by: Let me propose the PR to unset this flag : USESYSTEM_NCCL\nI guess we are using the system NCCL installation to be able to during the runtime. If we would use the module I assume we would link NCCL into the PyTorch binaries. If so, we should make sure to update the NCCL version whenever is updated. This could break our dependency on the NCCL pip package.\nYes. Also when we update this nccl here:\nWe used following path  What would be good path for nccl 2.19.3 and cuda 12.1 ? As per this issue: Will be using 12.0 ncck 2.19.3\nInstalling system NCCL into the nightly build docker container feels great, but we need to figure out, how to keep it in sync with PyTorch. I wonder if one can, for example, just check nccl version from this requirements file and build from source?\nyes probably we should build and upload nccl for our supported cuda versions, same way as we build docker containers, by following details here:\nVerified with final rc"], "neg": []}
{"query": "grad_input = torch.cross(other, grad_output, self.dim) grad_other = torch.cross(grad_output, input, self.dim) return grad_input, grad_other <ins> class Inverse(Function): @staticmethod def forward(ctx, input): inverse = torch.inverse(input) ctx.save_for_backward(inverse) return inverse @staticmethod def backward(ctx, grad_output): inverse, = ctx.saved_variables return -torch.mm(inverse.t(), torch.mm(grad_output, inverse.t())) </ins>", "pos": ["Command to install the latest nightly binary: installs : However, PyTorch itself reports : I guess this version tag might be defined during the build process from e.g. and then forwarded from , but also source builds should use based on the tag. Unsure, if this is just a reporting issue or if we are building the nightly binaries with and executing them with , so creating this issue to dig a bit further into it. CC cc\nThis is really weird, I've put up measure in place to validate against it... Will investigate today (and moving to pytorch repo)\ncan you please check that 2.2.0 reports (and is build) against the correct version of nccl library?\nConfirmed for nightlies : Same issue for RC:\nLooking at the build log Looks like we are compiling with: USESYSTEMNCCL\nOn Windows I see following exception: We compiling with should we return None rather then exception ?\nIn this case we are indeed using 2.18.1: Do you know why we are using the system-wide NCCL installation instead of the one? Should we delete the system NCCL and use by default?\nLooks like its set here: Looks like it was by: Let me propose the PR to unset this flag : USESYSTEM_NCCL\nI guess we are using the system NCCL installation to be able to during the runtime. If we would use the module I assume we would link NCCL into the PyTorch binaries. If so, we should make sure to update the NCCL version whenever is updated. This could break our dependency on the NCCL pip package.\nYes. Also when we update this nccl here:\nWe used following path  What would be good path for nccl 2.19.3 and cuda 12.1 ? As per this issue: Will be using 12.0 ncck 2.19.3\nInstalling system NCCL into the nightly build docker container feels great, but we need to figure out, how to keep it in sync with PyTorch. I wonder if one can, for example, just check nccl version from this requirements file and build from source?\nyes probably we should build and upload nccl for our supported cuda versions, same way as we build docker containers, by following details here:\nVerified with final rc"], "neg": []}
{"query": "template<typename FnType, FnType fn, typename ...Args> static PyObject* wrap_tuple_fn(Args ... args) { <del> PyObject *result = (*fn)(std::forward<Args>(args)...); </del> <ins> THPObjectPtr result((*fn)(std::forward<Args>(args)...)); </ins> if (!result) return NULL; <del> if (PyTuple_Check(result)) { return PyObject_CallFunctionObjArgs((PyObject*)&THPSizeType, result, NULL); </del> <ins> if (PyTuple_Check(result.get())) { return PyObject_CallFunctionObjArgs((PyObject*)&THPSizeType, result.get(), NULL); </ins> } <del> Py_INCREF(result); return result; </del> <ins> return result.release(); </ins> } static auto sq_concat = PyTuple_Type.tp_as_sequence->sq_concat;", "pos": ["I think I encountered a memory leak situation potentially due to . I create a to reproduce my observations. Basically, for three different (meaningless) models, once the function is used, the memory usage will keep increasing. To run the code, where the flag controls whether is used. Please run it long enough. I'm on Ubuntu 14.04 with python3.6. I tried two versions of pytorch, 0.1.122 and 0.1.12+. Side Note: Even without using , the memory usauge can also increase for a while, but it will finally stablize at a level. What's more, when (i.e. ) is used, it can take very long before the memory usage stablizes. I'm not sure whether this is a problem or not.\nI think I'm sure there is a memory leak due to , and I found a weird way to do a temporary fix under pytorch version \"0.1.122\". The simplest example is as follows: Basically, if we replace with , and pass in an unpacked list of numbers, (i.e., instead of ), the memory leak is gone. Using will also have the leak issue. Unfortunately, I don't know why this is working. Finally, since there are so many in pytorch code, to do a quick hacky fix, one can change the function in the file as follows:\nAfter some more research, I think we have a much bigger problem here. My current conjecture is that: whenever we use ( object) as an argument to a function, there will be a memory leak. An example is the function in , where of the code is which uses . Again, a minimal example for reproduction is as follows: As in the case, once I change the code to , memory leak is gone. So, in short, the problem is on the usage of .\nThe same issue. +1\nThanks for tracking this down and for the repro script!\nThanks for the quick fix! BTW, will the fix be included in the release version (0.1.12_?)?\nit'll be released in\nI see. Thanks again."], "neg": []}
{"query": "if momentum != 0: param_state = self.state[p] if 'momentum_buffer' not in param_state: <del> buf = param_state['momentum_buffer'] = d_p.clone() </del> <ins> buf = param_state['momentum_buffer'] = p.data.new().resize_as_(p.data).zero_() buf.mul_(momentum).add_(d_p) </ins> else: buf = param_state['momentum_buffer'] buf.mul_(momentum).add_(1 - dampening, d_p)", "pos": ["When using classic SGD optimizer with momentum with sparse embeddings the memory keeps garbage collecting / allocating leading to slow down and out of memory error eventually. ! ! The issue dissapears when momentum is not used ! or when embeddings are not sparse ! I'm using the last pytorch version on conda:\nI tried out your script with momentum 0.1 on master, it takes roughly 10800mb gpu memory max. This is caused by using sparse buffer. I'm sending out a PR for this."], "neg": []}
{"query": "input_size=(1, 2, 4, 4, 4)), OldModuleTest(nn.VolumetricMaxPooling, (2, 2, 2), <del> input_size=(2, 3, 5, 5, 5)), </del> <ins> input=(torch.randn(2, 3, 5, 5, 5) * 1000)), </ins> OldModuleTest(nn.VolumetricMaxPooling, (2, 2, 2, 2, 2, 2), input_size=(2, 3, 5, 5, 5),", "pos": ["A few times, VolumetricMaxPooling (in ) kept failing on the continuous builds, and needed to investigate that. So it turns out, there were these one set of indices that were in the same Pooling window that differed by less than epsilon. So, the numeric gradient was hitting boundary conditions (max-pooling is discontinuous of course). ! ! The obvious fix is to modify the test to not have the input tensor have these boundary conditions. However, as an aside, I am not sure why the boundary condition is only hit sometimes and not always, because the same random numbers are generated always, and division is unstable but it is deterministic (right? or have i always got this wrong)\nRepeat running will produce error occasional , more detail:"], "neg": []}
{"query": "above the diagonal. The argument :attr:`diagonal` controls which diagonal to consider. If <del> :attr:`diagonal` = 0, all elements on and below the main diagonal are </del> <ins> :attr:`diagonal` = 0, all elements on and above the main diagonal are </ins> retained. A positive value excludes just as many diagonals above the main diagonal, and similarly a negative value includes just as many diagonals below the main diagonal. The main diagonal are the set of indices", "pos": ["! It seems that the highlighted \"below\" should be \"upper\u201c. <!-- A clear and concise description of what content in is an issue. If this has to do with the general website, please file an issue at instead. If this has to do with , please file an issue at --\nHi, thank you for bringing this up. I'll send in a patch to fix it."], "neg": []}
{"query": "composed of several input planes. This module can be seen as the gradient of Conv2d with respect to its input. <del> It is sometimes (but incorrectly) refered to as a deconvolutional operation. </del> <ins> It is also known as a fractionally-strided convolution or a deconvolution (although it is not an actual deconvolution operation). </ins> | :attr:`stride` controls the stride for the cross-correlation. | If :attr:`padding` is non-zero, then the input is implicitly zero-padded on both sides <del> for :attr:`padding` number of points </del> <ins> for :attr:`padding` number of points. </ins> | If :attr:`output_padding` is non-zero, then the output is implicitly zero-padded on one side <del> for :attr:`output_padding` number of points | :attr:`dilation` controls the spacing between the kernel points. It is harder to describe, but this `link`_ has a nice visualization of what :attr:`dilation` does. | :attr:`groups` controls the connections between inputs and outputs. </del> <ins> for :attr:`output_padding` number of points. | :attr:`dilation` controls the spacing between the kernel points; also known as the \u00e0 trous algorithm. It is harder to describe, but this `link`_ has a nice visualization of what :attr:`dilation` does. | :attr:`groups` controls the connections between inputs and outputs. `in_channels` and `out_channels` must both be divisible by `groups`. </ins> |       At groups=1, all inputs are convolved to all outputs. |       At groups=2, the operation becomes equivalent to having two conv layers side by side, each seeing half the input channels, and producing half the output channels, and both subsequently concatenated. <ins> At groups=`in_channels`, each input channel is convolved with its own set of filters (of size `out_channels // in_channels`). </ins> The parameters :attr:`kernel_size`, :attr:`stride`, :attr:`padding`, :attr:`output_padding` can either be: <del> - a single ``int`` -- in which case the same value is used for the height and width dimension </del> <ins> - a single ``int`` -- in which case the same value is used for the height and width dimensions </ins> - a ``tuple`` of two ints -- in which case, the first `int` is used for the height dimension, and the second `int` for the width dimension", "pos": ["Hi, I have installed two versions of CUDA. But I want to use one of it, can I specify it during setup process?\nTry running\nI have tried it, () still returns False. Thx for your reply!\nDo you have the cuda driver installed?\nOf course I do. It raise assertion error which shows \"The NVIDIA driver on your system is too old(found version 8000)\".\nthat's the answer to your problem. upgrade your driver.\nThx!", "When building from source, I got this error. My system is OSX 10.12.4, conda, CUDA 8.0. Any thoughts? Thanks!\ncan you try adding to the file and see if that fixes things? if so i'll commit it.\nactually, that file already has memory included. Are you following instructions from Especially: This part is important:\nworks nicely! Thanks a lot!", "The suggest that takes a as a parameter, but it actually must be wrapped in a Variable first.\nyou are looking at the Tensor docs. When operating on Variable, it is implied from the tutorials that all functions operating on Variables have to use Variables. We didn't write separate docs for Variable itself.", "something wrong with import torchvision Traceback (most recent call last): File \"\", line 3, in <moduleimport torchvision File \"\", line 4, in <moduleimport torchvision.transforms as transforms ImportError: No module named transforms\nyou have a file called for some reason. delete it or change your pythonpath to ignore it.\nHey, I have the same problem persisting when I import torchvision. I have no other file by the same name. ImportError Traceback (most recent call last) <ipython-input-6-3baf6837d8f1in <module() 8 from import Dataset 9 from import DataLoader ---10 from torchvision import transforms ImportError: No module named torchvision Any idea why this is happening?\nDo you have torchvision installed (it is a separate package)?\npip Install torchvision try this\nfrom data import BaseTransform,VOC_CLASSES as labelmap this is not working\npip install torchvision", "Hi, I need to pass input through one nn.Module, then argmax the output, and then pass that output to a second nn.Module. I want to backprop through the argmax back to the weights of the first module. That's impossible right now, though, it seems. When I loop through the gradient of the weights of my first module (with the following code): firstmodel.zerograd() secondmodel.zerograd() loss.backward() # retainvariables=True for param in secondmodel.parameters(): print( 'data: ', ) print( 'grad data: ', ) for param in firstmodel.parameters(): print( 'data: ', ) print( 'grad data: ', ) I get the following error, stating that those gradients can no longer be viewed/exist: -main()\nargmax is not differentiable (and even if you try to think about it as a function differentiable almost everywhere, its derivative is always 0, so you won't get anything meaningful).", "def inithidden(self): h0 = Variable(FloatTensor(self.numlayers, self.batchlen, self.hiddendim).zero(),  requiresgrad=True) c0 = Variable(FloatTensor(self.numlayers, self.batchlen, self.hiddendim).zero(),  requiresgrad=True) return (h0, c0) def forward(self, sequences): lstmout, = ((1, self.batchlen, self.inputdim), ) ... = nn.LSTM(inputdim, hiddendim, numlayers, dropout = dropout, batchfirst=True) = self.inithidden() output = model(traininput) produces the following error although I defined weight and input dimensions as in the documentation: File \"/usr/local/lib/python2.7/dist-\", line 202, in call result = self.forward(input, kwargs) File \"/usr/local/lib/python2.7/dist-\", line 91, in forward output, hidden = func(input, self.allweights, hx) File \"/usr/local/lib/python2.7/dist-\", line 327, in forward return func(input, fargs, fkwargs) File \"/usr/local/lib/python2.7/dist-\", line 227, in forward nexth, output = func(input, hidden, weight) File \"/usr/local/lib/python2.7/dist-\", line 67, in forward hy, output = inner(input, hidden[l], weight[l]) File \"/usr/local/lib/python2.7/dist-\", line 96, in forward hidden = inner(input[i], hidden, *weight) File \"/usr/local/lib/python2.7/dist-\", line 22, in LSTMCell gates = F.linear(input, wih, bih) + F.linear(hx, whh, bhh) File \"/usr/local/lib/python2.7/dist-\", line 744, in add return (other) File \"/usr/local/lib/python2.7/dist-\", line 293, in add return (other, False) File \"/usr/local/lib/python2.7/dist-\", line 287, in add return Add(inplace)(self, other) File \"/usr/local/lib/python2.7/dist-\", line 20, in forward return a.add(b) RuntimeError: inconsistent tensor size at /data/users/soumith/builder/wheel/pytorch-src/torch/lib/TH/generic/THTensorMath.c:827\nI can't reproduce your problem. Could you include a short runnable example that exhibits this behavior?", "You also don't need to manually init hidden anymore, the following would work as well...\nThanks! After removing problem solved.", "Hi, I would like to implement a sharing weight 'alexnet' with two input batches of images, and the fully connected layers are concatenated together like this: However, when I run it, I come across the following error: Does anyone know about it? Thank you very much.\nYou forgot to call on the output of . In the future please remember that we're using the issues for bug reports and feature requests only. If you have questions or need help, please post ."], "neg": []}
{"query": "unset(CUDA_ARCH_PTX CACHE) endif() <del> if(DEFINED ENV{TORCH_CUDA_ARCH_LIST}) </del> <ins> if($ENV{TORCH_CUDA_ARCH_LIST}) </ins> # Pass CUDA architecture directly set(__cuda_arch_bin $ENV{TORCH_CUDA_ARCH_LIST}) message(STATUS \"Set CUDA arch from TORCH_CUDA_ARCH_LIST: ${__cuda_arch_bin}\")", "pos": ["This is most obvious by observing binary sizes, for example on pypi: (6.0 MB ) (33.8 MB ) Also, from the build log it's clear that 3.8 for some reason build with flag: CI cc\nWhich is a good old", "Einsum currently modifies variables in-place (without an explicit indication that it does so), which prevents pytorch from automatically backpropagating. Results in the following runtime error: Demonstration and discussion can also be found . PyTorch or Caffe2: PyTorch How you installed PyTorch (conda, pip, source): pip OS: PyTorch version: 0.4.0 Python version: 3.6.1\nOops. Thank you for reporting and the minimal example. I'll see to get that fixed.\nSaw the same issue over here. Thank you !\nIs cloning the tensor before passing to einsum a valid workaround?\nYes, but a somewhat expensive one. If you have the ability to recompile, you could also apply the PR fixing this. Somehow it seems to be stuck in the review queue because it hit an unrelated CI problem...\nOkay so how to use it then?\nJust get a recent enough master and it will work."], "neg": []}
{"query": "output_padding (int or tuple, optional): Zero-padding added to one side of the output groups (int, optional): Number of blocked connections from input channels to output channels bias (bool, optional): If True, adds a learnable bias to the output <ins> dilation (int or tuple, optional): Spacing between kernel elements </ins> Shape: - Input: :math:`(N, C_{in}, L_{in})`", "pos": ["Hi, I have installed two versions of CUDA. But I want to use one of it, can I specify it during setup process?\nTry running\nI have tried it, () still returns False. Thx for your reply!\nDo you have the cuda driver installed?\nOf course I do. It raise assertion error which shows \"The NVIDIA driver on your system is too old(found version 8000)\".\nthat's the answer to your problem. upgrade your driver.\nThx!", "When building from source, I got this error. My system is OSX 10.12.4, conda, CUDA 8.0. Any thoughts? Thanks!\ncan you try adding to the file and see if that fixes things? if so i'll commit it.\nactually, that file already has memory included. Are you following instructions from Especially: This part is important:\nworks nicely! Thanks a lot!", "The suggest that takes a as a parameter, but it actually must be wrapped in a Variable first.\nyou are looking at the Tensor docs. When operating on Variable, it is implied from the tutorials that all functions operating on Variables have to use Variables. We didn't write separate docs for Variable itself.", "something wrong with import torchvision Traceback (most recent call last): File \"\", line 3, in <moduleimport torchvision File \"\", line 4, in <moduleimport torchvision.transforms as transforms ImportError: No module named transforms\nyou have a file called for some reason. delete it or change your pythonpath to ignore it.\nHey, I have the same problem persisting when I import torchvision. I have no other file by the same name. ImportError Traceback (most recent call last) <ipython-input-6-3baf6837d8f1in <module() 8 from import Dataset 9 from import DataLoader ---10 from torchvision import transforms ImportError: No module named torchvision Any idea why this is happening?\nDo you have torchvision installed (it is a separate package)?\npip Install torchvision try this\nfrom data import BaseTransform,VOC_CLASSES as labelmap this is not working\npip install torchvision", "Hi, I need to pass input through one nn.Module, then argmax the output, and then pass that output to a second nn.Module. I want to backprop through the argmax back to the weights of the first module. That's impossible right now, though, it seems. When I loop through the gradient of the weights of my first module (with the following code): firstmodel.zerograd() secondmodel.zerograd() loss.backward() # retainvariables=True for param in secondmodel.parameters(): print( 'data: ', ) print( 'grad data: ', ) for param in firstmodel.parameters(): print( 'data: ', ) print( 'grad data: ', ) I get the following error, stating that those gradients can no longer be viewed/exist: -main()\nargmax is not differentiable (and even if you try to think about it as a function differentiable almost everywhere, its derivative is always 0, so you won't get anything meaningful).", "def inithidden(self): h0 = Variable(FloatTensor(self.numlayers, self.batchlen, self.hiddendim).zero(),  requiresgrad=True) c0 = Variable(FloatTensor(self.numlayers, self.batchlen, self.hiddendim).zero(),  requiresgrad=True) return (h0, c0) def forward(self, sequences): lstmout, = ((1, self.batchlen, self.inputdim), ) ... = nn.LSTM(inputdim, hiddendim, numlayers, dropout = dropout, batchfirst=True) = self.inithidden() output = model(traininput) produces the following error although I defined weight and input dimensions as in the documentation: File \"/usr/local/lib/python2.7/dist-\", line 202, in call result = self.forward(input, kwargs) File \"/usr/local/lib/python2.7/dist-\", line 91, in forward output, hidden = func(input, self.allweights, hx) File \"/usr/local/lib/python2.7/dist-\", line 327, in forward return func(input, fargs, fkwargs) File \"/usr/local/lib/python2.7/dist-\", line 227, in forward nexth, output = func(input, hidden, weight) File \"/usr/local/lib/python2.7/dist-\", line 67, in forward hy, output = inner(input, hidden[l], weight[l]) File \"/usr/local/lib/python2.7/dist-\", line 96, in forward hidden = inner(input[i], hidden, *weight) File \"/usr/local/lib/python2.7/dist-\", line 22, in LSTMCell gates = F.linear(input, wih, bih) + F.linear(hx, whh, bhh) File \"/usr/local/lib/python2.7/dist-\", line 744, in add return (other) File \"/usr/local/lib/python2.7/dist-\", line 293, in add return (other, False) File \"/usr/local/lib/python2.7/dist-\", line 287, in add return Add(inplace)(self, other) File \"/usr/local/lib/python2.7/dist-\", line 20, in forward return a.add(b) RuntimeError: inconsistent tensor size at /data/users/soumith/builder/wheel/pytorch-src/torch/lib/TH/generic/THTensorMath.c:827\nI can't reproduce your problem. Could you include a short runnable example that exhibits this behavior?", "You also don't need to manually init hidden anymore, the following would work as well...\nThanks! After removing problem solved.", "Hi, I would like to implement a sharing weight 'alexnet' with two input batches of images, and the fully connected layers are concatenated together like this: However, when I run it, I come across the following error: Does anyone know about it? Thank you very much.\nYou forgot to call on the output of . In the future please remember that we're using the issues for bug reports and feature requests only. If you have questions or need help, please post ."], "neg": []}
{"query": "#include \"THHalf.h\" <del> /* * Copyright 1993-2014 NVIDIA Corporation.  All rights reserved. * * NOTICE TO LICENSEE: * * This source code and/or documentation (\"Licensed Deliverables\") are * subject to NVIDIA intellectual property rights under U.S. and * international Copyright laws. * * These Licensed Deliverables contained herein is PROPRIETARY and * CONFIDENTIAL to NVIDIA and is being provided under the terms and * conditions of a form of NVIDIA software license agreement by and * between NVIDIA and Licensee (\"License Agreement\") or electronically * accepted by Licensee.  Notwithstanding any terms or conditions to * the contrary in the License Agreement, reproduction or disclosure * of the Licensed Deliverables to any third party without the express * written consent of NVIDIA is prohibited. * * NOTWITHSTANDING ANY TERMS OR CONDITIONS TO THE CONTRARY IN THE * LICENSE AGREEMENT, NVIDIA MAKES NO REPRESENTATION ABOUT THE * SUITABILITY OF THESE LICENSED DELIVERABLES FOR ANY PURPOSE.  IT IS * PROVIDED \"AS IS\" WITHOUT EXPRESS OR IMPLIED WARRANTY OF ANY KIND. * NVIDIA DISCLAIMS ALL WARRANTIES WITH REGARD TO THESE LICENSED * DELIVERABLES, INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY, * NONINFRINGEMENT, AND FITNESS FOR A PARTICULAR PURPOSE. * NOTWITHSTANDING ANY TERMS OR CONDITIONS TO THE CONTRARY IN THE * LICENSE AGREEMENT, IN NO EVENT SHALL NVIDIA BE LIABLE FOR ANY * SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL DAMAGES, OR ANY * DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, * WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS * ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE * OF THESE LICENSED DELIVERABLES. * * U.S. Government End Users.  These Licensed Deliverables are a * \"commercial item\" as that term is defined at 48 C.F.R. 2.101 (OCT * 1995), consisting of \"commercial computer software\" and \"commercial * computer software documentation\" as such terms are used in 48 * C.F.R. 12.212 (SEPT 1995) and is provided to the U.S. Government * only as a commercial end item.  Consistent with 48 C.F.R.12.212 and * 48 C.F.R. 227.7202-1 through 227.7202-4 (JUNE 1995), all * U.S. Government End Users acquire the Licensed Deliverables with * only those rights set forth herein. * * Any use of the Licensed Deliverables in individual and commercial * software must include, in the user documentation and internal * comments to the code, the above Disclaimer and U.S. Government End * Users Notice. */ </del> <ins> /* Copyright 1993-2014 NVIDIA Corporation.  All rights reserved. */ </ins> // Host functions for converting between FP32 and FP16 formats <del> // Paulius Micikevicius (pauliusm@nvidia.com) </del> float TH_half2float(THHalf h) {", "pos": ["When performing license check on Pytorch debian package I found this license problem: License of Pytorch is BSD-3-Clause, however THHalf.c is non-free according to its header.\nA possible replacement can be found in in NNPACK (BSD-3 licensed).\nActually NVIDIA's copyright declaration in THHalf.c makes it impossible for Pytorch to enter the main section of Debian archive. To promote to official package, we should have copyright problems cleared at first.\nWe're working on getting this cleared -- sorry for the inconvenience.\nMarking as done since was merged.\nIt's my bad. The fix entered torch/torch7 but is still not synced into pytorch.\nDid you want me to patch pytorch separately? I figured the file can now be copied over from torch by anyone.\ni'll sync the subtrees now!\nThank you guys."], "neg": []}
{"query": "'Cumprod', 'Gather', 'MaskedScatter', <del> 'Sort', 'Topk', </del> 'Norm', 'Prod', ))", "pos": ["Just a very small grammatical issue. ! Suggested fix: \"mutates\" to \"mutate\" <!-- A clear and concise description of what content in is an issue. If this has to do with the general website, please file an issue at instead. If this has to do with , please file an issue at --\nThanks for pointing this out. Please feel free to open a pull request for this.\nThanks for adding labels. Forgot to do that. I will make a pull request"], "neg": []}
{"query": "self += parameters def __getitem__(self, idx): <del> if not (-len(self) <= idx < len(self)): raise IndexError('index {} is out of range'.format(idx)) if idx < 0: idx += len(self) return self._parameters[str(idx)] </del> <ins> if isinstance(idx, slice): return ParameterList(list(self._parameters.values())[idx]) else: if not (-len(self) <= idx < len(self)): raise IndexError('index {} is out of range'.format(idx)) if idx < 0: idx += len(self) return self._parameters[str(idx)] </ins> def __setitem__(self, idx, param): return self.register_parameter(str(idx), param)", "pos": ["it'd be great to pick out elements of of a sequential with array slicing:\nI agree this idea!\nHowever,\nthis has come up again (last in ). Shall we reconsider?\nI attempted getting a to work. Is this fine? If this is fine, I can issue a PR too. Let me know what you think. I believe this modification needs to be done in the class.\ncc: This issue has been open for quite sometime."], "neg": []}
{"query": "upstream=\"$1\" pr=\"$2\" git diff --name-only \"$upstream\" \"$pr\" <del> git diff --name-only \"$upstream\" \"$pr\" | grep -Eq '^(aten/|caffe2/|.jenkins/pytorch|docs/(make.bat|Makefile|requirements.txt|source)|mypy|requirements.txt|setup.py|test/|third_party/|tools/|.gitmodules|torch/)' </del> <ins> # Now that PyTorch build depends on Caffe2, unconditionally trigger # for any changes. # TODO: Replace this with a NEGATIVE regex that allows us to blacklist # files (letting us skip builds when they are unnecessary) #git diff --name-only \"$upstream\" \"$pr\" | grep -Eq '^(aten/|caffe2/|.jenkins/pytorch|docs/(make.bat|Makefile|requirements.txt|source)|mypy|requirements.txt|setup.py|test/|third_party/|tools/|.gitmodules|torch/)' </ins>", "pos": ["This is most obvious by observing binary sizes, for example on pypi: (6.0 MB ) (33.8 MB ) Also, from the build log it's clear that 3.8 for some reason build with flag: CI cc\nWhich is a good old", "Einsum currently modifies variables in-place (without an explicit indication that it does so), which prevents pytorch from automatically backpropagating. Results in the following runtime error: Demonstration and discussion can also be found . PyTorch or Caffe2: PyTorch How you installed PyTorch (conda, pip, source): pip OS: PyTorch version: 0.4.0 Python version: 3.6.1\nOops. Thank you for reporting and the minimal example. I'll see to get that fixed.\nSaw the same issue over here. Thank you !\nIs cloning the tensor before passing to einsum a valid workaround?\nYes, but a somewhat expensive one. If you have the ability to recompile, you could also apply the PR fixing this. Somehow it seems to be stuck in the review queue because it hit an unrelated CI problem...\nOkay so how to use it then?\nJust get a recent enough master and it will work."], "neg": []}
{"query": "- **c_n** (num_layers * num_directions, batch, hidden_size): tensor containing the cell state for t=seq_len Attributes: <del> weight_ih_l[k] : the learnable input-hidden weights of the k-th layer `(W_ir|W_ii|W_in)`, of shape `(input_size x 3*hidden_size)` weight_hh_l[k] : the learnable hidden-hidden weights of the k-th layer `(W_hr|W_hi|W_hn)`, of shape `(hidden_size x 3*hidden_size)` bias_ih_l[k] : the learnable input-hidden bias of the k-th layer `(b_ir|b_ii|b_in)`, of shape `(3*hidden_size)` bias_hh_l[k] : the learnable hidden-hidden bias of the k-th layer `(W_hr|W_hi|W_hn)`, of shape `(3*hidden_size)` </del> <ins> weight_ih_l[k] : the learnable input-hidden weights of the k-th layer `(W_ii|W_if|W_ig|W_io)`, of shape `(input_size x 4*hidden_size)` weight_hh_l[k] : the learnable hidden-hidden weights of the k-th layer `(W_hi|W_hf|W_hg|W_ho)`, of shape `(hidden_size x 4*hidden_size)` bias_ih_l[k] : the learnable input-hidden bias of the k-th layer `(b_ii|b_if|b_ig|b_io)`, of shape `(4*hidden_size)` bias_hh_l[k] : the learnable hidden-hidden bias of the k-th layer `(W_hi|W_hf|W_hg|b_ho)`, of shape `(4*hidden_size)` </ins> Examples::", "pos": ["From : Happened because of some change in the environment as it is also affecting historical builds. cc\nThis issue seems to come from within libuv, a dependency of TensorPipe. First guess: there's two conflicting versions of libuv (e.g., a preinstalled one and the one shipped as a submodule of TensorPipe), and we build the one shipped by TensorPipe but somehow it ends up including the headers of the other one? In particular, the constant it's complaining about was in , which became part of libuv 1.41.0, which is the version that TensorPipe bundles. But it seems those jobs are installing version 1.40.0 via conda: So the question is why are the conda headers picked up and prioritized? I feel this will require some CMake digging, and it's possible that the fix will have to land in libuv's upstream CMake files. Ugh.\nBTW, if you want to validate this theory you can try setting this flag to : Which will cause TensorPipe to use the preinstalled libuv instead of building its own one. But please don't land this \"workaround\", as it will only lead to users coming to complain that they couldn't build PyTorch because their system has a libuv version that is too old for what TensorPipe needs.\nthis is definitely what is happening: libuv-1.40 is installed from conda, which does not have in the header: But I'm still confused why it started to happen just now [Edit] Can it be update that happened 4 hours ago?\nFinally found a PR that caused a regression: (and testing were clearly regressed in CI: )"], "neg": []}
{"query": "// a test for Py_None here; instead, you need to mark the argument // as *allowing none*; you can do this by writing 'Tensor?' instead // of 'Tensor' in the ATen metadata. <del> throw TypeError(\"expected Variable as argument %d, but got %s\", i, </del> <ins> throw TypeError(\"expected Tensor as argument %d, but got %s\", i, </ins> Py_TYPE(args[i])->tp_name); } return reinterpret_cast<THPVariable*>(args[i])->cdata;", "pos": ["Considering the merge of and , and the \"removal\" of from the API, there are still some messages that reference . This might prove to be confusing for new users starting off from , as they probably might not have heard of . I did a code search to find out multiple instances of this in . PyTorch or Caffe2: PyTorch How you installed PyTorch (conda, pip, source): source (@ ) Build command you used (if compiling from source): OS: Ubuntu 16.04.2 PyTorch version: Python version: 3.5 CUDA/cuDNN version: 8.0.44 GPU models and configuration: GeForce 940M GCC version (if compiling from source): 5.4.0 CMake version: 3.5.1"], "neg": []}
{"query": "self.assertEqual(torch.mm(flattened_tensor, flattened_tensor.t()), torch.eye(rows) * gain ** 2, prec=1e-6) <ins> # Generates rand tensor with non-equal values. This ensures that duplicate # values won't be causing test failure for modules like MaxPooling. # size should be small, otherwise randperm fails / long overflows. def _rand_tensor_non_equal(*size): total = reduce(mul, size, 1) return torch.randperm(total).view(*size).double() </ins> def add_test(test): test_name = test.get_name()", "pos": ["Intermittent test failures occurring on AdaptiveMaxPool3d. and I have seen this in the past where Max pooling has exactly the same values in the same window, then it's differently resolved on CPU/GPU. For example [0 2.5 2.5 3] can give a max index of or on CUDA depending on the runtime. So generating input values that are within epsilon of each other is important (especially at half precision). fix that one.\nWill do. Thanks for identifying the issue."], "neg": []}
{"query": "@skipIfNoLapack def test_gels(self): <del> def _test(a, b, expectedNorm): </del> <ins> def _test_underdetermined(a, b, expectedNorm): m = a.size()[0] n = a.size()[1] assert(m <= n) </ins> a_copy = a.clone() b_copy = b.clone() res1 = torch.gels(b, a)[0]", "pos": ["b = (600,1) A = (600,2) X, = (b,A) X.size() #torch.Size([2, 1]) Y, = (b.cuda(),A.cuda()) Y.size() #torch.Size([600, 1])\nthat is interesting. says that the first rows returned ( here) include the solution of the least-squares problem. I've verified that the first 2 rows of X and Y are the same so you can probably use that for now while we figure out what the rest of the output is."], "neg": []}
{"query": "return pad_bias<mem_eff_alignment>(attn_mask); } // Check and make the tensor contiguous if needed <del> if (attn_mask.sym_stride(0) % 16 != 0 || attn_mask.sym_stride(1) % 16 != 0 || attn_mask.sym_stride(2) % 16 != 0) { </del> <ins> auto needs_contig = [](const c10::SymInt& stride) { return (stride % 16 != 0) || (stride == 0); }; if (needs_contig(attn_mask.sym_stride(0)) || needs_contig(attn_mask.sym_stride(1)) || needs_contig(attn_mask.sym_stride(2)) || needs_contig(attn_mask.sym_stride(3))) { </ins> return attn_mask.contiguous(); }", "pos": ["Hi after more hours of debugging than I am comfortable to admit, I noticed the following breaking change between PyTorch 2.0.1 and PyTorch 2.1. The issue can be reproduced both with & . There is no issue on 2.0.1 For fp32 inputs to SDPA on CUDA device & passing a custom , we have the following: PyTorch 2.0.1: dispatches to math ( for other backends) PyTorch 2.1: dispatches to mem-efficient attention (likely following ). And it appears that the mem-efficient attention backend outputs wrong results when passing non-contigous key/value, while the math backend goes just fine. Reproduction: The result is: Note that these metrics are not super representative of how the drift is bad locally: ! The relevant tensors are attached in a zip here. Thank you! cc\nSo for this case the mask passed in is all Trues which will get converted to a \"zeros\" additive attention_bias mask.. in theory a no-op. When I run the same q,k,v input with no mask I get which leads me to believe that non-contiguous inputs alone are not the problem but non-contiguous + masking is no bueno\nIf I update the script to only call contiguous on either key or value but not both I still see the same results\nOkayy I got it, so this happens when bias_mask is not contiguous\nWeird - so not being contiguous was responsible of the issue?\nI can confirm this issue is fixed on\nUnfortunately this solutions makes the (1, 1, seqlenq, seqlenkv case) have to fully materialize the attention mask. I think that a third component of this bug is that it is occuring when seqlenq = 1, I tried with seqlen_q = 2 and it appears to not occur anymore\nI was acting somewhat rash with the impeding 2.1.1 deadline and I am unhappy with my solution, and I didn't fully think through its ramifications. For a very hot path I.e. a mask of (1, 1, seqlenq, seqlenkv) this causes a substantial memory regression.", "I am having a hard time figure out whats goin on with the repro mentioned here: The results of the repo appear to be dependent on order in which the ops are ran, which makes me thing there is some IMA happening, but alas I haven't found it.\nThanks a lot for looking into it sorry for having been a bit pushy about having this in the patch! I appear not to be able to reproduce the logits difference in using and A100-SXM4-80GB, is the output difference appearing using the memory patch\nTentatively marking hi-pri for the regression\nCould you help double check if the memory issue is fixed in 2.1.2? Here are the results and myself gets using the reproduction example in the issue: This is the same as what we get from 2.1.0\nSo the memory issue was \"fixed\" in 2.1.1 via this PR: While this did change the result of the repro, it caused a memory usage \"regression\" for a very common hot path for SDPA. This PR: Was used to fix the memory regression. I did further analysis on the initial correctness report. I determined that the repro wasn't behaving as I expected.. e.g the order with which aten ops were called impacted. I am also not entirely sure that a contiguous vs non contiguous input should produce identical results (it likely should but I need to reason if there is some potential float addmulitple ops that get reordered cc ) I was leaving this open in case we need to revisit this but the the memory reduction should be visible via usual memory profiling techniques\nI can also confirm that this issue is reintroduced in the 2.1.2 release, while it was fixed in 2.1.1. This is a bit bad for us as following 2.1.1 we SDPA by default in Transformers and a very common execution path has non-contiguous inputs: We'll likely have to do a patch and manually call , I guess. In 2.1.2, this codepath starts erroring out with . Here, silently SDPA starts outputting different results when inputs are not contiguous, which I think is not wishable. If the kernel requires contiguous inputs, why not check it on the C++ side and call if needed (maybe with a warning)?\ndo you have a repro for the illegal memory access?", "I wasn't able to re-repro this with the existing one in this issue.\nI don't have a minimal reproduction unfortunately, but you can try: and run with the following script: which errors at: This corresponds to this code: Note that on of Transformers you will not be able to reproduce as for now I just disabled the mem-efficient path for GPT BigCode due to this issue (fix at , removing the expand).\nIs it possible to have minimal repro for this ? Running using your instructions I get following:\nOh sorry. You can try with instead that is non-gated. I'll try to come up with a minimal repro ASAP. I edited my message above. For the record using\nRunning above command produced for me: Notable Environment differences: CUDA-12.1 PyTorch: Main GPU: H100 `\nWhile the repro isn't throwing a IMA compute-sanitizer is reporting errors:\nCurrent gists with compute-sanitizer outputs:\nI have a fix:\nI rerun the script to confirm that this is fixed in 2.2.0\nOkay I am going to close this issue for now"], "neg": []}
{"query": "self.assertEqual(reference[2, 2, 2, ...], 27, 0) self.assertEqual(reference[...], reference, 0) <del> reference_5d = self._consecutive((3, 3, 3, 3, 3)) </del> <ins> reference_5d = conv_fn(consec((3, 3, 3, 3, 3))) </ins> self.assertEqual(reference_5d[..., 1, 0], reference_5d[:, :, :, 1, 0], 0) self.assertEqual(reference_5d[2, ..., 1, 0], reference_5d[2, :, :, 1, 0], 0) self.assertEqual(reference_5d[2, 1, 0, ..., 1], reference_5d[2, 1, 0, :, 1], 0) self.assertEqual(reference_5d[...], reference_5d, 0) # LongTensor indexing <del> reference = self._consecutive((5, 5, 5)) idx = torch.LongTensor([2, 4]) </del> <ins> reference = conv_fn(consec((5, 5, 5))) idx = conv_fn(torch.LongTensor([2, 4])) </ins> self.assertEqual(reference[idx], torch.stack([reference[2], reference[4]])) # TODO: enable one indexing is implemented like in numpy # self.assertEqual(reference[2, idx], torch.stack([reference[2, 2], reference[2, 4]]))", "pos": ["Arguably, indexing a Tensor by a LongTensor with no dimension should return a LongTensor with no dimension. Consider the following case: This breaks indexing by indices when the list of indices is empty, which can be hard to detect. In my use-case, I use such a masking to select rows of a tensor corresponding to \"valid\" entries in a vector; I had not considered that it would not work when there are no valid entries.\nThanks! I'll take a look."], "neg": []}
{"query": "return linearIndex; } <ins> static bool hasEmptyTensor(TensorList tensors) { for (auto& tensor : tensors) { if (tensor.defined() && tensor.numel() == 0) { return true; } } return false; } </ins> static std::tuple<Tensor, Tensor> makeLinearIndex(Tensor self, TensorList orig) { <ins> checkIndexTensorTypes(orig); </ins> // first expand ByteTensor (boolean masks) into 1 or more LongTensors auto indices = expandByteTensors(self, orig); <ins> if (hasEmptyTensor(indices)) { return std::make_tuple(self, self.type().toScalarType(kLong).tensor()); } </ins> // next broadcast all index tensors together indices = expand_outplace(indices); // add missing null Tensors so that it matches self.dim()", "pos": ["In PyTorch master 0.4 , variables cannot be indexed. This used to work in PyTorch 0.2 :\nThis is breaking some of Pyro's inference algorithms on PyTorch master:"], "neg": []}
{"query": "return() endif() set(CAFFE2_FOUND_CUDA TRUE) <ins> message(STATUS \"Caffe2: CUDA detected: \" ${CUDA_VERSION}) message(STATUS \"Caffe2: CUDA nvcc is: \" ${CUDA_NVCC_EXECUTABLE}) message(STATUS \"Caffe2: CUDA toolkit directory: \" ${CUDA_TOOLKIT_ROOT_DIR}) if(CUDA_FOUND) # Sometimes, we may mismatch nvcc with the CUDA headers we are # compiling with, e.g., if a ccache nvcc is fed to us by CUDA_NVCC_EXECUTABLE # but the PATH is not consistent with CUDA_HOME.  It's better safe # than sorry: make sure everything is consistent. set(file \"${PROJECT_BINARY_DIR}/detect_cuda_version.c\") file(WRITE ${file} \"\" \"#include <cuda.h>n\" \"#include <stdio.h>n\" \"int main() {n\" \"  printf(\"%d.%d\", CUDA_VERSION / 1000, (CUDA_VERSION / 10) % 100);n\" \"  return 0;n\" \"}n\" ) try_run(run_result compile_result ${PROJECT_BINARY_DIR} ${file} CMAKE_FLAGS \"-DINCLUDE_DIRECTORIES=${CUDA_INCLUDE_DIRS}\" LINK_LIBRARIES ${CUDA_LIBRARIES} RUN_OUTPUT_VARIABLE cuda_version_from_header COMPILE_OUTPUT_VARIABLE output_var ) if(NOT compile_result) message(FATAL_ERROR \"Caffe2: Couldn't determine version from header: \" ${output_var}) endif() message(STATUS \"Caffe2: Header version is: \" ${cuda_version_from_header}) if(NOT ${cuda_version_from_header} STREQUAL ${CUDA_VERSION}) # Force CUDA to be processed for again next time # TODO: I'm not sure if this counts as an implementation detail of # FindCUDA set(${cuda_version_from_findcuda} ${CUDA_VERSION}) unset(CUDA_TOOLKIT_ROOT_DIR_INTERNAL CACHE) # Not strictly necessary, but for good luck. unset(CUDA_VERSION CACHE) # Error out message(FATAL_ERROR \"FindCUDA says CUDA version is ${cuda_version_from_findcuda} (usually determined by nvcc), \" \"but the CUDA headers say the version is ${cuda_version_from_header}.  This often occurs \" \"when you set both CUDA_HOME and CUDA_NVCC_EXECUTABLE to \" \"non-standard locations, without also setting PATH to point to the correct nvcc.  \" \"Perhaps, try re-running this command again with PATH=${CUDA_TOOLKIT_ROOT_DIR}/bin:$PATH.  \" \"See above log messages for more diagnostics, and see https://github.com/pytorch/pytorch/issues/8092 for more details.\") endif() endif() </ins> # Find cuDNN. if(CAFFE2_STATIC_LINK_CUDA)", "pos": ["Here's the situation: Ideally, you'd say something like to make a build of PyTorch with a non-standard version of CUDA. When this happens, needs to find the correct copy of nvcc to use. If you are using ccache as per the instructions in you'll have set to your ccache binary, so that ccache happens. cmake will correctly respect this variable. However, here's the problem: ccache's nvcc will use your path to determine what the real nvcc to use is, this means that you'll pick up whatever nvcc is in your PATH, which is probably not the one you want, because you're trying to use a different nvcc. So in sum, you have to do the non-intuitive to get the whole shebang to work. The worst part is that it won't tell you if you've got it wrong, it will happily use the wrong nvcc. I haven't gone through user submitted issues to see if anyone has succumbed to this particular bug.\nA long-term solution might be getting to work\nworks now."], "neg": []}
{"query": "(4.53, 3.83, -6.64))).t() b = torch.Tensor(((8.58, 8.26, 8.48), (9.35, -4.43, -0.70))).t() <del> _test(a, b, expectedNorm) </del> <ins> _test_underdetermined(a, b, expectedNorm) </ins> # test reuse expectedNorm = 0", "pos": ["b = (600,1) A = (600,2) X, = (b,A) X.size() #torch.Size([2, 1]) Y, = (b.cuda(),A.cuda()) Y.size() #torch.Size([600, 1])\nthat is interesting. says that the first rows returned ( here) include the solution of the least-squares problem. I've verified that the first 2 rows of X and Y are the same so you can probably use that for now while we figure out what the rest of the output is."], "neg": []}
{"query": "n_dist = THTensor_(size)(prob_dist, 0); n_categories = THTensor_(size)(prob_dist, 1); <del> THArgCheck(n_sample > 0, 2, \"cannot sample n_sample < 0 samples\"); </del> <ins> THArgCheckWithCleanup(n_sample > 0, THCleanup(if (start_dim == 1) THTensor_(resize1d)(prob_dist, n_categories);), 2, \"cannot sample n_sample < 0 samples\"); </ins> if (!with_replacement) { <del> THArgCheck((!with_replacement) && (n_sample <= n_categories), 2,  \"cannot sample n_sample > prob_dist:size(1) samples without replacement\"); </del> <ins> THArgCheckWithCleanup((!with_replacement) && (n_sample <= n_categories), THCleanup(if (start_dim == 1) THTensor_(resize1d)(prob_dist, n_categories);), 2, \"cannot sample n_sample > prob_dist:size(1) samples without replacement\"); </ins> } /* cumulative probability distribution vector */", "pos": ["I was reading the documentation of method in the . I tried this in IPython environment. Of course, error occurs because 5 is larger than the size of weights and I didn't set . The error message is like this: But after this, the dimensions of weights changed. If I run as following without runtime exception, then the dimensions of weights do not change. This is caused by not cleaning up after exception raised in this . Maybe the solution is adding cleaning up code in the .\nThe problem emerged in two other situations as well, here are the examples: I'm sending a pull request with the fix. Thanks for reporting the problem. :)"], "neg": []}
{"query": "return r; } <del> std::vector<Tensor> as_tensor_list(std::vector<Variable> &vars) { </del> <ins> static std::vector<Tensor> as_tensor_list(std::vector<Variable> &vars) { </ins> std::vector<Tensor> tensors; for (auto& v : vars) { tensors.emplace_back(std::move(v));", "pos": ["See Docker run fails with , which is due to the fact that is specified in docker container: Which results in: CI cc\nCheck that adding to manywheels image results in a failure:\nTentatively marking hi-pri since nightly testing are failing\nThis is resolved, here is the successful run:"], "neg": []}
{"query": "}; struct PythonEngine : public Engine { <del> virtual void thread_main(std::shared_ptr<ReadyQueue> queue) override { </del> <ins> virtual void thread_main(std::shared_ptr<ReadyQueue> queue, int device) override { </ins> // Create a PyThreadState, but release the GIL. This lets AutoGIL calls // inside thread_main acquire the GIL without having to create a new // PyThreadState each time. AutoGIL gil; AutoNoGIL no_gil; <del> Engine::thread_main(queue); </del> <ins> Engine::thread_main(queue, device); </ins> } virtual void thread_on_exception(FunctionTask& task, std::exception& e) override {", "pos": ["Consider following code, and run it with multi gpus (e.g. 4): It will output: That is, the default device would always be zero even in DataParallel. My pytorch version is 0.1.122. I think this is not the desired behavior. It would cause some troubles. I tried to insert my own cuda kernel into backward to calculate the gradients, it become very slow, and I fixed it by (grad.getdevice()). Anyway, I think currentdevice in forward and backward should be the same, or could anyone explain to me why they are different?\nthis seems like something we should fix: ?\n<!-- drci-comment-start --:pagefacingup: Preview :pagefacingup: Preview :question: Need help or want to give feedback on the CI? Visit our Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. <!-- drci-comment-end --"], "neg": []}
{"query": "print('-- Detected CUDA at ' + CUDA_HOME) else: print('-- Not using CUDA') <del> if WITH_NCCL and SYSTEM_NCCL: print('-- Using system provided NCCL library') </del> <ins> if WITH_NCCL and WITH_SYSTEM_NCCL: print('-- Using system provided NCCL library at ' + NCCL_LIB_DIR + ', ' + NCCL_INCLUDE_DIR) </ins> elif WITH_NCCL: print('-- Building NCCL library') else:", "pos": ["Currently, you can have a libnccl on your system, and pytorch setup favors that instead of compiling one itself from source. The code that takes care of this is: However, this breaks gloo build because gloo gets a path to NCCL via:\nI got the error during the installation: Is it related to this problem?\nthat's a dupe of\nno, that's a consequence of , now you have to do before compiling. I think you don't want to do recursive submodule update, because that will pull in gtest and another copy of nccl.\nAlternatively we can pass new flags to gloo to prevent it from using its own NCCL\nThanks works for me.\nYou can prevent it from using its own nccl, but its cmake still has to be able to find some nccl.\nI think should be sufficient. I'm testing.\nfixes it when we dont have an external NCCL and pytorch compiles it's own NCCL?\nI don't see how NCCLEXTERNAL fixes anything, looks like in gloo NCCLEXTERNAL is used just to set dependency if gloo is compiling its own nccl (which we don't want in any case).\nThe only thing I tested when I was that the vendored nccl was not being built (even if its submodule was checked out); I tested this by adding an to one of the headers. I'll test with the other configurations.\nActually, there are a few bugs here. First, even though GLOOFLAGS looks like it's setting NCCLROOTDIR: it actually isn't because the gloo invocation is which expands to but only is passed to cmake. Suppose you fix this. nccl is still not going to be picked up, because INSTALLDIR is PyTorch's local install dir, but if you disabled nccl builds, there will never actually be a nccl here. So cmake is just going to do it's standard search to find a nccl somewhere. Inversely, if you ever built nccl, and then turn it off, we will keep picking up that locally built version until you delete it. So we shouldn't do that. Suppose you fix this.", "Then I think gloo will (continue to) correctly pick up system nccl and be built with it, and the final thing to fix is that build of PyTorch csrc doesn't know how to search for nccl headers somewhere. This is probably related to . I wonder, maybe we shouldn't do any autodetection at all, and force the user to pass us the include and library paths of nccl if they want to build with system.\nFYI, I committed a change to gloo to stop bundling NCCL itself. It will now happily pick up whatever NCCL you give it through NCCLROOTDIR or by trying to find it in the standard systems paths. See facebookincubator/gloo You do have to set to make Gloo try and find it and compile in support.\nSo to fix this, you should be able to not specify anything aside from if you want to use system NCCL, and add if you have some custom location (either a PyTorch bundled version or a custom install altogether)."], "neg": []}
{"query": "going to be volatile. Volatility spreads accross the graph much easier than non-requiring gradient - you only need a **single** volatile leaf to have a volatile output, while you need **all** leaves to not require gradient to <del> have an output the doesn't require gradient. Using volatile flag you don't </del> <ins> have an output that doesn't require gradient. Using volatile flag you don't </ins> need to change any settings of your model parameters to use it for inference. It's enough to create a volatile input, and this will ensure that no intermediate states are saved.", "pos": ["Platforms: mac, macos This test was disabled because it is failing on master (). This fails again , so I disable this test pending the discussion in on how best to clean up these runners. cc\n<!-- validation-comment-start --<bodyHello there! From the DISABLED prefix in this issue title, it looks like you are attempting to disable a test in PyTorch CI. The information I have parsed is below: Test name: Platforms for which to skip the test: mac, macos Within ~15 minutes, will be disabled in PyTorch CI for these platforms: mac, macos. Please verify that your test name looks correct, e.g., . To modify the platforms list, please include a line in the issue body, like below. The default action will disable the test for all platforms if no platforms list is specified. We currently support the following platforms: asan, linux, mac, macos, rocm, win, windows.</body<!-- validation-comment-end --\nI don't think this is really sustainable (nor productive) as all tests will eventually fail with no space left on device error... Instead, we should be reclaiming space on those nodes...\nYeah, this is only a temporary measure while we continue the investigation in This test writes a 2GB file to disk I think. I'll bring up the issue again tomorrow to get help from the team. At the moment, I have no idea which one is responsible for filling up the disk.\nmac M1 runners has been cleaned up, resolving this one\nfailing again\nGot a fix here to increase the minimum space required on MacOS to be 6GB. The current threshold of 4GB is not big enough"], "neg": []}
{"query": "IF ($ENV{PYTORCH_BINARY_BUILD}) MESSAGE(STATUS \"PYTORCH_BINARY_BUILD detected. Statically linking libstdc++\") SET(CMAKE_CXX_FLAGS \"-static-libstdc++ ${CMAKE_CXX_FLAGS}\") <ins> IF (UNIX AND NOT APPLE) # hiding statically linked library symbols, this flag is not available for the linker under MACOSX SET(CMAKE_CXX_FLAGS \"-Wl,--exclude-libs,ALL ${CMAKE_CXX_FLAGS}\") ENDIF(UNIX AND NOT APPLE) </ins> ENDIF() ADD_LIBRARY(shm SHARED core.cpp)", "pos": ["see the commit comment: To reproduce install ros-kinetic-tf 1.11.8-0xenial---0800 package from the deb xenial main repository. my system is: lsb_release -a No LSB modules are available. Distributor ID: Ubuntu Description: Ubuntu 16.04.1 LTS Release: 16.04 Codename: xenial cuda version is 8.0"], "neg": []}
{"query": "self.assertRaisesRegex(RuntimeError, 'modified by an inplace operation', lambda: z.backward()) <ins> def test_no_grad_input(self): class MyFunction(Function): @staticmethod def forward(self, x): return x @staticmethod def backward(self, grad_output): return grad_output x = torch.randn(5, requires_grad=True) with torch.no_grad(): y = MyFunction.apply(x) self.assertTrue(x.requires_grad) self.assertIsNone(y.grad_fn) </ins> def test_backward_copy(self): # This tests checks backward engine for a very subtle bug that appreared # in one of the initial versions of autograd. Gradients tensors were", "pos": ["OS: ubuntu 16.04 PyTorch version: 0.4.0a0+ How you installed PyTorch (conda, pip, source): source Python version: 3.6.3 CUDA/cuDNN version: v9.1, cudnn-9.1-linux-x64-v7 GPU models and configuration: 4x Tesla M60 GCC version (if compiling from source): 5.4.0 When using context () on a model with dataparallel operation (same with ) on 1 GPU the flag requires_grad does not come back to True. If only one GPU is used then everything works fine. Script: On 2 GPUs, this outputs which is wrong in the last line. On 1 GPU (set ) the output is correct.\nis looking into it"], "neg": []}
{"query": "dim_args_idx = test[4] if len(test) == 5 else [] <ins> skipTestIf = test[5] if len(test) == 6 else [] </ins> for dim_perm in product([-1, 1], repeat=len(dim_args_idx)): test_name = basic_test_name new_constructor_args = [arg * dim_perm[dim_args_idx.index(i)] if i in dim_args_idx else arg", "pos": ["Command to install the latest nightly binary: installs : However, PyTorch itself reports : I guess this version tag might be defined during the build process from e.g. and then forwarded from , but also source builds should use based on the tag. Unsure, if this is just a reporting issue or if we are building the nightly binaries with and executing them with , so creating this issue to dig a bit further into it. CC cc\nThis is really weird, I've put up measure in place to validate against it... Will investigate today (and moving to pytorch repo)\ncan you please check that 2.2.0 reports (and is build) against the correct version of nccl library?\nConfirmed for nightlies : Same issue for RC:\nLooking at the build log Looks like we are compiling with: USESYSTEMNCCL\nOn Windows I see following exception: We compiling with should we return None rather then exception ?\nIn this case we are indeed using 2.18.1: Do you know why we are using the system-wide NCCL installation instead of the one? Should we delete the system NCCL and use by default?\nLooks like its set here: Looks like it was by: Let me propose the PR to unset this flag : USESYSTEM_NCCL\nI guess we are using the system NCCL installation to be able to during the runtime. If we would use the module I assume we would link NCCL into the PyTorch binaries. If so, we should make sure to update the NCCL version whenever is updated. This could break our dependency on the NCCL pip package.\nYes. Also when we update this nccl here:\nWe used following path  What would be good path for nccl 2.19.3 and cuda 12.1 ? As per this issue: Will be using 12.0 ncck 2.19.3\nInstalling system NCCL into the nightly build docker container feels great, but we need to figure out, how to keep it in sync with PyTorch. I wonder if one can, for example, just check nccl version from this requirements file and build from source?\nyes probably we should build and upload nccl for our supported cuda versions, same way as we build docker containers, by following details here:\nVerified with final rc"], "neg": []}
{"query": "{ char transpose_r, transpose_m1, transpose_m2; THTensor *r__, *m1_, *m2_; <ins> int free_m1 = 0; int free_m2 = 0; </ins> if( (m1->nDimension != 2) || (m2->nDimension != 2)) THError(\"matrices expected, got %dD, %dD tensors\", m1->nDimension, m2->nDimension);", "pos": ["Here is an example of what seems to be a memory leak in pytorch, although it could be numpy? The leak was only occurring when using the or array unsqueezing methods in numpy. Also worth noting that the issue is resolved by removing the from and casting the torch tensor to type. An easy workaround is to reshape the array after turning it into a torch Tensor."], "neg": []}
{"query": "raise assert not hasattr(TestAutograd, test_name), 'Two tests have the same name: ' + test_name <ins> for skip in skipTestIf: do_test = skip(do_test) </ins> setattr(TestAutograd, test_name, do_test)", "pos": ["Command to install the latest nightly binary: installs : However, PyTorch itself reports : I guess this version tag might be defined during the build process from e.g. and then forwarded from , but also source builds should use based on the tag. Unsure, if this is just a reporting issue or if we are building the nightly binaries with and executing them with , so creating this issue to dig a bit further into it. CC cc\nThis is really weird, I've put up measure in place to validate against it... Will investigate today (and moving to pytorch repo)\ncan you please check that 2.2.0 reports (and is build) against the correct version of nccl library?\nConfirmed for nightlies : Same issue for RC:\nLooking at the build log Looks like we are compiling with: USESYSTEMNCCL\nOn Windows I see following exception: We compiling with should we return None rather then exception ?\nIn this case we are indeed using 2.18.1: Do you know why we are using the system-wide NCCL installation instead of the one? Should we delete the system NCCL and use by default?\nLooks like its set here: Looks like it was by: Let me propose the PR to unset this flag : USESYSTEM_NCCL\nI guess we are using the system NCCL installation to be able to during the runtime. If we would use the module I assume we would link NCCL into the PyTorch binaries. If so, we should make sure to update the NCCL version whenever is updated. This could break our dependency on the NCCL pip package.\nYes. Also when we update this nccl here:\nWe used following path  What would be good path for nccl 2.19.3 and cuda 12.1 ? As per this issue: Will be using 12.0 ncck 2.19.3\nInstalling system NCCL into the nightly build docker container feels great, but we need to figure out, how to keep it in sync with PyTorch. I wonder if one can, for example, just check nccl version from this requirements file and build from source?\nyes probably we should build and upload nccl for our supported cuda versions, same way as we build docker containers, by following details here:\nVerified with final rc"], "neg": []}
{"query": "np.int64, np.int32, np.int16, <del> np.uint8 </del> <ins> np.uint8, np.longlong, </ins> ] for dtype in dtypes: array = np.array([1, 2, 3, 4], dtype=dtype)", "pos": ["Minimal example: Torch initializes the tensor with values which are different from the numpy array. I've seen issues for similar kinds of problems (see ). Maybe the easiest solution to all these initialization problems might be, to check for the same values of the original and initialized array/tensor instead of initializing tensors silently with wrong numbers. What do you think? Furthermore, I've opened an issue which is also related this one in terms of initializing a numpy array with a number of type and then initializing a torch tensor from that array, see Feel free to merge them if you think that both issues can be solved together.\nCould you please provide some information about what version of pytorch and numpy you're using? ! It works as expected for me on master\nI'm using torch 0.3.0.post4 (latest one on website of PyTorch) and numpy 1.13.3.\nI'm running the same thing, this is strange. In the first line , I assumed that came from . Is this the case?\nNope, it's indeed the python built-in type. Actually this issue came up when I tried to get the shape of (i.e. , , returns ).\nThanks for the clarification, I'm now seeing your results. Looks like the tensor constructor doesn't work well with python2.7 long types in this case\nSure, no problem. Thanks for looking into it!\nThe problem seems to be that np.longlong and np.int64 are the same thing (at least on a 64bit system, I think), but the NPYLONGLONG and NPYINT64 macros are different... I'll put up a pr for this soon.", "A minimal example: Although has and this type is listed as supported, the array can't be converted. Consider that is initialized with an upper bound of type . I'm not completely sure, whether this issue is related to torch or numpy, but numpy initializes the array with values as expected, whereas torch throws an inconsistent error message.\nEdit: nevermind, I didn't run the same exact thing ! Works for me on master and 0.3. Seems like it could be related to"], "neg": []}
{"query": "(IndexCopy,     (0,),               ((S, S), index_variable(2, S), (2, S))      ), (IndexFill,     (0, 2),             ((S, S), index_variable(2, S))              ), (IndexSelect,   (0,),               ((S, S), index_variable(2, S))              ), <ins> (Gather,        (0,),               ((M, S), gather_variable((S, S), 1, M))     ), (Gather,        (1,),               ((M, S), gather_variable((M, S//2), 0, S)), 'dim1'), (Scatter,       (0,),               ((M, S), gather_variable((S, S), 1, M), (S, S))), (Scatter,       (1,),               ((M, S), gather_variable((M, S//2), 0, S), (M, S//2)), 'dim1'), </ins> (Concat,        (0,),               ((1, S, S), (2, S, S), (3, S, S))           ), (Resize,        (S*S, S),           ((S, S, S),)                                ), (Diag,          (),                 ((S, S),),                  '2d'            ),", "pos": ["<!-- A clear and concise description of what content in is an issue. If this has to do with the general website, please file an issue at instead. If this has to do with , please file an issue at --In the getting started section section the following quote appears: I guess this is from previous versions where CUDA support was not built-in and this can be removed, correct? cc\nThis seems to be easy fixable docs issue\nIf gets merged this is fixed.\nThanks for the PR Just merged it."], "neg": []}
{"query": "lib_path = os.path.join(cwd, \"torch\", \"lib\") <del> # Check if you remembered to check out submodules def check_file(f): if not os.path.exists(f): print(\"Could not find {}\".format(f)) print(\"Did you run 'git submodule update --init'?\") sys.exit(1) check_file(os.path.join(lib_path, \"gloo\", \"CMakeLists.txt\")) check_file(os.path.join(lib_path, \"nanopb\", \"CMakeLists.txt\")) check_file(os.path.join(lib_path, \"pybind11\", \"CMakeLists.txt\")) </del> tmp_install_path = lib_path + \"/tmp_install\" include_dirs += [ cwd,", "pos": ["It seems like running has a dependency on first running . This command also seems to also error out when installing via a wheel. This creates a problem for tools like  which rely on the output of to examine dependencies. Is there anyway to allow this command to not error out when installing via a wheel? Full stack trace: I understand the preferred package manager for this project is conda but being officially endorsed from , pipenv has a fairly large audience:\nthanks do you know when's the next planned release?\nHopefully within a month"], "neg": []}
{"query": "THCDeviceTensor<Dtype, 1> gradOutput, THCDeviceTensor<Dtype, 2> gradInput, Dtype *weights, <ins> int n_classes, </ins> int ignore_index) { CUDA_KERNEL_LOOP(index, batch_size) {", "pos": ["Claim: When I use an invalid label with F.crossentropy with reduce=False, I don't receive an error message. I do receive an error message when reduce=True. Code: Result: No error message received. Expected error message: Context: I was using , accidentally used an invalid label, and got extremely large losses and extremely small losses, and inconsistently so. Environment: OS: Ubuntu 16.04 on Docker and host PyTorch version: 0.3.0.post4 How you installed PyTorch: conda Python version: 3.6.3"], "neg": []}
{"query": "if (is_padding_neg()) throw std::runtime_error(\"negative padding is not supported\"); if (is_output_padding_neg()) throw std::runtime_error(\"negative output_padding is not supported\"); <del> AutoGPU guard(input_.data->getDevice()); auto input = input_.unpack_data()->contiguous(); </del> <ins> auto input = input_.unpack_data(); AutoGPU guard(input->getDevice()); input = input->contiguous(); </ins> std::unique_ptr<Tensor> weight(weight_.unpack_data()->clone_shallow()); auto bias = bias_.unpack_data(); auto grad_output = grad_outputs[0]->data->contiguous();", "pos": ["Usually calling backwards twice gives this message: But conv2d just segfaults. $ pt A B Segmentation fault (core dumped)"], "neg": []}
{"query": "{\"_cuda_getDevice\",   (PyCFunction)THCPModule_getDevice_wrap,   METH_NOARGS,  NULL}, {\"_cuda_getDeviceCount\", (PyCFunction)THCPModule_getDeviceCount_wrap, METH_NOARGS, NULL}, {\"_cuda_getDeviceName\", (PyCFunction)THCPModule_getDeviceName_wrap, METH_O,   NULL}, <ins> {\"_cuda_getDeviceCapability\", (PyCFunction)THCPModule_getDeviceCapability_wrap, METH_O,   NULL}, </ins> {\"_cuda_getCurrentStream\", (PyCFunction)THCPModule_getCurrentStream_wrap, METH_NOARGS, NULL}, {\"_cuda_getCurrentBlasHandle\", (PyCFunction)THCPModule_getCurrentBlasHandle_wrap, METH_NOARGS, NULL}, {\"_cuda_setStream\",    (PyCFunction)THCPModule_setStream_wrap,  METH_O, NULL}, {\"_cuda_isDriverSufficient\", (PyCFunction)THCPModule_isDriverSufficient, METH_NOARGS, NULL}, {\"_cuda_getDriverVersion\", (PyCFunction)THCPModule_getDriverVersion, METH_NOARGS, NULL}, <ins> {\"_cuda_getCompiledVersion\", (PyCFunction)THCPModule_getCompiledVersion, METH_NOARGS, NULL}, </ins> {\"_cuda_getRNGState\", (PyCFunction)THCPModule_getRNGState,      METH_NOARGS,  NULL}, {\"_cuda_setRNGState\", (PyCFunction)THCPModule_setRNGState,      METH_O,       NULL}, {\"_cuda_emptyCache\", (PyCFunction) THCPModule_emptyCache,       METH_NOARGS,  NULL},", "pos": ["There's quite a few user-reports of installing the CUDA 7.5 binary of pytorch and then running on GTX 1080 / 1070, and there will be reports of installing 7.5 / 8.0 binaries and running on V100. Raising a user-warning in such situations and letting them know will fix the problem.\nThe version of CUDA used to compile the binary should be available thanks to this PR:\nI think is a more robust way of getting the CUDA used to compile PyTorch, as currently if there are multiple CUDA installed in the machine, the heuristics used in might not match those used by CMake FindCUDA. I think that we should modify the to return the cuda version used during compilation. I have the impression that it will not require changing much because doesn't require cuda to execute. What do you think?"], "neg": []}
{"query": "extra_link_args = [] if IS_WINDOWS: <del> extra_compile_args = ['/Z7', '/EHa', '/DNOMINMAX' </del> <ins> extra_compile_args = ['/Z7', '/EHa', '/DNOMINMAX', '/wd4267', '/wd4251', '/wd4522', '/wd4522', '/wd4838', '/wd4305', '/wd4244', '/wd4190', '/wd4101', '/wd4996', '/wd4275' </ins> # /Z7 turns on symbolic debugging information in .obj files # /EHa is about native C++ catch support for asynchronous # structured exception handling (SEH) # /DNOMINMAX removes builtin min/max functions <ins> # /wdXXXX disables warning no. XXXX </ins> ] if sys.version_info[0] == 2: # /bigobj increases number of sections in .obj file, which is needed to link", "pos": ["That's pretty big! In contrast, an OS X build log is only 1.3M. Is there a reason they have to be this big? CC\nMy guess is that it's because of warnings. I remember having to get through thousands of them when looking for the error that failed the build.\nYes, the major part of them are warnings. And some of them is not fixable. We can hide them using some flags on the compilers.\nAnd this build configuration actually contains both the CPU and GPU build, which may also be a possible reason.\nThe reduced log size after PR is about 1.2M."], "neg": []}
{"query": "0.0000  0.0000 -1.6250 [torch.FloatTensor of size 3x3] <del> >>> torch.triu(a, k=1) </del> <ins> >>> torch.triu(a, diagonal=1) </ins> 0.0000  1.7304  1.4573 0.0000  0.0000 -0.1809 0.0000  0.0000  0.0000 [torch.FloatTensor of size 3x3] <del> >>> torch.triu(a, k=-1) </del> <ins> >>> torch.triu(a, diagonal=-1) </ins> 1.3225  1.7304  1.4573 -0.3052 -0.3111 -0.1809", "pos": ["The states that the argument controls which diagonal to consider. However, this argument is actually named in pytorch 0.2.0."], "neg": []}
{"query": "torch.mm(a, b, out=self.output) else: if self.transA: <del> a = a.transpose(2, 3) </del> <ins> a = a.transpose(1, 2) </ins> if self.transB: <del> b = b.transpose(2, 3) </del> <ins> b = b.transpose(1, 2) </ins> self.output.resize_(a.size(0), a.size(1), b.size(2)) torch.bmm(a, b, out=self.output)", "pos": ["Hi I think that in lines 28 and 30: if self.transA: a = a.transpose(2, 3) if self.transB: b = b.transpose(2, 3) should be: if self.transA: a = a.transpose(1, 2) if self.transB: b = b.transpose(1, 2) Indeed in that branch the tensor has 3 dimension and thus the code crashes. Maybe an indexing error when translating from lua ? Thank you very much,\nYes, you're right. Could you send a PR with a fix? Thanks!\nYes of course, just done, thank you very much."], "neg": []}
{"query": "last_epoch (int): The index of last epoch. Default: -1. Example: <del> >>> # Assuming optimizer uses lr = 0.5 for all groups </del> <ins> >>> # Assuming optimizer uses lr = 0.05 for all groups </ins> >>> # lr = 0.05     if epoch < 30 >>> # lr = 0.005    if 30 <= epoch < 60 >>> # lr = 0.0005   if 60 <= epoch < 90", "pos": ["Hi, In the is written: \"lr values are [0.05, 0.005, ...].\" If I understand it correctly, it should be [0.5, 0.05, 0.005, ...].\nThat is correct. Please feel free to send a PR to fix it! (otherwise, I can)\nIf you don't mind to change it yourself, I'm not contributor yet and don't want to start just for one comment. I guess is the same case."], "neg": []}
{"query": "else: if exp_max > prec + 1 or exp_max < 0: sz = max(min_sz, 7) <del> scale = math.pow(10, exp_max - 1) </del> <ins> scale = math.pow(10, max(exp_max - 1, __MIN_LOG_SCALE)) </ins> else: if exp_max == 0: sz = 7", "pos": ["on print(torch.DoubleTensor(1)) I got ZeroDivisionError: float division by zero but torch.DoubleTensor(0)) or torch.DoubleTensor(2)) work just fine. I work in Jupyter notebook pytorch 0.1.12\nNote that returns undefined memory. Could you find the specific content of the tensor that gave this error, by printing ? I'm not sure if this bug has been fixed in the latest release already.\nthe error is pretty interesting. I restarted Jupyter notebook. # many freaking signatures: print(torch.FloatTensor(1)) print(torch.IntTensor(1)) print(torch.DoubleTensor(1)) # no error right after initial run. #print(torch.DoubleTensor(0)) print(torch.DoubleTensor(2)) print(torch.ByteTensor(1)) print(torch.CharTensor(1)) print(torch.ShortTensor(1)) print(torch.LongTensor(1)) steps to reproduce the error: 1). run, no errors 2) repeat, no error 3) uncomment ..DoubleTensor(0), run, got an error (see below): 4) comment ..DouleTensor(0), run still have this error 2.+24 [torch.FloatTensor of size 1] 7.+08 [torch.IntTensor of size 1] ZeroDivisionError Traceback (most recent call last) &lt;ipython-input-5-3629c250c505in &lt;module() 2 print(torch.FloatTensor(1)) 3 print(torch.IntTensor(1)) ----4 print(torch.DoubleTensor(1)) # ZeroDivisionError: float division by zero 5 #print(torch.DoubleTensor(0)) 6 print(torch.DoubleTensor(2)) C:UsersleogAppDataLocalcondacondaenvspy35libsite-packagestorch in str(self) 126 # characters to replace unicode characters with.", "127 if sys.versioninfo (3,): --128 return (self) 129 else: 130 if hasattr(, 'encoding'): C:UsersleogAppDataLocalcondacondaenvspy35libsite-packagestorch in str(self) 284 return '[{} with no dimension]n'.format(torch.typename(self)) 285 elif self.ndimension() == 1: --286 strt = vectorstr(self) 287 elif self.ndimension() == 2: 288 strt = matrixstr(self) C:UsersleogAppDataLocalcondacondaenvspy35libsite-packagestorch in vectorstr(self) 270 if () &lt; PRINTOPTS.threshold: 271 return (strt + --272 'n'.join(ident + (val / scale) for val in self) + 273 'n') 274 else: C:UsersleogAppDataLocalcondacondaenvspy35libsite-packagestorch in &lt;genexpr(.0) 270 if () &lt; PRINT_OPTS.threshold: 271 return (strt + --272 'n'.join(ident + (val / scale) for val in self) + 273 'n') 274 else: ZeroDivisionError: float division by zero\nOk, I got a repro. Here is the failing case: Fails with the error you mentioned. This seems to be an edge case though, as both and works. The issue is in , where becomes zero because is a too large negative number. Maybe we should clip to be positive?"], "neg": []}
{"query": "CMAKE_POLICY(SET CMP0026 OLD) ENDIF() <ins> IF(MSVC AND NOT \"${CMAKE_BUILD_TYPE}\" MATCHES \"Debug\") SET(MSVC_OPT_FLAG \"/Ox \") ELSE() SET(MSVC_OPT_FLAG \"\") ENDIF() </ins> ######################## # SET_SOURCE_FILES_PROPERTIES must be in the same CMakeLists.txt file as the target that includes the file # so we need to set these commands here rather than in src/TH IF(C_SSE4_1_FOUND AND C_SSE4_2_FOUND) IF(MSVC) <del> SET_SOURCE_FILES_PROPERTIES(${PROJECT_SOURCE_DIR}/src/TH/generic/simd/convolve5x5_sse.c PROPERTIES COMPILE_FLAGS \"/Ox /fp:fast\") </del> <ins> SET_SOURCE_FILES_PROPERTIES(${PROJECT_SOURCE_DIR}/src/TH/generic/simd/convolve5x5_sse.c PROPERTIES COMPILE_FLAGS \"${MSVC_OPT_FLAG}/fp:fast\") </ins> ELSE(MSVC) SET_SOURCE_FILES_PROPERTIES(${PROJECT_SOURCE_DIR}/src/TH/generic/simd/convolve5x5_sse.c PROPERTIES COMPILE_FLAGS \"-O3 -ffast-math\") ENDIF(MSVC) ENDIF(C_SSE4_1_FOUND AND C_SSE4_2_FOUND) IF(C_AVX_FOUND) IF(MSVC) <del> SET_SOURCE_FILES_PROPERTIES(${PROJECT_SOURCE_DIR}/src/TH/generic/simd/convolve5x5_avx.c PROPERTIES COMPILE_FLAGS \"/Ox /fp:fast ${C_AVX_FLAGS}\") SET_SOURCE_FILES_PROPERTIES(${PROJECT_SOURCE_DIR}/src/TH/vector/AVX.c PROPERTIES COMPILE_FLAGS \"/Ox /arch:AVX ${C_AVX_FLAGS}\") </del> <ins> SET_SOURCE_FILES_PROPERTIES(${PROJECT_SOURCE_DIR}/src/TH/generic/simd/convolve5x5_avx.c PROPERTIES COMPILE_FLAGS \"${MSVC_OPT_FLAG}/fp:fast ${C_AVX_FLAGS}\") SET_SOURCE_FILES_PROPERTIES(${PROJECT_SOURCE_DIR}/src/TH/vector/AVX.c PROPERTIES COMPILE_FLAGS \"${MSVC_OPT_FLAG}/arch:AVX ${C_AVX_FLAGS}\") </ins> ELSE(MSVC) SET_SOURCE_FILES_PROPERTIES(${PROJECT_SOURCE_DIR}/src/TH/generic/simd/convolve5x5_avx.c PROPERTIES COMPILE_FLAGS \"-O3 -ffast-math ${C_AVX_FLAGS}\") SET_SOURCE_FILES_PROPERTIES(${PROJECT_SOURCE_DIR}/src/TH/vector/AVX.c PROPERTIES COMPILE_FLAGS \"-O3 ${C_AVX_FLAGS}\")", "pos": ["The current master triggers the following error messages from Visual Studio 2015 (Debug build): 2------ Build started: Project: ATen, Configuration: Debug x64 ------ 2cl : Command line error D8016: '/Ox' and '/RTC1' command-line options are incompatible ... 14LINK : fatal error LNK1104: cannot open file '....srcATenDebug' It appears that they are triggered by *.c files in the ATen project, like \"srcTHvectorAVX.c\"\nWhy do you need Debug build?\nIsn't ATen meant to support both Debug and Release builds? FYI I'm involved with the project at which does support both build configurations, and we would like to be able to link our project with ATen.\nWell, you can remove the flags in as a workground.\nWe should patch this file to ignore this flag when building the library in debug mode. It shouldn't be too hard\nI catched similar errors , when I Tryed to build Debug Version I try to build Release and catched next errors e:pytorchatensrcatencpuvec256vec256int.h(184): note: see reference to function template instantiation 'at::vec256::anonymous-namespace'::Vec256<int16t::blend<1(at::vec256::anonymous-namespace'::Vec256<int16t)' being compiled 28e:pytorchatensrcatencpuvec256vec256int.h(147): error C3861: 'mm256extract_epi16': identifier not found What I should do?\nThere're two issues in your description. The first question is easy. You'll also have to specify the flag to 1. The second question is about your compiler. I guess you are using VS2015 or earlier versions of VS2017. Make sure the version in the output is 19.11 after your enter in CMD.\nYes I am using VS2015. My Compiler is 19.0.2. Must I use VS 2017 in any case?\nSorry, but the answer is yes. The AVX2 functions was in VS2017 and they are used in PyTorch. So VS2017 19.11 is a must.\nNo problem:) Thank you."], "neg": []}
{"query": "def inverse(self): return Inverse.apply(self) <ins> def gesv(self, a): return Gesv.apply(self, a) </ins> def multinomial(self, num_samples=1, with_replacement=False): return Multinomial(num_samples, with_replacement)(self)", "pos": ["There seems to be regression in the dynamo export of the MONAI VNet model as regards . Export script: The versions I'm using is torch 2.5.0.dev20240625+cu121 Previous torch versions such as were able to export this model successfully. Perhaps dynamo was generating FX graph without the expm1 node previously? cc: PyTorch version: 2.5.0.dev20240625+cu121 Is debug build: False CUDA used to build PyTorch: 12.1 ROCM used to build PyTorch: N/A OS: Ubuntu 22.04.4 LTS (x8664) GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 Clang version: Could not collect CMake version: version 3.29.0 Libc version: glibc-2.35 Python version: 3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0] (64-bit runtime) Python platform: Linux-6.5.0-35-generic-x8664-with-glibc2.35 Is CUDA available: True CUDA runtime version: 12.4.131 CUDAMODULELOADING set to: LAZY GPU models and configuration: GPU 0: NVIDIA RTX A5000 Nvidia driver version: 535.171.04 cuDNN version: Probably one of the following: /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- HIP runtime version: N/A MIOpen runtime version: N/A Is XNNPACK available: True CPU: Architecture: x8664 CPU op-mode(s): 32-bit, 64-bit Address sizes: 46 bits physical, 48 bits virtual Byte Order: Little Endian CPU(s): 20 On-line CPU(s) list: 0-19 Vendor ID: GenuineIntel Model name: Intel(R) Core(TM) i9-9820X CPU @ 3.", "30GHz CPU family: 6 Model: 85 Thread(s) per core: 2 Core(s) per socket: 10 Socket(s): 1 Stepping: 4 CPU max MHz: 4200.0000 CPU min MHz: 1200.0000 BogoMIPS: 6599.98 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constanttsc art archperfmon pebs bts repgood nopl xtopology nonstoptsc cpuid aperfmperf pni pclmulqdq dtes64 monitor dscpl vmx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse41 sse42 x2apic movbe popcnt tscdeadlinetimer aes xsave avx f16c rdrand lahflm abm 3dnowprefetch cpuidfault epb catl3 cdpl3 invpcidsingle pti ssbd mba ibrs ibpb stibp tprshadow flexpriority ept vpid eptad fsgsbase tscadjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm mpx rdta avx512f avx512dq rdseed adx smap clflushopt clwb intelpt avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqmllc cqmoccupllc cqmmbmtotal cqmmbmlocal dtherm ida arat pln pts hwp hwpactwindow hwpepp hwppkgreq vnmi mdclear flushl1d archcapabilities Virtualization: VT-x L1d cache: 320 KiB (10 instances) L1i cache: 320 KiB (10 instances) L2 cache: 10 MiB (10 instances) L3 cache: 16.5 MiB (1 instance) NUMA node(s): 1 NUMA node0 CPU(s): 0-19 Vulnerability Gather data sampling: Mitigation;", "Microcode Vulnerability Itlb multihit: KVM: Mitigation: VMX disabled Vulnerability L1tf: Mitigation; PTE Inversion; VMX conditional cache flushes, SMT vulnerable Vulnerability Mds: Mitigation; Clear CPU buffers; SMT vulnerable Vulnerability Meltdown: Mitigation; PTI Vulnerability Mmio stale data: Mitigation; Clear CPU buffers; SMT vulnerable Vulnerability Retbleed: Mitigation; IBRS Vulnerability Spec rstack overflow: Not affected Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and _user pointer sanitization Vulnerability Spectre v2: Mitigation; IBRS; IBPB conditional; STIBP conditional; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Mitigation; Clear CPU buffers; SMT vulnerable Versions of relevant libraries: [pip3] numpy==1.24.4 [pip3] onnx==1.16.0 [pip3] onnxscript==0.1.0.dev20240625 [pip3] optree==0.11.0 [pip3] pytorch-quantization==2.1.2 [pip3] pytorch-triton==3.0.0+ [pip3] torch==2.5.0.dev20240625+cu121 [pip3] torch-tensorrt==2.3.0a0 [pip3] torchaudio==2.4.0.dev20240625+cu121 [pip3] torchdata==0.7.1a0 [pip3] torchtext==0.17.0a0 [pip3] torchvision==0.20.0.dev20240625+cu121 [conda] Could not collect\nI am not very sure, but this looks like the kind of ops that should get decomp'ed. do you know? could you help take a look as part of the op support work you are doing? Thanks!\nYes, it looks like there is a graph-changed in dynamo, which dynamo now decomposes elu for some reasons. Using can preserve elu and then successfully export though.", "Yes can run but we would need to be resolved first since the weights are still not exported properly.\nI can take a look at this.\nshould fix this.", "Here, I can't export trivial net with sort() using dynamoexport(). Legacy () works fine, dynamoexport errors out : Unsupported FX nodes: {'callfunction': ['']}. Pytorch nightly 06/03 cc\na missing op, could you take this?\nNot sure if related:\nHmmm. Ideally, ONNX should add a sort op. It doesn't have one. I wonder what the legacy exporter converts it to.\nshould fix this issue.", "Are there any plans to implement something like tensorflow's and ? exists but it doesn't have the backward step implemented.\nOne workaround for the moment would be to write as , which should support backprop since . Adding support for should be trivial, the gradient for it can be found in\nI'll add backward support for , which should be equivalent to . I'll try to do it later today.\nHow difficult would it be to add in batched versions (where only the final two dimensions are square) of GESV and inverse? Are there any plans for this?\nIf we have batched version of , then adapting the autograd part is very simple. There is an open issue already that discuss implementing batched gesv\nActually, there is an open PR for that, so it might be coming soon\nI'll rebase the batched version of gesv. It'll come in soon\nMay I ask why gesv only support . As I saw from , it support that A is a MxN matrix, and B is a NxK matrix, then solve a matrix X (NxK). If I want to use a non-square matrix A in gesv, would you mind to let me know how to implement in PyTorch?\nthis doesn't exactly answer your question but you can add zero-rows to the A matrix until it becomes square.\nthanks a lot, I find the solution, can solve the problem.\nAny plan to add gradient calculation for gels?\nI am also wondering if will have differentiation support. My current workaround is to use instead\nI actually implemented overdetermined case following tensorflow implementation. For backward pass I only did basic grad test. On Monday, March 18, 2019, Johnson Zhong <:\nThanks, seems to work fine!"], "neg": []}
{"query": "Tensor forward(const Tensor& input) { this->_check_input_dim(input); <del> return F::detail::instance_norm( </del> <ins> return torch::nn::functional::detail::instance_norm( </ins> input, this->running_mean, this->running_var, this->weight, this->bias, this->is_training() || !this->options.track_running_stats(), this->options.momentum(), this->options.eps()); }", "pos": ["As part of pull request , F is redefined globally in nn/modules/batchnorm.h F is then used once in batchnorm.h and again in instancenorm.h not a bug, but a problem for anyone using F in some other way. it's solvable, but I don't see this pattern anywhere else in libtorch, was wondering if this use is an anomaly, and if we could use thanks cc\nThanks a lot for reporting! Yes this is an anomaly, and will be fixed by"], "neg": []}
{"query": ", dependencies() {} }; <del> struct FunctionTask { BackwardTask* base; std::shared_ptr<Function> fn; GradBuffer grad; FunctionTask(BackwardTask* base, std::shared_ptr<Function> fn, GradBuffer grad) : base(base) , fn(fn) , grad(std::move(grad)) {} }; </del> auto ReadyQueue::push_front(FunctionTask item) -> void { { std::lock_guard<std::mutex> lock(mutex);", "pos": ["Version include <stringin some cpp files: define order in The pre-defined struct is not passed by clang in macOS, so we should reorder them.\nSimple fix"], "neg": []}
{"query": "self.assertFalse(found_data[data_point_idx]) found_data[data_point_idx] += 1 break <del> self.assertEqual(target, self.labels.narrow(0, data_point_idx, 1)) </del> <ins> self.assertEqual(target, self.labels[data_point_idx]) </ins> found_labels[data_point_idx] += 1 self.assertEqual(sum(found_data.values()), (i + 1) * batch_size) self.assertEqual(sum(found_labels.values()), (i + 1) * batch_size)", "pos": ["TensorDataset adds casts 1D target tensor to 2D which is not acceptable by NLL loss layer.\nFixed in ."], "neg": []}
{"query": "self.assertRaises(AssertionError, nn.Embedding, num_embeddings=10, embedding_dim=20, padding_idx=25) self.assertRaises(AssertionError, nn.Embedding, num_embeddings=10, embedding_dim=20, padding_idx=-25) <ins> # test backward when input contains padding_idx padding_idx = 0 embedding = nn.Embedding(5, 2, padding_idx=padding_idx) for n in (1, 2): for other_indices in ([], [1, 3], [2]): indices = torch.LongTensor(other_indices + [padding_idx] * n) pre = embedding.weight[padding_idx].clone() embedding(indices).sum().backward() after = (embedding.weight + embedding.weight.grad)[padding_idx] embedding.zero_grad() self.assertEqual(after, pre) </ins> def test_embedding_max_norm(self): embedding = nn.Embedding(22, 5, max_norm=1.0) input = Variable(torch.LongTensor([2, 8, 8, 6]))", "pos": ["I have a straightforward use case. Given a token (or word), lookup the embedding of the word, pass that through a linear layer and perform L2 regression. If the token corresponds to the of the embedding layer, throws an error saying OS: macOS High Sierra PyTorch version: 0.3.1 How you installed PyTorch (conda, pip, source): Tried both conda and pip Python version: 3.6.4 CUDA/cuDNN version: CPU GPU models and configuration: NA GCC version (if compiling from source): NA A script to reproduce the bug:\nI got the same trouble, and it throw this error after some epochs of training. Does anyone have some solutions?"], "neg": []}
{"query": "super(TestCase, self).assertEqual(len(x), len(y), message) for x_, y_ in zip(x, y): self.assertEqual(x_, y_, prec, message) <ins> elif isinstance(x, bool) and isinstance(y, bool): super(TestCase, self).assertEqual(x, y, message) </ins> elif isinstance(x, Number) and isinstance(y, Number): if abs(x) == float('inf') or abs(y) == float('inf'): if allow_inf:", "pos": ["PyTorch version: 0.3.0 Python version: 3.5 GPU models and configuration: 8x NVIDIA Titan I found that requiresgrad setting in module is ignored when module is wrapped with nn.DataParallel. For example, This code should be executed because very small part of network activation (only fc layer) is stored for backward computation, but it results to run-time error: Internally, I found a crack in function which is in . In replicate function, it copies all parameter in module (# of replica times) with . In broadcasting code, it just defines new with default constructor parameter, which is always set to True. I think there can be some choices to fix this issue. is intended behavior for DataParallel, so we should use volatile (or ) to implement transfer learning in DataParallel model should be fixed to handle requires_grad when copying parameters. should be synchronizing in I think any of solution doesn't need lots of effort to fix it."], "neg": []}
{"query": "Attributes: weight_ih_l[k] : the learnable input-hidden weights of the k-th layer (W_ir|W_ii|W_in), of shape <del> (input_size x 3*hidden_size) </del> <ins> `(input_size x 3*hidden_size)` </ins> weight_hh_l[k] : the learnable hidden-hidden weights of the k-th layer (W_hr|W_hi|W_hn), of shape <del> (hidden_size x 3*hidden_size) bias_ih_l[k] : the learnable input-hidden bias of the k-th layer (b_ir|b_ii|b_in), of shape (3*hidden_size) bias_hh_l[k] : the learnable hidden-hidden bias of the k-th layer (W_hr|W_hi|W_hn), of shape (3*hidden_size) </del> <ins> `(hidden_size x 3*hidden_size)` bias_ih_l[k] : the learnable input-hidden bias of the k-th layer (b_ir|b_ii|b_in), of shape `(3*hidden_size)` bias_hh_l[k] : the learnable hidden-hidden bias of the k-th layer (W_hr|W_hi|W_hn), of shape `(3*hidden_size)` </ins> Examples:: >>> rnn = nn.GRU(10, 20, 2)", "pos": ["From : Happened because of some change in the environment as it is also affecting historical builds. cc\nThis issue seems to come from within libuv, a dependency of TensorPipe. First guess: there's two conflicting versions of libuv (e.g., a preinstalled one and the one shipped as a submodule of TensorPipe), and we build the one shipped by TensorPipe but somehow it ends up including the headers of the other one? In particular, the constant it's complaining about was in , which became part of libuv 1.41.0, which is the version that TensorPipe bundles. But it seems those jobs are installing version 1.40.0 via conda: So the question is why are the conda headers picked up and prioritized? I feel this will require some CMake digging, and it's possible that the fix will have to land in libuv's upstream CMake files. Ugh.\nBTW, if you want to validate this theory you can try setting this flag to : Which will cause TensorPipe to use the preinstalled libuv instead of building its own one. But please don't land this \"workaround\", as it will only lead to users coming to complain that they couldn't build PyTorch because their system has a libuv version that is too old for what TensorPipe needs.\nthis is definitely what is happening: libuv-1.40 is installed from conda, which does not have in the header: But I'm still confused why it started to happen just now [Edit] Can it be update that happened 4 hours ago?\nFinally found a PR that caused a regression: (and testing were clearly regressed in CI: )"], "neg": []}
{"query": "bidirectional=bidirectional, batch_first=batch_first) <del> outputs_cpu = forward_backward( False, rnn, input_val, hx_val, rnn.all_weights) rnn_gpu = module(input_size, hidden_size, num_layers, bias=bias, dropout=dropout, bidirectional=bidirectional, batch_first=batch_first) </del> <ins> outputs_gpu = forward_backward( True, rnn_gpu, input_val, hx_val, rnn.all_weights) </ins> <del> outputs_gpu = forward_backward( True, rnn_gpu, input_val, hx_val, rnn.all_weights) compare_cpu_gpu(outputs_cpu, outputs_gpu) </del> <ins> compare_cpu_gpu(outputs_cpu, outputs_gpu) </ins> for nonlinearity in ('tanh', 'relu'): hx_val = torch.randn(num_layers, batch, hidden_size)", "pos": ["cuDNN backend for RNN stabilizes at large mem usage. The following code stabilizes for me at 14+GiB reported by nvidia-smi. The same RNN without cuDNN uses <2.5GiB.\nOops... any idea what's going on with the garbage collector in this case?\nare you running this on a P100 GPU?\nYes, I observe similar behavior on Titan X Pascal\nhmmm, do cudnn RNNs have weird memory-oversubscribing properties? If it wasn't for CuDNN RNN, then it's not unreasonable. we use a custom CUDA allocator that will oversubscribe if needed, but will free the memory blocks it's holding onto if it hits OOM / failed allocations.\nTried the same RNN in torch and am getting ~2.6GiB on nvidia-smi.\nWith cuDNN I'm getting a constant usage of 2241MB on my machine using PyTorch (M40).\nare you using cudnnv5, or v6?\nv5\nno, cudnn does not do memory oversubscribing, and already tried the same network in lua torch with cudnn, where it gets expected memory usage. Also, he is getting OOMs on slightly bigger networks, where apparently caching allocator should have tried to free previous blocks. Interesting that can not repro this.\nJust to make sure did you set when you ran? I left it as False in the repro.\nI can repro in the docker image built from the current version on P100, and I go OOM after 8 batches with the same image on 1080, which generally should not happen - if there is enough memory to go through 1 batch, memory usage should not increase afterwards.\nNo, I just copied your script and ran it. Weird thing is that it goes OOM after 18 batches once I enable . Without it memory remains stable.\nA lot of cudaMalloc allocations go not through THCudaMalloc, and never get freed:\nSo not going through THCudaMalloc is actually expected for storages allocated with THCStorage_(newWithSize), but what's not expected is to every time cudaMalloc bytes of rnn workspace, instead of reusing earlier allocations. (~784 MB allocations are for workspace, ~156MB allocations happen somewhere in criterion).\nMy guess is that the bug is at the Python level.", "Either: Something is holding on to a Tensor longer than it should Something is creating a circular reference which prevents quick clean-up\nThe issue is that in creates cyclic references. It's made worse by the fact that it holds on to more memory than it should (like the workspace). You can see this by adding a as the first line in your for loop. With the explicit call to the GC every iteration, the memory usage doesn't explode.\nhow can these cyclic references be fixed?\nNot sure yet, but I will fix it tomorrow with"], "neg": []}
{"query": "// torch.SparseTensor() if (num_args == 0) { self->cdata = THSTensor_(new)(LIBRARY_STATE_NOARGS); <del> return (PyObject*)self.release(); } PyObject *first_arg = PyTuple_GET_ITEM(args, 0); </del> <ins> } else { PyObject *first_arg = PyTuple_GET_ITEM(args, 0); </ins> <del> // torch.SparseTensor(size) if (num_args == 1 && THPUtils_checkLong(first_arg)) { long size = THPUtils_unpackLong(first_arg); self->cdata = THSTensor_(newWithSize1d)(LIBRARY_STATE size); } // torch.SparseTensor(torch.Size sizes) else if (num_args == 1 && THPSize_Check(first_arg)) { THLongStoragePtr sizes(THPUtils_unpackSize(first_arg)); self->cdata = THSTensor_(newWithSize)(LIBRARY_STATE sizes.get()); } // torch.SparseTensor(torch.LongTensor indices, torch.LongTensor values) else if (num_args == 2 && THPIndexTensor_Check(first_arg)) { PyObject *second_arg = PyTuple_GET_ITEM(args, 1); if (!THPTensor_(Check)(second_arg)) goto invalid_arguments; </del> <ins> // torch.SparseTensor(size) if (num_args == 1 && THPUtils_checkLong(first_arg)) { long size = THPUtils_unpackLong(first_arg); self->cdata = THSTensor_(newWithSize1d)(LIBRARY_STATE size); } // torch.SparseTensor(torch.Size sizes) else if (num_args == 1 && THPSize_Check(first_arg)) { THLongStoragePtr sizes(THPUtils_unpackSize(first_arg)); self->cdata = THSTensor_(newWithSize)(LIBRARY_STATE sizes.get()); } // torch.SparseTensor(torch.LongTensor indices, torch.LongTensor values) else if (num_args == 2 && THPIndexTensor_Check(first_arg)) { PyObject *second_arg = PyTuple_GET_ITEM(args, 1); if (!THPTensor_(Check)(second_arg)) goto invalid_arguments; </ins> <del> THIndexTensor *indices = ((THPIndexTensor*)first_arg)->cdata; THTensor *values = ((THPTensor*)second_arg)->cdata; </del> <ins> THIndexTensor *indices = ((THPIndexTensor*)first_arg)->cdata; THTensor *values = ((THPTensor*)second_arg)->cdata; </ins> #ifdef THC_GENERIC_FILE <del> THCAssertSameGPU(THSTensor_(checkGPU)(LIBRARY_STATE 0, 2, indices, values)); </del> <ins> THCAssertSameGPU(THSTensor_(checkGPU)(LIBRARY_STATE 0, 2, indices, values)); </ins> #endif <del> self->cdata = THSTensor_(newWithTensor)(LIBRARY_STATE indices, values); } // torch.SparseTensor(torch.LongTensor indices, //                    torch.Tensor values, //                    torch.Size sizes) else if (num_args > 2 && THPIndexTensor_Check(first_arg)) { PyObject *second_arg = PyTuple_GET_ITEM(args, 1); PyObject *third_arg = PyTuple_GET_ITEM(args, 2); if (!THPTensor_(Check)(second_arg)) goto invalid_arguments; if (!THPSize_Check(third_arg)) goto invalid_arguments; </del> <ins> self->cdata = THSTensor_(newWithTensor)(LIBRARY_STATE indices, values); } // torch.SparseTensor(torch.LongTensor indices, //                    torch.Tensor values, //                    torch.Size sizes) else if (num_args > 2 && THPIndexTensor_Check(first_arg)) { PyObject *second_arg = PyTuple_GET_ITEM(args, 1); PyObject *third_arg = PyTuple_GET_ITEM(args, 2); if (!THPTensor_(Check)(second_arg)) goto invalid_arguments; if (!THPSize_Check(third_arg)) goto invalid_arguments; </ins> <del> THIndexTensor *indices = ((THPIndexTensor*)first_arg)->cdata; THTensor *values = ((THPTensor*)second_arg)->cdata; THLongStoragePtr sizes(THPUtils_unpackSize(third_arg)); </del> <ins> THIndexTensor *indices = ((THPIndexTensor*)first_arg)->cdata; THTensor *values = ((THPTensor*)second_arg)->cdata; THLongStoragePtr sizes(THPUtils_unpackSize(third_arg)); </ins> #ifdef THC_GENERIC_FILE <del> THCAssertSameGPU(THSTensor_(checkGPU)(LIBRARY_STATE 0, 2, indices, values)); </del> <ins> THCAssertSameGPU(THSTensor_(checkGPU)(LIBRARY_STATE 0, 2, indices, values)); </ins> #endif <del> self->cdata = THSTensor_(newWithTensorAndSize)( LIBRARY_STATE indices, values, sizes); } // torch.SparseTensor(int ...) else if (THPUtils_tryUnpackLongVarArgs(args, 0, sizes)) { self->cdata = THSTensor_(newWithSize)(LIBRARY_STATE sizes.get()); </del> <ins> self->cdata = THSTensor_(newWithTensorAndSize)( LIBRARY_STATE indices, values, sizes); } // torch.SparseTensor(int ...) else if (THPUtils_tryUnpackLongVarArgs(args, 0, sizes)) { self->cdata = THSTensor_(newWithSize)(LIBRARY_STATE sizes.get()); } else goto invalid_arguments; // All other cases </ins> } <del> else goto invalid_arguments; // All other cases THSPTensor_(initStorage)(self->cdata); </del> <ins> THSTensor_(initStorage)(self->cdata); </ins> return (PyObject*)self.release(); invalid_arguments:", "pos": ["This causes to not work properly on newly created empty sparse tensors. In the example below, both should print 1 instead."], "neg": []}
{"query": "flat_weight = first_data.new().set_(first_data.storage(), 0, torch.Size([self._param_buf_size])) else: flat_weight = None <ins> self.check_forward_args(input, hx, batch_sizes) </ins> func = self._backend.RNN( self.mode, self.input_size,", "pos": ["I run RNN sucessfully with the code. and I run embedding example. Then, got RuntimeError. I set to use nothing seems wrong. And the error info seems not the dim problem? The error is so strange.\nWhat exactly do you mean by \"embedding example\"? Could you provide a code snippet that causes the RuntimeError?\nTo better reproduce the error, I use the official code example of The origin runs well: Due to use , I have to make batch first. And then , it raises error.\nThank you for the example! The say that setting means that the first dim of input is the batch, ie, input has size . however is unchanged, so the following should work: I do agree the error message could be better, though"], "neg": []}
{"query": "return platform <ins> def get_python_platform(): import platform return platform.platform() def get_libc_version(): import platform if get_platform() != 'linux': return 'N/A' return '-'.join(platform.libc_ver()) </ins> def get_pip_packages(run_lambda): \"\"\"Returns `pip list` output. Note: will also find conda-installed pytorch and numpy packages.\"\"\"", "pos": ["TensorPipe uses some advanced features of the Linux kernel. It does so on a \"best effort\" basis (i.e., use them if available, fall back on something else if not). This logic is unfortunately hard to test because it would require specific kernel versions on our CI machines. So it occasionally happens that we find out about issues when users report them. To help investigate these reports we've found it useful to know the kernel version and the glibc version used by the user (both when building and when running, if they're different). It would be great if these could be to the report produced by collect_env.\nCC"], "neg": []}
{"query": "{ /* Get normalized cumulative distribution from prob distribution */ double sum = 0; <ins> double val; </ins> for (j=0; j<n_categories; j++) { <del> sum += THStorage_(get)(  </del> <ins> val = THStorage_(get)(  </ins> prob_dist->storage,  prob_dist->storageOffset+i*prob_dist->stride[0]+j*prob_dist->stride[1]  ); <ins> THArgCheckWithCleanup((val >= 0), THCleanup(THDoubleTensor_free(cum_dist); if (start_dim == 1) THTensor_(resize1d)(prob_dist, n_categories);), 2, \"invalid multinomial distribution (encountering probability entry < 0)\"); sum += val; </ins> THDoubleStorage_set( cum_dist->storage,  cum_dist->storageOffset+j*cum_dist->stride[0], ", "pos": ["Looking at the values returned from tn.multinomial(1) and counting them, it seems like it returns the values of the multinomial distribution taken from t, i.e. that it has ignored the signs of the values. However, it seem like always returns a tensor filled with zeros.\nlooking into this"], "neg": []}
{"query": "dict( module_name='AdaptiveMaxPool1d', constructor_args=(3,), <del> input=torch.rand(1, 3, 5), </del> <ins> input=_rand_tensor_non_equal(1, 3, 5), </ins> ), dict( module_name='AdaptiveMaxPool2d', constructor_args=(3,), <del> input=torch.rand(1, 3, 5, 6), </del> <ins> input=_rand_tensor_non_equal(1, 3, 5, 6), </ins> desc='single', ), dict( module_name='AdaptiveMaxPool2d', constructor_args=((3, 4),), <del> input=torch.rand(1, 3, 5, 6), </del> <ins> input=_rand_tensor_non_equal(1, 3, 5, 6), </ins> desc='tuple', ), dict( module_name='AdaptiveMaxPool3d', constructor_args=(3,), <del> input=torch.rand(2, 3, 5, 6, 7), </del> <ins> input=_rand_tensor_non_equal(2, 3, 5, 6, 7), </ins> desc='single', ), dict( module_name='AdaptiveMaxPool3d', constructor_args=((3, 4, 5),), <del> input=torch.rand(2, 3, 5, 6, 7), </del> <ins> input=_rand_tensor_non_equal(2, 3, 5, 6, 7), </ins> desc='tuple', ), dict( module_name='AdaptiveMaxPool3d', constructor_args=(3,), <del> input=torch.rand(2, 3, 12, 9, 3), </del> <ins> input=_rand_tensor_non_equal(2, 3, 12, 9, 3), </ins> desc='single_nonatomic', ), dict( module_name='AdaptiveMaxPool3d', constructor_args=((3, 4, 5),), <del> input=torch.rand(2, 3, 6, 4, 10), </del> <ins> input=_rand_tensor_non_equal(2, 3, 6, 4, 10), </ins> desc='tuple_nonatomic', ), dict(", "pos": ["Intermittent test failures occurring on AdaptiveMaxPool3d. and I have seen this in the past where Max pooling has exactly the same values in the same window, then it's differently resolved on CPU/GPU. For example [0 2.5 2.5 3] can give a max index of or on CUDA depending on the runtime. So generating input values that are within epsilon of each other is important (especially at half precision). fix that one.\nWill do. Thanks for identifying the issue."], "neg": []}
{"query": "'symmetric_pd', NO_ARGS, [skipIfNoLapack]), ('logdet', lambda: make_nonzero_det(random_fullrank_matrix_distinct_singular_value(S), 1, 0), NO_ARGS, 'distinct_singular_values', NO_ARGS, [skipIfNoLapack]), <del> ('slogdet', lambda: make_nonzero_det(torch.randn(1, 1), 1), NO_ARGS, '1x1_pos_det', NO_ARGS, [skipIfNoLapack], [1]), </del> <ins> ('slogdet', lambda: make_nonzero_det(torch.randn(1, 1), 1), NO_ARGS, '1x1_pos_det', NO_ARGS, [skipIfNoLapack], itemgetter(1)), </ins> ('slogdet', lambda: make_nonzero_det(torch.randn(1, 1), -1), NO_ARGS, <del> '1x1_neg_det', NO_ARGS, [skipIfNoLapack], [1]), ('slogdet', lambda: make_nonzero_det(torch.randn(S, S), 1), NO_ARGS, 'pos_det', NO_ARGS, [skipIfNoLapack], [1]), ('slogdet', lambda: make_nonzero_det(torch.randn(S, S), -1), NO_ARGS, 'neg_det', NO_ARGS, [skipIfNoLapack], [1]), </del> <ins> '1x1_neg_det', NO_ARGS, [skipIfNoLapack], itemgetter(1)), ('slogdet', lambda: make_nonzero_det(torch.randn(S, S), 1), NO_ARGS, 'pos_det', NO_ARGS, [skipIfNoLapack], itemgetter(1)), ('slogdet', lambda: make_nonzero_det(torch.randn(S, S), -1), NO_ARGS, 'neg_det', NO_ARGS, [skipIfNoLapack], itemgetter(1)), </ins> ('slogdet', lambda: make_nonzero_det(random_symmetric_matrix(S)), NO_ARGS, <del> 'symmetric', NO_ARGS, [skipIfNoLapack], [1]), ('slogdet', lambda: random_symmetric_pd_matrix(S), NO_ARGS, 'symmetric_pd', NO_ARGS, [skipIfNoLapack], [1]), </del> <ins> 'symmetric', NO_ARGS, [skipIfNoLapack], itemgetter(1)), ('slogdet', lambda: random_symmetric_pd_matrix(S), NO_ARGS, 'symmetric_pd', NO_ARGS, [skipIfNoLapack], itemgetter(1)), </ins> ('slogdet', lambda: random_fullrank_matrix_distinct_singular_value(S), NO_ARGS, <del> 'distinct_singular_values', NO_ARGS, [skipIfNoLapack], [1]), </del> <ins> 'distinct_singular_values', NO_ARGS, [skipIfNoLapack], itemgetter(1)), </ins> ('svd', lambda: random_fullrank_matrix_distinct_singular_value(S), NO_ARGS, '', NO_ARGS, [skipIfNoLapack]), <del> ('svd', lambda: random_fullrank_matrix_distinct_singular_value(M), NO_ARGS, 'large', NO_ARGS, [skipIfNoLapack]), </del> <ins> ('svd', lambda: random_fullrank_matrix_distinct_singular_value(S)[:(S - 2)], NO_ARGS, 'wide', NO_ARGS, [skipIfNoLapack]), ('svd', lambda: random_fullrank_matrix_distinct_singular_value(S)[:, :(S - 2)], NO_ARGS, 'tall', NO_ARGS, [skipIfNoLapack]), ('svd', lambda: random_fullrank_matrix_distinct_singular_value(S)[:(S - 2)], (False,), 'wide_all', NO_ARGS, [skipIfNoLapack], lambda usv: (usv[0], usv[1], usv[2][:, :(S - 2)])), ('svd', lambda: random_fullrank_matrix_distinct_singular_value(S)[:, :(S - 2)], (False,), 'tall_all', NO_ARGS, [skipIfNoLapack], lambda usv: (usv[0][:, :(S - 2)], usv[1], usv[2])), ('svd', lambda: random_fullrank_matrix_distinct_singular_value(M), NO_ARGS, 'large', NO_ARGS, [skipIfNoLapack]), </ins> ('gesv', (S, S), ((S, S),), '', NO_ARGS, [skipIfNoLapack]), ('fill_', (S, S, S), (1,), 'number'), ('fill_', (), (1,), 'number_scalar'),", "pos": ["I am having problem in running the backward propagation with svd. In particular, it seems working when the first dimension of the matrix x (below) is greater than the second. But I get the following error if the first is smaller than the second. PyTorch version: 2018.04.20 Is debug build: No CUDA used to build PyTorch: 8.0.61 OS: Ubuntu 16.04.4 LTS GCC version: (Ubuntu 5.4.0-6ubuntu1~16.04.9) 5.4.0 CMake version: Could not collect Python version: 3.5 Is CUDA available: No CUDA runtime version: No CUDA GPU models and configuration: No CUDA Nvidia driver version: No CUDA cuDNN version: No CUDA Versions of relevant libraries: [pip3] numpy (1.13.3) [pip3] torch (2018.4.20) [pip3] torchvision (0.2.0) [conda] pytorch 0.3.1 py35cuda8.0.61cudnn7.0.52 pytorch [conda] pytorch-nightly 2018.04.20 py35cuda8.0.61cudnn7.1.21 pytorch [conda] torchvision 0.2.0 py35heaa392f_1 pytorch\nRepro'ed on 0.4 branch. works btw.\nWhat does \"Repro'ed\" means? Should I install from that branch? I need to compute the kernel of a matrix, therefore I need the last columns of the matrix V (which is why I have to pass ).\nIt's a shorthand for \"reproduced\", which means that the bug still hasn't been fixed. Sorry, we'll try to get to it soon."], "neg": []}
{"query": "modules = self.__dict__['_modules'] if name in modules: return modules[name] <del> return object.__getattribute__(self, name) </del> <ins> return object.__getattr__(self, name) </ins> def __setattr__(self, name, value): <ins> def remove_from(*dicts): for d in dicts: if name in d: del d[name] </ins> params = self.__dict__.get('_parameters') if isinstance(value, Parameter): if params is None: raise AttributeError( \"cannot assign parameters before Module.__init__() call\") <ins> remove_from(self.__dict__, self._buffers, self._modules) </ins> self.register_parameter(name, value) elif params is not None and name in params: if value is not None:", "pos": ["They do one pair of these calls for each slicing operation.\nDoes this actually have a performance impact? These calls tend to be very quick\nYes, it's a major reason why small unrolled LSTMs (which involve in addition to two GEMMs: a chunk into four gates, for 4 pairs of calls, five unary pointwise ops, for 5 more, and five binary pointwise ops, for another 5) are much slower than they are in Theano and TensorFlow.\nAlso, there are excessive cudaGetDevice calls for each pointwise operation (9 cudaGetDevice) even without chunking, as in the following repro code: As for small unrolled LSTMs it does show up.\nThis should no longer be an issue, now that is implemented as a native ATen function. I think we're still using the Python version for tensors, but this will be resolved soon when we stop using tensors altogether."], "neg": []}
{"query": "grad.zero_() def updateParameters(self, learningRate): <del> params, gradParams = self.parameters() if params: for p, gp in zip(params, gradParams): p.add_(-learningRate, gp) </del> <ins> if self.parameters() is not None: params, gradParams = self.parameters() if params: for p, gp in zip(params, gradParams): p.add_(-learningRate, gp) </ins> def training(self): self.train = True", "pos": ["I implemented a model using because I have to do particular things and I'm starting from a PyTorch implementation. I think this method may be broken. If I run this code: I get this error: This happens whenever the Sequential model has layers without parameters. Removing ReLUs and LogSoftMax solves the problem. I think this happens because returns for s which don't have parameters and raises the error. I tried this both on Python2.7 and Python 3.5 and it's the same.\nSolved by"], "neg": []}
{"query": "CMAKE_PREFIX_PATH=\"$(dirname $(which conda))/../\"  pip install -v . <ins> RUN git clone https://github.com/pytorch/vision.git && cd vision && pip install -v . </ins> WORKDIR /workspace RUN chmod -R a+w /workspace", "pos": ["Torchvision is considered a basic requirement of the tutorials. Perhaps it makes sense to include it in the docker build.\nIt's easy to install it by command ;)\nIt's easy to install torch itself using a similar command? :). I was thinking a docker image should have torchvision included.\nRuntime docker file already has torchvision."], "neg": []}
{"query": "#define THSTensorPtr  TH_CONCAT_3(THS,Real,TensorPtr) #define THSPTensorPtr  TH_CONCAT_3(THSP,Real,TensorPtr) <del> template<class T> class THPPointer { public: THPPointer(): ptr(nullptr) {}; THPPointer(T *ptr): ptr(ptr) {}; THPPointer(THPPointer &&p) { free(); ptr = p.ptr; p.ptr = nullptr; }; ~THPPointer() { free(); }; T * get() { return ptr; } T * release() { T *tmp = ptr; ptr = NULL; return tmp; } operator T*() { return ptr; } THPPointer& operator =(T *new_ptr) { free(); ptr = new_ptr; return *this; } THPPointer& operator =(THPPointer &&p) { free(); ptr = p.ptr; p.ptr = nullptr; return *this; } T * operator ->() { return ptr; } operator bool() { return ptr != nullptr; } private: void free(); T *ptr = nullptr; }; typedef THPPointer<PyObject> THPObjectPtr; </del> typedef THPPointer<THPGenerator> THPGeneratorPtr; template <typename T>", "pos": ["Go to Look at the html. It contains the following: ! This problem manifests in google search. Searching for 1.8.1 docs brings up the \"1.8.0 docs\": ! We can probably fix this by doing a quick sed on the website but we should develop a solution for future releases. I think this has happened before in the past and am not sure what causes it. cc"], "neg": []}
{"query": "output_var_.requires_grad = self->cdata.requires_grad; auto dummy_prev_fn = std::make_shared<Variable>( std::unique_ptr<Tensor>(output_var_.data->clone_shallow()), false, false); <del> self->cdata.previous_functions[i] = std::make_pair<>(dummy_prev_fn, 0); </del> <ins> // Replace all references to the variable auto& previous_functions = self->cdata.previous_functions; for (int inp = 0; inp < self->num_inputs; inp++) { if (previous_functions[inp].first.get() == &output_var_) { previous_functions[inp] = std::make_pair<>(dummy_prev_fn, 0); } } </ins> } else { // output_var_.requires_grad throw std::runtime_error(\"a leaf Variable that requires grad has been used in an in-place operation.\"); }", "pos": ["import torch as th from torch.autograd import Variable x = Variable(th.Tensor([1 , 2 , 3]) , requiresgrad = False) y = Variable(th.Tensor([100]) , requiresgrad = True) x = x.cuda(0) y = y.cuda(0) x[0] = y print(x.requiresgrad) If ((0) (0)) is executed then the output is False, True otherwise. I think the result that x.requiresgrad = True is more natural.\nI further checked some in-place operator in GPU mode(say, x = Variable(...).cuda(0) ), and find out that after the operation is done, the requiresgrad of the Variable doesn't change. For example: import torch as th from torch.autograd import Variable x = Variable(torch.Tensor([1]) , requiresgrad = False) y = Variable(torch.Tensor([2]) , requiresgrad = True) x = x.cuda(0) y = y.cuda(0) x.add(y) print(x.requires_grad) the output is False\nIt doesn't have to do with Variable holding a CUDA tensor, but rather with the fact that it's no longer a leaf. I've sent a fix in .\nThanks for the report!"], "neg": []}
{"query": "#include <vector> #include <stack> #include <tuple> <del> #include \"THCP.h\" </del> <ins> #include \"torch/csrc/cuda/THCP.h\" </ins> <del> #include \"override_macros.h\" </del> <ins> #include \"torch/csrc/cuda/override_macros.h\" </ins> #include \"torch/csrc/copy_utils.h\" #include \"DynamicTypes.h\" <del> #define THC_GENERIC_FILE \"torch/csrc/generic/Tensor.cpp\" #include <THC/THCGenerateAllTypes.h> </del> <ins> //generic_include THC torch/csrc/generic/Tensor.cpp </ins> <del> #include \"undef_macros.h\" #include \"restore_macros.h\" </del> <ins> #include \"torch/csrc/cuda/undef_macros.h\" #include \"torch/csrc/cuda/restore_macros.h\" </ins>", "pos": ["Trying to run on github actions via AWS instance gives me: I will of course set up now, but it shouldn't crash like that. May be it should use as a fallback? pt-2.0 cc\nIs this a multitenant setup? Also curious which distro are you running, I was under the impression that most linux distros set home by default\nthis is AWS image. one default user. I think it's user, but it wants you to run as root, which is weird. I have manually now set to overcome this.\nWe should update triton to use pathlib or something"], "neg": []}
{"query": "THArgCheck(THCTensor_(nDimension)(state, src) <= MAX_CUTORCH_DIMS, 2, CUTORCH_DIM_WARNING); THArgCheck(THCudaLongTensor_nDimension(state, index) <= MAX_CUTORCH_DIMS, 2, CUTORCH_DIM_WARNING); <del> int srcDims = THCTensor_(nDimension)(state, src); THArgCheck(srcDims > 0, 2, \"Source tensor is empty\"); </del> <ins> if (numIndices == 0) { return; } </ins> if (accumulate) { // wrap indices so to replace negative indices", "pos": ["odd difference between cuda Tensor and Variable: encountered in torch.distributions.Categorical, built yesterday from master\nrelated to hmm, here's the issue is for variables, so probably not (immediately) related\ndifferent code path but similar issue\nThis is a regression from the latest 0.3 binaries, which don't have this error."], "neg": []}
{"query": "`U, S, V = torch.svd(A)` returns the singular value decomposition of a real matrix `A` of size `(n x m)` such that :math:`A = USV^T`. <del> `U` is of shape :math:`(n times min(n, m))`. </del> <ins> `U` is of shape :math:`(n times n)`. </ins> <del> `S` is a diagonal matrix of shape :math:`(min(n, m) times min(n, m))`, represented as a vector of size :math:`min(n, m)` containing the diagonal entries. </del> <ins> `S` is a diagonal matrix of shape :math:`(n times m)`, represented as a vector of size :math:`min(n, m)` containing the diagonal entries. </ins> <del> `V` is of shape :math:`(m times min(n, m))`. </del> <ins> `V` is of shape :math:`(m times m)`. </ins> <del> :attr:`some` represents the number of singular values to be computed. If `some=True`, it computes some and `some=False` computes all. </del> <ins> If :attr:`some` is ``True`` (default), the returned `U` and `V` matrices will contain only :math:`min(n, m)` orthonormal columns. </ins> .. note:: Irrespective of the original strides, the returned matrix `U` will be transposed, i.e. with strides `(1, n)` instead of `(n, 1)`.", "pos": ["The documentation of the parameter to is one of the strangest I've seen... what does (the default) do???\nHa it is somewhat weird indeed. When , the returned U and V matrices are square orthonormal matrices. I will send out a PR to clarify this."], "neg": []}
{"query": "return self.clone(); } <del> std::vector<int64_t> to_arg_sizes(TensorList tensors, int64_t dim) { </del> <ins> static std::vector<int64_t> to_arg_sizes(TensorList tensors, int64_t dim) { </ins> std::vector<int64_t> arg_sizes(tensors.size()); for (size_t i = 0; i < tensors.size(); ++i) { arg_sizes[i] = tensors[i].size(dim);", "pos": ["See Docker run fails with , which is due to the fact that is specified in docker container: Which results in: CI cc\nCheck that adding to manywheels image results in a failure:\nTentatively marking hi-pri since nightly testing are failing\nThis is resolved, here is the successful run:"], "neg": []}
{"query": "self._test_broadcast_helper(group, group_id, rank) @unittest.skipIf(BACKEND != 'gloo', \"Only Gloo backend supports CUDA allReduce\") <ins> @skip_if_no_cuda_distributed </ins> def test_broadcast_cuda(self): group, group_id, rank = self._init_global_test() self._test_broadcast_helper(group, group_id, rank, True)", "pos": ["is causing a crash when we try to build pytorch. Crash looks like this: (all times in US/West) Our 7AM, 8AM builds and earlier were all fine but 9AM, 10AM, 11AM builds all crashed with that same traceback. The new import was in which I think went in between 8AM and 9AM today See command below: See here if needed for : <!-- A clear and concise description of what you expected to happen. --Please copy and paste the output from our (or fill out the checklist below manually). You can get the script and run it with: PyTorch Version (e.g., 1.0): building from master OS (e.g., Linux): linux How you installed PyTorch (, , source): source Build command you used (if compiling from source): see above under 'reproduce' Python version: 3.6\nmaybe easier to read for build command used:\nplease try in the CircleCi environment?\nin our build script seems to have fixed this. Thanks !"], "neg": []}
{"query": "self.assertEqual(res1, res2, 0) self.assertEqual(res1, res3, 0) <ins> def _test_overdetermined(a, b, expectedNorm): m = a.size()[0] n = a.size()[1] assert(m > n) def check_norm(a, b, expected_norm, gels_result): # Checks |ax - b| and the residual info from the result n = a.size()[1] # The first n rows is the least square solution. # Rows n to m-1 contain residual information. x = gels_result[:n] resid_info = gels_result[n:] resid_norm = (torch.mm(a, x) - b).norm() self.assertEqual(resid_norm, expectedNorm, 1e-8) self.assertEqual(resid_info.norm(), resid_norm, 1e-8) a_copy = a.clone() b_copy = b.clone() res1 = torch.gels(b, a)[0] self.assertEqual(a, a_copy, 0) self.assertEqual(b, b_copy, 0) check_norm(a, b, expectedNorm, res1) ta = torch.Tensor() tb = torch.Tensor() res2 = torch.gels(b, a, out=(tb, ta))[0] self.assertEqual(a, a_copy, 0) self.assertEqual(b, b_copy, 0) check_norm(a, b, expectedNorm, res2) res3 = torch.gels(b, a, out=(b, a))[0] check_norm(a_copy, b_copy, expectedNorm, res3) self.assertEqual(res1, tb, 0) self.assertEqual(res1, b, 0) self.assertEqual(res1, res2, 0) self.assertEqual(res1, res3, 0) </ins> # basic test expectedNorm = 0 a = torch.Tensor(((1.44, -9.96, -7.55, 8.34),", "pos": ["b = (600,1) A = (600,2) X, = (b,A) X.size() #torch.Size([2, 1]) Y, = (b.cuda(),A.cuda()) Y.size() #torch.Size([600, 1])\nthat is interesting. says that the first rows returned ( here) include the solution of the least-squares problem. I've verified that the first 2 rows of X and Y are the same so you can probably use that for now while we figure out what the rest of the output is."], "neg": []}
{"query": "static struct PyGetSetDef conv_forward_properties[] = { THP_FUNCTION_DEFAULT_PROPERTIES, {(char*)\"stride\", (getter)getTupleAttr<ConvForward, std::vector<int>, ConvParams, <del> &ConvParams::stride, int64_t, PyInt_FromLong>, NULL, NULL, NULL}, </del> <ins> &ConvParams::stride, int64_t, PortablePyInt_FromLong>, NULL, NULL, NULL}, </ins> {(char*)\"padding\", (getter)getTupleAttr<ConvForward, std::vector<int>, ConvParams, <del> &ConvParams::padding, int64_t, PyInt_FromLong>, NULL, NULL, NULL}, </del> <ins> &ConvParams::padding, int64_t, PortablePyInt_FromLong>, NULL, NULL, NULL}, </ins> {(char*)\"dilation\", (getter)getTupleAttr<ConvForward, std::vector<int>, ConvParams, <del> &ConvParams::dilation, int64_t, PyInt_FromLong>, NULL, NULL, NULL}, </del> <ins> &ConvParams::dilation, int64_t, PortablePyInt_FromLong>, NULL, NULL, NULL}, </ins> {(char*)\"transposed\", (getter)getValueAttr<ConvForward, bool, ConvParams, &ConvParams::transposed, long, PyBool_FromLong>, NULL, NULL, NULL}, {(char*)\"output_padding\", (getter)getTupleAttr<ConvForward, std::vector<int>, ConvParams, <del> &ConvParams::output_padding, int64_t, PyInt_FromLong>, NULL, NULL, NULL}, </del> <ins> &ConvParams::output_padding, int64_t, PortablePyInt_FromLong>, NULL, NULL, NULL}, </ins> {(char*)\"groups\", (getter)getValueAttr<ConvForward, int, ConvParams, <del> &ConvParams::groups, int64_t, PyInt_FromLong>, NULL, NULL, NULL}, </del> <ins> &ConvParams::groups, int64_t, PortablePyInt_FromLong>, NULL, NULL, NULL}, </ins> {NULL} }; static struct PyGetSetDef conv_backward_properties[] = { THP_FUNCTION_DEFAULT_PROPERTIES, {(char*)\"stride\", (getter)getTupleAttr<ConvBackward, std::vector<int>, ConvParams, <del> &ConvParams::stride, int64_t, PyInt_FromLong>, NULL, NULL, NULL}, </del> <ins> &ConvParams::stride, int64_t, PortablePyInt_FromLong>, NULL, NULL, NULL}, </ins> {(char*)\"padding\", (getter)getTupleAttr<ConvBackward, std::vector<int>, ConvParams, <del> &ConvParams::padding, int64_t, PyInt_FromLong>, NULL, NULL, NULL}, </del> <ins> &ConvParams::padding, int64_t, PortablePyInt_FromLong>, NULL, NULL, NULL}, </ins> {(char*)\"dilation\", (getter)getTupleAttr<ConvBackward, std::vector<int>, ConvParams, <del> &ConvParams::dilation, int64_t, PyInt_FromLong>, NULL, NULL, NULL}, </del> <ins> &ConvParams::dilation, int64_t, PortablePyInt_FromLong>, NULL, NULL, NULL}, </ins> {(char*)\"transposed\", (getter)getValueAttr<ConvBackward, bool, ConvParams, &ConvParams::transposed, long, PyBool_FromLong>, NULL, NULL, NULL}, {(char*)\"output_padding\", (getter)getTupleAttr<ConvBackward, std::vector<int>, ConvParams, <del> &ConvParams::output_padding, int64_t, PyInt_FromLong>, NULL, NULL, NULL}, </del> <ins> &ConvParams::output_padding, int64_t, PortablePyInt_FromLong>, NULL, NULL, NULL}, </ins> {(char*)\"groups\", (getter)getValueAttr<ConvBackward, int, ConvParams, <del> &ConvParams::groups, int64_t, PyInt_FromLong>, NULL, NULL, NULL}, </del> <ins> &ConvParams::groups, int64_t, PortablePyInt_FromLong>, NULL, NULL, NULL}, </ins> {NULL} }; static struct PyGetSetDef conv_backward_backward_properties[] = { THP_FUNCTION_DEFAULT_PROPERTIES, {(char*)\"stride\", (getter)getTupleAttr<ConvBackwardBackward, std::vector<int>, ConvParams, <del> &ConvParams::stride, int64_t, PyInt_FromLong>, NULL, NULL, NULL}, </del> <ins> &ConvParams::stride, int64_t, PortablePyInt_FromLong>, NULL, NULL, NULL}, </ins> {(char*)\"padding\", (getter)getTupleAttr<ConvBackwardBackward, std::vector<int>, ConvParams, <del> &ConvParams::padding, int64_t, PyInt_FromLong>, NULL, NULL, NULL}, </del> <ins> &ConvParams::padding, int64_t, PortablePyInt_FromLong>, NULL, NULL, NULL}, </ins> {(char*)\"dilation\", (getter)getTupleAttr<ConvBackwardBackward, std::vector<int>, ConvParams, <del> &ConvParams::dilation, int64_t, PyInt_FromLong>, NULL, NULL, NULL}, </del> <ins> &ConvParams::dilation, int64_t, PortablePyInt_FromLong>, NULL, NULL, NULL}, </ins> {(char*)\"transposed\", (getter)getValueAttr<ConvBackwardBackward, bool, ConvParams, &ConvParams::transposed, long, PyBool_FromLong>, NULL, NULL, NULL}, {(char*)\"output_padding\", (getter)getTupleAttr<ConvBackwardBackward, std::vector<int>, ConvParams, <del> &ConvParams::output_padding, int64_t, PyInt_FromLong>, NULL, NULL, NULL}, </del> <ins> &ConvParams::output_padding, int64_t, PortablePyInt_FromLong>, NULL, NULL, NULL}, </ins> {(char*)\"groups\", (getter)getValueAttr<ConvBackwardBackward, int, ConvParams, <del> &ConvParams::groups, int64_t, PyInt_FromLong>, NULL, NULL, NULL}, </del> <ins> &ConvParams::groups, int64_t, PortablePyInt_FromLong>, NULL, NULL, NULL}, </ins> {NULL} };", "pos": ["One representative error: The cause of the problem is : It seems that on clang, changing the type parameter here is sufficient to cause template instantiation to fail. Maybe the easiest way to fix this is to write a more portable version of PyIntFromLong (and friends) which always returns ."], "neg": []}
{"query": "id_map = {old_id: p for old_id, p in zip(chain(*(g['params'] for g in saved_groups)), chain(*(g['params'] for g in groups)))} <del> state = defaultdict( dict, {id_map.get(k, k): v for k, v in state_dict['state'].items()}) </del> <ins> def cast(param, value): \"\"\"Make a deep copy of value, casting all tensors to device of param.\"\"\" if torch.is_tensor(value): # Floating-point types are a bit special here. They are the only ones # that are assumed to always match the type of params. if any(tp in type(param.data).__name__ for tp in {'Half', 'Float', 'Double'}): value = value.type_as(param.data) value = value.cuda(param.get_device()) if param.is_cuda else value.cpu() return value elif isinstance(value, dict): return {k: cast(param, v) for k, v in value.items()} elif isinstance(value, Iterable): return type(value)(cast(param, v) for v in value) else: return value # Copy state assigned to params (and cast tensors to appropriate types). # State that is not assigned to params is copied as is (needed for # backward compatibility). state = defaultdict(dict) for k, v in state_dict['state'].items(): if k in id_map: param = id_map[k] state[param] = cast(param, v) else: state[k] = v </ins> # Update parameter groups, setting their 'params' value def update_group(group, new_group):", "pos": ["Hi, I encountered this bug: The code skeleton is like: It seems the loaded are , and I've tried a workaround to move to , but it still has the same bug.\nCould you provide a full script to reproduce the problem?\nmaybe you can try like this, () (beta1).add_(1 - beta1, ())\nSorry, I missed the reply email. I am afraid that I am unable to provide a reproducer now. It is a work I am doing for the OpenNMT-py , trying to use for doing lr update. And I encoutered this problem when testing the case. So I factored out the code skeleton about this problem above. I've tried several methods, including tricks like what suggests, but it still happens. Per my analysis, it is because the previous training was done on gpu, so when saving the , the stored states(tensors) are of version. During resuming, when we load the saved optimizer, loads this version to cpu(the model() can be moved to gpu easily, but seems lacking this ability?) , so this problem emerges.\nTry moving optimizer state to the GPU memory manually after loading it from the checkpoint. I agree that having an method for this operation would be nice.\nthanks. My work was suspended due to other problems, when resumed, I will try your method.\nI'm afraid solution won't work. It will make the error go away, but your optimizer will no longer be training the model. You should recreate optimizers after casting modules to a different type or device, and you can use to restore the state from a previous copy. This currently doesn't work, but we should fix it (by copying from the data from the state dict, instead of using the tensors directly - this allows for cross-device or cross-type updates).\n, yep, your method is what I currently use, it works. But I will wait for upstream to fix this problem though. Thanks for your great works!\nAh, my bad. I forgot to update the line where the optimizer is recreated. But otherwise, the following should do the job, right?\nah, right. That should work\nExcept that you should use instead of\nI had a similar problem. When I save the optimizer state from a GPU other than GPU 0 and then load the state it still loads everything to GPU 0.", "Specifying in didn't work either. 's solution solves this though.\nHi guys, I have a very similar problem as the one in this thread, here's my code: And then once I resume, I got KeyErrors on my optimizer: Do you guys know how to fix this issue? BTW, I have 8 GPUs used, I'm guessing if this issue was because of that?\nwere you able to solve your issue with multiple-GPUs?\nThanks for asking. Yeah, I'm able to, turns out the issue is because I used an early version of PyTorch, after I updated the version, it works like a charm!\nJust a comment, this problem is caused by makes all state tensor are moved into GPU0 so by moving the state of an optimizer to specific GPU will fix this problem.\nHi how do you know the specific GPU location of different tensors in advance?\nWould a feature where all () calls always makes use of an automatically generated CPU version be feasible ? And then at resume the () would make use of the \"current\" device being used (or any better strategy). At the moment it seems we need a lot of boilerplater code to ensure saving and loading are consistents across devices for models/optimizer/scheduler/etc.\nI have met similar problem, I recreated Adam optimizer without () after reloading model, () and DataParallel(model) according to solution.\nthanks, it work!\nHi, as you say that: every time moving model to other device, we should build optimizer, but, if we move the model to other device and move back, should we build the optimizer again? here is an example code: does optimizer run as expected? also, if doing , should we rebuild optimizer ?\nIs there a problem if you switch the order to something like this? Meaning moving the model to 'cuda' but only loading it's state dict from checkpoint after loading the optimizer's state dict first?\nThe problem can be concluded that the optimizer's state will be loaded to the device as same as the model. You must load the model to GPU at first, and then load the optimizer's state. So that both the model and the optimizer's state are loaded in GPU.", "Instead of moving optimizer to cuda after loading it in cpu, you could load the checkpoint directly in cuda:\nI've independently rediscovered that this works :) Should read until the end of the thread next time\nI find my codes still have the problem. I tried my best to range the modules as the examples shown in the above. can anyone give me some hints? I also tried this\nFor me, doesn't solve the issue, but the one by works. Any idea why? I create the optimizer, load the state (with map location to CUDA), pass to the train loop, where the mode is pushed to device (though I assume it's not needed if the loaded model is pushed to cuda by maplocation). Specifically, I would get stuck at call, throwing the device mismatch error.\nStill it doesn't work for me. Could you please help me?"], "neg": []}
{"query": "pytorch-linux-trusty-py3.6-gcc5.4-test pytorch-linux-trusty-py3.6-gcc7.2-build pytorch-linux-trusty-py3.6-gcc7.2-test <ins> pytorch-linux-trusty-py3.6-gcc7-build pytorch-linux-trusty-py3.6-gcc7-test </ins> pytorch-linux-trusty-pynightly-build pytorch-linux-trusty-pynightly-test pytorch-win-ws2016-cuda9-cudnn7-py3-build", "pos": ["We're not testing against CUDA 9.2 and we should. C.f. We should test against GCC 7 (because this is how Anaconda is building the packages), AND we should use test with this in OS X, because CUDA 9.2 now supports the default xcode shipped on OS X.\nBuilding PyTorch with CUDA 9.2 for 7.0 compute compatibility devices produces quite worrisome warnings: warning : Instruction 'vote' without '.sync' may produce unpredictable results on sm70 and later architectures warning : Instruction 'vote' without '.sync' is deprecated since PTX ISA version 6.0 and will be discontinued in a future PTX ISA version warning : Instruction 'shfl' without '.sync' is deprecated since PTX ISA version 6.0 and will be discontinued in a future PTX ISA version warning : Instruction 'shfl' without '.sync' may produce unpredictable results on sm70 and later architectures\nCUDA 9.2 CI on macOS is done in"], "neg": []}
{"query": ".. automethod:: rsqrt .. automethod:: rsqrt_ .. automethod:: scatter_ <ins> .. automethod:: scatter_add_ </ins> .. automethod:: select .. automethod:: set_ .. automethod:: share_memory_", "pos": ["is not available in the docs page, but the doc string has been in . This was mentioned earlier in", "Oh, bilinear, I missed that last time I looked at it. I'll add a docstring.", "returns nothing. Might be useful to have some linter in the build process to find Python functions missing", "Pytorch 0.4.0 gives a segmentation fault with dilated 2D convolutions over a specific size. import torch import as nn model = nn.Conv2d(inchannels=64, outchannels=64, kernelsize=5, dilation=2) model(([1, 64, 1166, 1166])) # This runs model(([1, 64, 1167, 1167])) # This crashes PyTorch version: 0.4.0 Is debug build: No CUDA used to build PyTorch: 9.1.85 OS: CentOS Linux 7 (Core) GCC version: (GCC) 4.8.5 (Red Hat 4.8.5-28) CMake version: version 2.8.12.2 Python version: 3.6 Is CUDA available: Yes CUDA runtime version: 8.0.61 GPU models and configuration: GPU 0: GeForce GTX 1080 Ti GPU 1: TITAN X (Pascal) GPU 2: Quadro K2200 Nvidia driver version: 390.59 cuDNN version: Probably one of the following: /usr/local/cuda- /usr/local/cuda- /usr/local/cuda-7.5/lib64/libcudnnstatic.a /usr/local/cuda- /usr/local/cuda-8.0/lib64/libcudnn_static.a\ndo you think this is the same as the segfault you showed me a week ago on convolution?\nah yes. int overflow. I only fixed conv3d because I didn\u2019t want to hurt perf. I\u2019ll fix all of them.", "This should be an easy one.. The function is implemented, just no sign of it in the docs."], "neg": []}
{"query": "if config.triton.autotune_at_compile_time: self.kernel_autotune_calls.writeline(line) if zero_fill: <del> # TODO: remove this function to use the default WrapperCodegen behavior after service platform has zero_() symbol # default behavior is f\"workspace.zero_(){self.ending}\" self.writeline( f\"AOTI_TORCH_ERROR_CODE_CHECK(aoti_torch_zero_(workspace.get())){self.ending}\" ) </del> <ins> if isinstance(nbytes, sympy.Expr): self.writeline(f\"workspace.zero_(){self.ending}\") else: # TODO: remove this function to use the default WrapperCodegen behavior after service platform has zero_() symbol # default behavior is f\"workspace.zero_(){self.ending}\" self.writeline( f\"AOTI_TORCH_ERROR_CODE_CHECK(aoti_torch_zero_(workspace.get())){self.ending}\" ) </ins> if config.triton.autotune_at_compile_time: self.kernel_autotune_calls.writeline(f\"workspace.zero_(){self.ending}\")", "pos": ["Repro: Error: cc\nUn-assign myself to let work on it.\n, I can't reproduce the issue. The is no longer in the codebase. I gave a try of , which worked well. Could you double check if you can reproduce the issue?\nYes, I can still reproduce with your cmdline. Can you retry after deleting your local cache?"], "neg": []}
{"query": "INCLUDE(CheckCXXSourceCompiles) <ins> # disable some verbose warnings IF (MSVC) set(CMAKE_CXX_FLAGS \"/wd4267 /wd4251 /wd4522 /wd4522 /wd4838 /wd4305 /wd4244 /wd4190 /wd4101 /wd4996 /wd4275 ${CMAKE_CXX_FLAGS}\") ENDIF(MSVC) </ins> # windef.h will define max/min macros if NOMINMAX is not defined IF(MSVC) add_definitions(/DNOMINMAX)", "pos": ["That's pretty big! In contrast, an OS X build log is only 1.3M. Is there a reason they have to be this big? CC\nMy guess is that it's because of warnings. I remember having to get through thousands of them when looking for the error that failed the build.\nYes, the major part of them are warnings. And some of them is not fixable. We can hide them using some flags on the compilers.\nAnd this build configuration actually contains both the CPU and GPU build, which may also be a possible reason.\nThe reduced log size after PR is about 1.2M."], "neg": []}
{"query": "END_HANDLE_TH_ERRORS } <ins> PyObject * THCPModule_getDeviceCapability_wrap(PyObject *self, PyObject *arg) { HANDLE_TH_ERRORS THPUtils_assert(THPUtils_checkLong(arg), \"invalid argument to getDeviceCapability\"); long device = THPUtils_unpackLong(arg); cudaDeviceProp prop; THCudaCheck(cudaGetDeviceProperties(&prop, device)); return Py_BuildValue(\"(ii)\", prop.major, prop.minor); END_HANDLE_TH_ERRORS } </ins> PyObject * THCPModule_getCurrentStream_wrap(PyObject *self) { HANDLE_TH_ERRORS", "pos": ["There's quite a few user-reports of installing the CUDA 7.5 binary of pytorch and then running on GTX 1080 / 1070, and there will be reports of installing 7.5 / 8.0 binaries and running on V100. Raising a user-warning in such situations and letting them know will fix the problem.\nThe version of CUDA used to compile the binary should be available thanks to this PR:\nI think is a more robust way of getting the CUDA used to compile PyTorch, as currently if there are multiple CUDA installed in the machine, the heuristics used in might not match those used by CMake FindCUDA. I think that we should modify the to return the cuda version used during compilation. I have the impression that it will not require changing much because doesn't require cuda to execute. What do you think?"], "neg": []}
{"query": "Other Operations ~~~~~~~~~~~~~~~~~~~~~~ <ins> .. autofunction:: bincount </ins> .. autofunction:: cross .. autofunction:: diag .. autofunction:: diagflat", "pos": ["is not available in the docs page, but the doc string has been in . This was mentioned earlier in", "Oh, bilinear, I missed that last time I looked at it. I'll add a docstring.", "returns nothing. Might be useful to have some linter in the build process to find Python functions missing", "Pytorch 0.4.0 gives a segmentation fault with dilated 2D convolutions over a specific size. import torch import as nn model = nn.Conv2d(inchannels=64, outchannels=64, kernelsize=5, dilation=2) model(([1, 64, 1166, 1166])) # This runs model(([1, 64, 1167, 1167])) # This crashes PyTorch version: 0.4.0 Is debug build: No CUDA used to build PyTorch: 9.1.85 OS: CentOS Linux 7 (Core) GCC version: (GCC) 4.8.5 (Red Hat 4.8.5-28) CMake version: version 2.8.12.2 Python version: 3.6 Is CUDA available: Yes CUDA runtime version: 8.0.61 GPU models and configuration: GPU 0: GeForce GTX 1080 Ti GPU 1: TITAN X (Pascal) GPU 2: Quadro K2200 Nvidia driver version: 390.59 cuDNN version: Probably one of the following: /usr/local/cuda- /usr/local/cuda- /usr/local/cuda-7.5/lib64/libcudnnstatic.a /usr/local/cuda- /usr/local/cuda-8.0/lib64/libcudnn_static.a\ndo you think this is the same as the segfault you showed me a week ago on convolution?\nah yes. int overflow. I only fixed conv3d because I didn\u2019t want to hurt perf. I\u2019ll fix all of them.", "This should be an easy one.. The function is implemented, just no sign of it in the docs."], "neg": []}
{"query": "ENDFUNCTION() ################################################################################ <ins> if(${CMAKE_VERSION} VERSION_LESS \"2.8.12\") set(CMAKE_CXX_FLAGS \"${CMAKE_CXX_FLAGS} -std=c++11\") else(${CMAKE_VERSION} VERSION_LESS \"2.8.12\") if(${CMAKE_VERSION} VERSION_LESS \"3.1\") add_compile_options(-std=c++11) # CMake 2.8.12 to 3.1 endif(${CMAKE_VERSION} VERSION_LESS \"3.1\") endif(${CMAKE_VERSION} VERSION_LESS \"2.8.12\") </ins> IF(NOT Torch_FOUND) FIND_PACKAGE(Torch REQUIRED)", "pos": ["Currently, CMake 3.0 or higher is required for THPP. This is annoying for building from source on Ubuntu 14.04 systems, as default CMake is 2.8, and updating it requires either compiling from source or adding some random repos:\nit should work fine if you change it to 2.8. I'm testing this out in a sec, and will push it into master\nNope, it does not for me."], "neg": []}
{"query": "import torch from functools import reduce from ._utils import _range <ins> from sys import float_info __MIN_LOG_SCALE = math.ceil(math.log(float_info.min * float_info.epsilon, 10)) </ins> class __PrinterOptions(object):", "pos": ["on print(torch.DoubleTensor(1)) I got ZeroDivisionError: float division by zero but torch.DoubleTensor(0)) or torch.DoubleTensor(2)) work just fine. I work in Jupyter notebook pytorch 0.1.12\nNote that returns undefined memory. Could you find the specific content of the tensor that gave this error, by printing ? I'm not sure if this bug has been fixed in the latest release already.\nthe error is pretty interesting. I restarted Jupyter notebook. # many freaking signatures: print(torch.FloatTensor(1)) print(torch.IntTensor(1)) print(torch.DoubleTensor(1)) # no error right after initial run. #print(torch.DoubleTensor(0)) print(torch.DoubleTensor(2)) print(torch.ByteTensor(1)) print(torch.CharTensor(1)) print(torch.ShortTensor(1)) print(torch.LongTensor(1)) steps to reproduce the error: 1). run, no errors 2) repeat, no error 3) uncomment ..DoubleTensor(0), run, got an error (see below): 4) comment ..DouleTensor(0), run still have this error 2.+24 [torch.FloatTensor of size 1] 7.+08 [torch.IntTensor of size 1] ZeroDivisionError Traceback (most recent call last) &lt;ipython-input-5-3629c250c505in &lt;module() 2 print(torch.FloatTensor(1)) 3 print(torch.IntTensor(1)) ----4 print(torch.DoubleTensor(1)) # ZeroDivisionError: float division by zero 5 #print(torch.DoubleTensor(0)) 6 print(torch.DoubleTensor(2)) C:UsersleogAppDataLocalcondacondaenvspy35libsite-packagestorch in str(self) 126 # characters to replace unicode characters with.", "127 if sys.versioninfo (3,): --128 return (self) 129 else: 130 if hasattr(, 'encoding'): C:UsersleogAppDataLocalcondacondaenvspy35libsite-packagestorch in str(self) 284 return '[{} with no dimension]n'.format(torch.typename(self)) 285 elif self.ndimension() == 1: --286 strt = vectorstr(self) 287 elif self.ndimension() == 2: 288 strt = matrixstr(self) C:UsersleogAppDataLocalcondacondaenvspy35libsite-packagestorch in vectorstr(self) 270 if () &lt; PRINTOPTS.threshold: 271 return (strt + --272 'n'.join(ident + (val / scale) for val in self) + 273 'n') 274 else: C:UsersleogAppDataLocalcondacondaenvspy35libsite-packagestorch in &lt;genexpr(.0) 270 if () &lt; PRINT_OPTS.threshold: 271 return (strt + --272 'n'.join(ident + (val / scale) for val in self) + 273 'n') 274 else: ZeroDivisionError: float division by zero\nOk, I got a repro. Here is the failing case: Fails with the error you mentioned. This seems to be an edge case though, as both and works. The issue is in , where becomes zero because is a too large negative number. Maybe we should clip to be positive?"], "neg": []}
{"query": "std::vector<bool> *is_variable_input; char has_freed_buffers; <ins> // See a comment in THPFucntion_asFunction for details about this field. std::weak_ptr<torch::autograd::PyFunction> cdata_ptr; </ins> torch::autograd::PyFunction cdata; };", "pos": ["there is a bug. the code is from The code for installing cuda10.2 should be corrected to The code for installing cuda10.2 should be corrected to cc\nThank you for reporting the typo", "This is how I implement the decoder of a sequence to sequence model I'm not sure about the new autograd mechanics but this worked in the previous version. If I didn't make the unrolling codes a function it will work. It will also work with CPU. I compiled from source with Python 2.7, CUDA 8.0 and Cudnn 6\nThanks for the minimal reproduction! It broke OpenNMT and a big internal model but we weren't looking forward to isolating what parts of those had the issue.\nThat's great, I'll take a look today. Thanks!\nI've pushed a fix to .\nHow does the inference work in this case? I mean, once you train it how do you test it in the test phase?"], "neg": []}
{"query": "IF \"%CMAKE_GENERATOR%\"==\"Ninja\" ( IF \"%CC%\"== \"\" set CC=cl.exe IF \"%CXX%\"== \"\" set CXX=cl.exe <del> set MAKE_COMMAND=cmake --build . --target install --config Release -- -j%NUMBER_OF_PROCESSORS% || exit /b 1 </del> <ins> set MAKE_COMMAND=cmake --build . --target install --config %BUILD_TYPE% -- -j%NUMBER_OF_PROCESSORS% || exit /b 1 </ins> ) ELSE ( <del> set MAKE_COMMAND=msbuild INSTALL.vcxproj /p:Configuration=Release </del> <ins> set MAKE_COMMAND=msbuild INSTALL.vcxproj /p:Configuration=%BUILD_TYPE% </ins> ) )", "pos": ["The current master triggers the following error messages from Visual Studio 2015 (Debug build): 2------ Build started: Project: ATen, Configuration: Debug x64 ------ 2cl : Command line error D8016: '/Ox' and '/RTC1' command-line options are incompatible ... 14LINK : fatal error LNK1104: cannot open file '....srcATenDebug' It appears that they are triggered by *.c files in the ATen project, like \"srcTHvectorAVX.c\"\nWhy do you need Debug build?\nIsn't ATen meant to support both Debug and Release builds? FYI I'm involved with the project at which does support both build configurations, and we would like to be able to link our project with ATen.\nWell, you can remove the flags in as a workground.\nWe should patch this file to ignore this flag when building the library in debug mode. It shouldn't be too hard\nI catched similar errors , when I Tryed to build Debug Version I try to build Release and catched next errors e:pytorchatensrcatencpuvec256vec256int.h(184): note: see reference to function template instantiation 'at::vec256::anonymous-namespace'::Vec256<int16t::blend<1(at::vec256::anonymous-namespace'::Vec256<int16t)' being compiled 28e:pytorchatensrcatencpuvec256vec256int.h(147): error C3861: 'mm256extract_epi16': identifier not found What I should do?\nThere're two issues in your description. The first question is easy. You'll also have to specify the flag to 1. The second question is about your compiler. I guess you are using VS2015 or earlier versions of VS2017. Make sure the version in the output is 19.11 after your enter in CMD.\nYes I am using VS2015. My Compiler is 19.0.2. Must I use VS 2017 in any case?\nSorry, but the answer is yes. The AVX2 functions was in VS2017 and they are used in PyTorch. So VS2017 19.11 is a must.\nNo problem:) Thank you."], "neg": []}
{"query": "libs = ['TH', 'THS', 'THNN'] if WITH_CUDA: libs += ['THC', 'THCS', 'THCUNN'] <del> if WITH_NCCL and not SYSTEM_NCCL: </del> <ins> if WITH_NCCL and not WITH_SYSTEM_NCCL: </ins> libs += ['nccl'] libs += ['THPP', 'libshm', 'ATen', 'nanopb'] if WITH_DISTRIBUTED:", "pos": ["Currently, you can have a libnccl on your system, and pytorch setup favors that instead of compiling one itself from source. The code that takes care of this is: However, this breaks gloo build because gloo gets a path to NCCL via:\nI got the error during the installation: Is it related to this problem?\nthat's a dupe of\nno, that's a consequence of , now you have to do before compiling. I think you don't want to do recursive submodule update, because that will pull in gtest and another copy of nccl.\nAlternatively we can pass new flags to gloo to prevent it from using its own NCCL\nThanks works for me.\nYou can prevent it from using its own nccl, but its cmake still has to be able to find some nccl.\nI think should be sufficient. I'm testing.\nfixes it when we dont have an external NCCL and pytorch compiles it's own NCCL?\nI don't see how NCCLEXTERNAL fixes anything, looks like in gloo NCCLEXTERNAL is used just to set dependency if gloo is compiling its own nccl (which we don't want in any case).\nThe only thing I tested when I was that the vendored nccl was not being built (even if its submodule was checked out); I tested this by adding an to one of the headers. I'll test with the other configurations.\nActually, there are a few bugs here. First, even though GLOOFLAGS looks like it's setting NCCLROOTDIR: it actually isn't because the gloo invocation is which expands to but only is passed to cmake. Suppose you fix this. nccl is still not going to be picked up, because INSTALLDIR is PyTorch's local install dir, but if you disabled nccl builds, there will never actually be a nccl here. So cmake is just going to do it's standard search to find a nccl somewhere. Inversely, if you ever built nccl, and then turn it off, we will keep picking up that locally built version until you delete it. So we shouldn't do that. Suppose you fix this.", "Then I think gloo will (continue to) correctly pick up system nccl and be built with it, and the final thing to fix is that build of PyTorch csrc doesn't know how to search for nccl headers somewhere. This is probably related to . I wonder, maybe we shouldn't do any autodetection at all, and force the user to pass us the include and library paths of nccl if they want to build with system.\nFYI, I committed a change to gloo to stop bundling NCCL itself. It will now happily pick up whatever NCCL you give it through NCCLROOTDIR or by trying to find it in the standard systems paths. See facebookincubator/gloo You do have to set to make Gloo try and find it and compile in support.\nSo to fix this, you should be able to not specify anything aside from if you want to use system NCCL, and add if you have some custom location (either a PyTorch bundled version or a custom install altogether)."], "neg": []}
{"query": "static void setattr(jit::Node* n, jit::Symbol name, const at::IntList& v)  { n->is_(name, v); } static void setattr(jit::Node* n, jit::Symbol name, bool v)                { n->i_(name, v); } static void setattr(jit::Node* n, jit::Symbol name, double v)              { n->f_(name, v); } <del> template<unsigned long N> </del> <ins> template<std::size_t N> </ins> static void setattr(jit::Node* n, jit::Symbol name, std::array<bool, N> v) { n->is_(name, std::vector<int64_t>(v.begin(), v.end())); } VariableType::VariableType(Context* context, Type* baseType)", "pos": ["Hello, I'm trying to build pytorch (I tried both v3_buildfix and master branch) on a raspberry pi 3 but i'm getting that error: Note: I'm building without CUDA and DISTRIBUTED support. OS: Ubuntu 16.04 (Mate) gcc: arm-linux-gnueabihf-gcc (Ubuntu/Linaro 5.4.0-6ubuntu1~16.04.5) 5.4.0 cmake: v3.5.1\nCan you try building with and let me know if that fixes it?\nThanks, it compiles now ;)"], "neg": []}
{"query": "last_epoch (int): The index of last epoch. Default: -1. Example: <del> >>> # Assuming optimizer uses lr = 0.5 for all groups </del> <ins> >>> # Assuming optimizer uses lr = 0.05 for all groups </ins> >>> # lr = 0.05     if epoch < 30 >>> # lr = 0.005    if 30 <= epoch < 80 >>> # lr = 0.0005   if epoch >= 80", "pos": ["Hi, In the is written: \"lr values are [0.05, 0.005, ...].\" If I understand it correctly, it should be [0.5, 0.05, 0.005, ...].\nThat is correct. Please feel free to send a PR to fix it! (otherwise, I can)\nIf you don't mind to change it yourself, I'm not contributor yet and don't want to start just for one comment. I guess is the same case."], "neg": []}
{"query": "\"test_mm_plus_mm2_cuda_dynamic_shapes\": test_torchinductor.TestFailure( (\"cuda_wrapper\",), is_skip=True ), <del> \"test_consecutive_split_cumprod_cuda_dynamic_shapes\": test_torchinductor.TestFailure( (\"cuda_wrapper\",), is_skip=True ), </del> }", "pos": ["Repro: Error: cc\nUn-assign myself to let work on it.\n, I can't reproduce the issue. The is no longer in the codebase. I gave a try of , which worked well. Could you double check if you can reproduce the issue?\nYes, I can still reproduce with your cmdline. Can you retry after deleting your local cache?"], "neg": []}
{"query": "<del> from collections import defaultdict </del> <ins> from collections import defaultdict, Iterable </ins> import torch from copy import deepcopy", "pos": ["Hi, I encountered this bug: The code skeleton is like: It seems the loaded are , and I've tried a workaround to move to , but it still has the same bug.\nCould you provide a full script to reproduce the problem?\nmaybe you can try like this, () (beta1).add_(1 - beta1, ())\nSorry, I missed the reply email. I am afraid that I am unable to provide a reproducer now. It is a work I am doing for the OpenNMT-py , trying to use for doing lr update. And I encoutered this problem when testing the case. So I factored out the code skeleton about this problem above. I've tried several methods, including tricks like what suggests, but it still happens. Per my analysis, it is because the previous training was done on gpu, so when saving the , the stored states(tensors) are of version. During resuming, when we load the saved optimizer, loads this version to cpu(the model() can be moved to gpu easily, but seems lacking this ability?) , so this problem emerges.\nTry moving optimizer state to the GPU memory manually after loading it from the checkpoint. I agree that having an method for this operation would be nice.\nthanks. My work was suspended due to other problems, when resumed, I will try your method.\nI'm afraid solution won't work. It will make the error go away, but your optimizer will no longer be training the model. You should recreate optimizers after casting modules to a different type or device, and you can use to restore the state from a previous copy. This currently doesn't work, but we should fix it (by copying from the data from the state dict, instead of using the tensors directly - this allows for cross-device or cross-type updates).\n, yep, your method is what I currently use, it works. But I will wait for upstream to fix this problem though. Thanks for your great works!\nAh, my bad. I forgot to update the line where the optimizer is recreated. But otherwise, the following should do the job, right?\nah, right. That should work\nExcept that you should use instead of\nI had a similar problem. When I save the optimizer state from a GPU other than GPU 0 and then load the state it still loads everything to GPU 0.", "Specifying in didn't work either. 's solution solves this though.\nHi guys, I have a very similar problem as the one in this thread, here's my code: And then once I resume, I got KeyErrors on my optimizer: Do you guys know how to fix this issue? BTW, I have 8 GPUs used, I'm guessing if this issue was because of that?\nwere you able to solve your issue with multiple-GPUs?\nThanks for asking. Yeah, I'm able to, turns out the issue is because I used an early version of PyTorch, after I updated the version, it works like a charm!\nJust a comment, this problem is caused by makes all state tensor are moved into GPU0 so by moving the state of an optimizer to specific GPU will fix this problem.\nHi how do you know the specific GPU location of different tensors in advance?\nWould a feature where all () calls always makes use of an automatically generated CPU version be feasible ? And then at resume the () would make use of the \"current\" device being used (or any better strategy). At the moment it seems we need a lot of boilerplater code to ensure saving and loading are consistents across devices for models/optimizer/scheduler/etc.\nI have met similar problem, I recreated Adam optimizer without () after reloading model, () and DataParallel(model) according to solution.\nthanks, it work!\nHi, as you say that: every time moving model to other device, we should build optimizer, but, if we move the model to other device and move back, should we build the optimizer again? here is an example code: does optimizer run as expected? also, if doing , should we rebuild optimizer ?\nIs there a problem if you switch the order to something like this? Meaning moving the model to 'cuda' but only loading it's state dict from checkpoint after loading the optimizer's state dict first?\nThe problem can be concluded that the optimizer's state will be loaded to the device as same as the model. You must load the model to GPU at first, and then load the optimizer's state. So that both the model and the optimizer's state are loaded in GPU.", "Instead of moving optimizer to cuda after loading it in cpu, you could load the checkpoint directly in cuda:\nI've independently rediscovered that this works :) Should read until the end of the thread next time\nI find my codes still have the problem. I tried my best to range the modules as the examples shown in the above. can anyone give me some hints? I also tried this\nFor me, doesn't solve the issue, but the one by works. Any idea why? I create the optimizer, load the state (with map location to CUDA), pass to the train loop, where the mode is pushed to device (though I assume it's not needed if the loaded model is pushed to cuda by maplocation). Specifically, I would get stuck at call, throwing the device mismatch error.\nStill it doesn't work for me. Could you please help me?"], "neg": []}
{"query": "throw std::runtime_error(ss.str()); } <ins> static void checkIndexTensorTypes(TensorList indices) { for (auto& tensor : indices) { if (tensor.defined()) { auto& type = tensor.type(); auto scalarType = type.scalarType(); if (scalarType != kLong && scalarType != kByte) { throw std::runtime_error(\"tensors used as indices must be long or byte tensors\"); } } } } </ins> static std::vector<Tensor> expandByteTensors(const Tensor & self, TensorList indices) { // Expands byte tensors (masks) into the equivalent indexing by LongTensors std::vector<Tensor> result;", "pos": ["In PyTorch master 0.4 , variables cannot be indexed. This used to work in PyTorch 0.2 :\nThis is breaking some of Pyro's inference algorithms on PyTorch master:"], "neg": []}
{"query": "<ins> From PyTorch: Copyright (c) 2016-     Facebook, Inc            (Adam Paszke) Copyright (c) 2014-     Facebook, Inc            (Soumith Chintala) Copyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert) Copyright (c) 2012-2014 Deepmind Technologies    (Koray Kavukcuoglu) Copyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu) Copyright (c) 2011-2013 NYU                      (Clement Farabet) Copyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou, Iain Melvin, Jason Weston) Copyright (c) 2006      Idiap Research Institute (Samy Bengio) Copyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert, Samy Bengio, Johnny Mariethoz) From Caffe2: Copyright (c) 2016-present, Facebook Inc. All rights reserved. All contributions by Facebook: Copyright (c) 2016 Facebook Inc. All contributions by Google: Copyright (c) 2015 Google Inc. All rights reserved. All contributions by Yangqing Jia: Copyright (c) 2015 Yangqing Jia All rights reserved. All contributions from Caffe: Copyright(c) 2013, 2014, 2015, the respective contributors All rights reserved. All other contributions: Copyright(c) 2015, 2016 the respective contributors All rights reserved. Caffe2 uses a copyright model similar to Caffe: each contributor holds copyright over their contributions to Caffe2. The project versioning records all such contribution and copyright details. If a contributor wants to further mark their specific copyright on a particular contribution, they should indicate their copyright solely in the commit message of the change when it is committed. </ins> All rights reserved. Redistribution and use in source and binary forms, with or without", "pos": ["The three-clause BSD license in file LICENSE says in clause 1 \"Redistributions of source code must retain the above copyright notice...\" However, there is no longer a copyright notice in the file; it appears to have been removed in commit (see ).\ncc"], "neg": []}
{"query": "# Target only our CI GPU machine's CUDA arch to speed up the build export TORCH_CUDA_ARCH_LIST=5.2 <ins> if [[ \"$BUILD_ENVIRONMENT\" == *trusty-py3.6-gcc5.4* ]]; then export DEBUG=1 fi </ins> WERROR=1 python setup.py install # Add the test binaries so that they won't be git clean'ed away", "pos": ["Turning on DEBUG=1 turns off inlining. However, if you declare a function as , it won't get object code generated for it, which means there will be undefined symbol errors. You might not notice this with a release build, since everything gets inlined away. We can catch errors like this by doing CI with DEBUG=1. Probably just pick one platform and run that only."], "neg": []}
{"query": "} } <del> variable_list flatten(const TensorList& tensors) { </del> <ins> static variable_list flatten(const TensorList& tensors) { </ins> return cast_tensor_list(tensors); } <del> variable_list flatten(const Tensor& x, const TensorList& y) { </del> <ins> static variable_list flatten(const Tensor& x, const TensorList& y) { </ins> std::vector<Variable> r; r.reserve(1 + y.size()); r.emplace_back(x);", "pos": ["See Docker run fails with , which is due to the fact that is specified in docker container: Which results in: CI cc\nCheck that adding to manywheels image results in a failure:\nTentatively marking hi-pri since nightly testing are failing\nThis is resolved, here is the successful run:"], "neg": []}
{"query": "if (info != 0) THError(\"MAGMA orgqr : Argument %d : illegal value.\", -info); <ins> THCTensor_(free)(state, a); </ins> THCTensor_(free)(state, work); magma_free_pinned(tau_data);", "pos": ["It seems like the function leaks memory. I'm using 0.2.0_4 on anaconda python 3.6 CUDA 8.0 using the standard conda installation. So the memory footprint increases up to 1.9G though each takes about 100M. compared to where it takes up more memory but doesn't leak memory and the memory foot print fluctuates between 414M and 432M and back to 414M. Aside from that, is there a reason for using Magma instead of cuBLAS?\nThanks for catching this! Looks like we're leaking : I think at the time we the Magma bindings cuBLAS didn't provide a QR solver. I'm not sure there's any reason to prefer it now."], "neg": []}
{"query": "return value.reshape(required_shape).sum(-1) <del> def log_sum_exp(tensor, keepdim=True): </del> <ins> def _log_sum_exp(tensor, keepdim=True): </ins> r\"\"\" Numerically stable implementation for the `LogSumExp` operation. The summing is done along the last dimension.", "pos": ["Leaving this issue here for higher visibility. Originally reported by here: Also mentioned by Piggybacking off of this issue, has strange behavior (reported here: )\nI have fixed this locally. Could I send a PR?\nYes, go ahead!"], "neg": []}
{"query": "l = torch.randn(15) source = TensorDataset(t, l) for i in range(15): <del> self.assertEqual(t[i:i + 1], source[i][0]) self.assertEqual(l[i:i + 1], source[i][1]) </del> <ins> self.assertEqual(t[i], source[i][0]) self.assertEqual(l[i], source[i][1]) </ins> class ErrorDataset(Dataset):", "pos": ["TensorDataset adds casts 1D target tensor to 2D which is not acceptable by NLL loss layer.\nFixed in ."], "neg": []}
{"query": "#include <cstdint> <del> namespace F = torch::nn::functional; </del> namespace torch { namespace nn {", "pos": ["As part of pull request , F is redefined globally in nn/modules/batchnorm.h F is then used once in batchnorm.h and again in instancenorm.h not a bug, but a problem for anyone using F in some other way. it's solvable, but I don't see this pattern anywhere else in libtorch, was wondering if this use is an anomaly, and if we could use thanks cc\nThanks a lot for reporting! Yes this is an anomaly, and will be fixed by"], "neg": []}
{"query": "* Inception (warning: this model is highly sensitive to changes in operator implementation) * ResNet <del> * SqueezeNet </del> * SuperResolution * VGG * `word_language_model <https://github.com/pytorch/examples/tree/master/word_language_model>`_", "pos": ["Squeezenet is not exportable via onnx since it uses in which is not supported by ONNX. Please update documentation accordingly."], "neg": []}
{"query": "return res def emit_record_trace(env, declaration): <ins> # Note [clang-802.0.42 tuple overload bug] # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ # Originally, my plan for emit_$ecord_trace was to keep it as # simple as possible, if at the expense of some somewhat ugly # overloads.  So this meant we had a 'recordTrace' function # with overloads like this: # #   recordTrace(..., const Variable& out) #   recordTrace(..., const std::tuple<Variable, Variable>& out) # # Unfortunately, this triggers a bug in clang-802.0.42 # (widely used in macOS Sierra 10.12.6) wherein a Variable is # implicitly convertible into a std::tuple<Variable, Variable>; # a minimal repro can be seen below here: # #   #include <tuple> #   struct T {}; #   void f(const std::tuple<T, T>&) {} #   void g(T& x) { f(x); } # # To work around this bug, the code generator is a bit more # complicated, and is taught how to handle this situation. </ins> local = {} <ins> arguments = declaration['arguments'] tensor_args = [arg for arg in arguments if arg['simple_type'] in {'Tensor', 'TensorList'}] if len(tensor_args) == 1 and tensor_args[0]['simple_type'] == 'TensorList': # Special case for TensorList.  This only works when there # is a single argument local['trace_inputs'] = \"cast_tensor_list({})\".format(declaration['arguments'][0]['name']) else: local['trace_inputs'] = CodeTemplate(\"{ ${tensor_args} }\").substitute(env) </ins> local['record_attributes'] = [] for arg in declaration['arguments']: if arg['simple_type'] in {'Tensor', 'TensorList'}:", "pos": ["when executing:\nI'm seeing this as well as of last night. Its preceded by many warnings about : But the first error is the one about . I went back, and I get a good install as late as commit"], "neg": []}
{"query": "v = self.ValueTensor([5]).cuda(0) self.assertRaises(RuntimeError, lambda: self.SparseTensor(i, v, torch.Size([3]))) <ins> def _test_new_device(self, size, device): with torch.cuda.device(device): x = torch.cuda.sparse.DoubleTensor(*size) self.assertEqual(x.get_device(), device) x1 = x.new() x2 = x.new(2, 3) self.assertEqual(x1.get_device(), device) self.assertEqual(x2.get_device(), device) </ins> @cuda_only <del> @unittest.skipIf(torch.cuda.device_count() < 2, \"only one GPU detected\") def test_new_device(self): with torch.cuda.device(1): x = torch.cuda.sparse.DoubleTensor(30, 2) self.assertEqual(x.get_device(), 1) </del> <ins> def test_new_device_single_gpu(self): self._test_new_device((), 0) self._test_new_device((30, 20), 0) self._test_new_device((30, 20, 10), 0) </ins> <del> y = x.new(2, 3) self.assertEqual(y.get_device(), 1) </del> <ins> @cuda_only @unittest.skipIf(torch.cuda.device_count() < 2, \"only one GPU detected\") def test_new_device_multi_gpu(self): self._test_new_device((), 1) self._test_new_device((30, 20), 1) self._test_new_device((30, 20, 10), 1) </ins> class TestUncoalescedSparse(TestSparse):", "pos": ["This causes to not work properly on newly created empty sparse tensors. In the example below, both should print 1 instead."], "neg": []}
{"query": "if bias is not None: self.bias.data.uniform_(-0.1, 0.1) <del> def forward(self, input): # See the autograd section for explanation of what happens here. return Linear()(input, self.weight, self.bias) </del> <ins> def forward(self, input): # See the autograd section for explanation of what happens here. return Linear()(input, self.weight, self.bias) </ins> Writing custom C extensions", "pos": ["Broken download links for libtorch v1.8.0 with CUDA 10.2 for Linux (pre-cxx11 and cxx11 ABI) and Windows (Release and Debug versions) on the PyTorch website. Steps to reproduce the behavior: on any of the following links leads to an \"Access Denied\" error: The chosen libtorch version is downloaded. PyTorch Version: .1.8.0 OS: Linux and Windows CUDA version: 10.2\nThe correct links are listed below. cc\nThanks!\nSame issue also happens for CUDA 11.1 when trying to download windows precompiled binaries Changing to in the links suggested by still leads to Access Denied page\nOkay I wrote a simple script to test if all of our libtorch packages are actually downloadable, sharing results here: <details<summarypkgs </summaryCommand I used to generate these (from the root of ): </detailsOutput being: I did find that this issue is correct in that the installation instructions were incorrect for CUDA 10.2 and I created a PR to resolve this:"], "neg": []}
{"query": "} private: <del> friend class VariableVersion; </del> <ins> friend struct VariableVersion; </ins> int expected_version; std::shared_ptr<VersionBlock> version_block;  // may be null };", "pos": ["when executing:\nI'm seeing this as well as of last night. Its preceded by many warnings about : But the first error is the one about . I went back, and I get a good install as late as commit"], "neg": []}
{"query": "- [NVIDIA CUDA](https://developer.nvidia.com/cuda-downloads) 7.5 or above - [NVIDIA CuDNN](https://developer.nvidia.com/cudnn) v5.x <ins> If you want to disable CUDA support, export environment variable `NO_CUDA=1`. </ins> #### Install optional dependencies On Linux", "pos": ["Caffe and TensorFlow supports both CPUONLY and CUDA mode. The meaning of CPUONLY mode is to enable users to run pytorch on architectures except for AMD64 and PPC64EL (Seems that CUDA is only available on AMD64 and PPC64EL). For example, the CPU_ONLY version of is available for (not only) amd64, armel, i386, ppc64el and s390x, while the CUDA version only supports amd64 and ppc64el.\nI think we don't plan to invest too much time in supporting many different architectures. The vast majority of the users is on x86 and this will remain our main focus. If someone is willing to spend their time supporting other archs, we'll happily accept PRs, but we have a lot of other priorities. We also support CPU only and CUDA mode. You can use or build pytorch without any GPU. We also have CPU-only binaries (see the website). You can force building without CUDA with env variable, as I pointed out in another PR (it's broken because of THPP at the moment, but I'll fix it soon). The only limitation for supporting different archs that I see is what our backends are limited to, and the fact that you need a C++11 compiler to build the extension. Apart from this, I don't think we depend on x86 anywhere. Can you please elaborate on what kind of guidance would you like to see? More detailed instructions on how to build pytorch somewhere in the docs?\nAs reported in [1], Pytorch's backend library (libTH) is now healthy in portability. I suggest that the environment variable should be referred in . [1]", "I'm trying to implement distributed adversarial training in PyTorch. Thus, in my program pipeline I need to forward the output of one DDP model to another one. When I run the code in distributed setting, the following error is throwed: frame : c10::Error::Error(c10::SourceLocation, std::string const&) + 0x45 (0x7fa248f906d5 in /export/home/haoran/anaconda3/envs/torch1.2/lib/python3.6/site-) frame : c10d::Reducer::prepareforbackward(std::vector<torch::autograd::Variable, std::allocator<torch::autograd::Variableconst&) + 0x666 (0x7fa26f9d0896 in /export/home/haoran/anaconda3/envs/torch1.2/lib/python3.6/site-) frame : <unknown function+ 0x6df3b8 (0x7fa26f9c43b8 in /export/home/haoran/anaconda3/envs/torch1.2/lib/python3.6/site-) frame : <unknown function+ 0x1fc7f0 (0x7fa26f4e17f0 in /export/home/haoran/anaconda3/envs/torch1.2/lib/python3.6/site-) frame : main + 0x16d (0x400c1d in /export/home/haoran/anaconda3/envs/torch1.2/bin/python) frame : _libcstartmain + 0xf5 (0x7fa285e06c05 in ) frame : /export/home/haoran/anaconda3/envs/torch1.2/bin/python() [0x4009e9] Running the following code with DistributedDataParallel <!-- If you have a code sample, error messages, stack traces, please provide it here as well --The code could run successfully without throwing error. Please copy and paste the output from our (or fill out the checklist below manually).", "You can get the script and run it with: PyTorch version: 1.2.0.dev20190701 Is debug build: No CUDA used to build PyTorch: 9.0.176 OS: CentOS Linux 7 (Core) GCC version: (GCC) 5.4.0 CMake version: version 3.6.2 Python version: 3.6 Is CUDA available: Yes CUDA runtime version: 9.0.176 GPU models and configuration: GPU 0: Tesla P40 GPU 1: Tesla P40 GPU 2: Tesla P40 GPU 3: Tesla P40 Nvidia driver version: 390.25 cuDNN version: Could not collect Versions of relevant libraries: [pip] numpy==1.16.4 [pip] torch==1.2.0.dev20190701 [conda] blas 1.0 mkl defaults [conda] mkl 2019.4 243 defaults [conda] mklfft 1.0.12 py36ha843d7b0 defaults [conda] mklrandom 1.0.2 py36hd81dba30 defaults [conda] pytorch-nightly 1.2.0.dev20190701 py3.6cuda9.0.176cudnn7.5.1_0 pytorch <!-- Add any other context about the problem here. --\nPlease, try one of the suggested options: (1) passing the keyword argument findunusedparameters=True to ; (2) making sure all forward function outputs participate in calculating loss. If you already have done the above two steps, then the distributed data parallel module wasn't able to locate the output tensors in the return value of your module's forward function. Please include the loss function and the structure of the return value of forward of your module when reporting this issue (e.g. list, dict, iterable). If the issue will still exist, please, reopen this issue.\nI double checked your repro to figure out if this is a real issue or not and found that even setting won't help you here. It's a simple example of course, but you're not using the variable in your loss computation. The model's forward function returns it which is an indication for DDP that it will receive a gradient.", "Then when it doesn't, and you try to run the next iteration, it raises an error, because the previous iteration didn't finish. I recommend you return only the tensor you plan to use in the loss computation.\nIs there any way to see which parameters on earth are unused? It is useless to simply give this RuntimeError or set findunusedparameters to True.\nI had the same problem with just one GPUI have modified the code according to what you said\nI met the same problem and I found this topic. It works for me. (though it's been a long time for you)"], "neg": []}
{"query": "BASE_DIR=$(pwd) cd torch/lib INSTALL_DIR=\"$(pwd)/tmp_install\" <del> BASIC_C_FLAGS=\" -DTH_INDEX_BASE=0 -I$INSTALL_DIR/include -I$INSTALL_DIR/include/TH -I$INSTALL_DIR/include/THC \" </del> <ins> BASIC_C_FLAGS=\" -DTH_INDEX_BASE=0 -DTH_GENERIC_USE_HALF=1 -DCUDA_HAS_FP16=1 -I$INSTALL_DIR/include -I$INSTALL_DIR/include/TH -I$INSTALL_DIR/include/THC \" </ins> LDFLAGS=\"-L$INSTALL_DIR/lib \" if [[ $(uname) == 'Darwin' ]]; then LDFLAGS=\"$LDFLAGS -Wl,-rpath,@loader_path\"", "pos": ["Hi, I try to run the following example from here, but run into some issues: ld: warning: ignoring file , file was built for unsupported file format ( 0x7F 0x45 0x4C 0x46 0x02 0x01 0x01 0x03 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 ) which is not the architecture being linked (x8664): ld: warning: ignoring file , file was built for unsupported file format ( 0x7F 0x45 0x4C 0x46 0x02 0x01 0x01 0x03 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 ) which is not the architecture being linked (x8664): ld: warning: ignoring file , file was built for unsupported file format ( 0x7F 0x45 0x4C 0x46 0x02 0x01 0x01 0x03 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 ) which is not the architecture being linked (x86_64): full error log: i installed pytorch: im running pytorch on: osx 10.11.6 i installed libtorch following these steps: here the cmake output:\nshould i use: wget -c but the problem is that libtorch-macos- does not contain any libraries in ./lib folder. how to create it? instead of: wget -c when i use libtorch-macos-, i get this error:\nI've got the same image not found example in this same spot, by following the minimal example of the docs. Running macOS Mojave here.\nwe need to build libtorch manually because it doesnt come with all libs...\nGetting the same issue as with macOS High Sierra 10.13.5. When I look in I see but not as it requires.\ncc: looks like a missing copying of in our OSX libtorch builds.\nSeeing the same thing.\nFixed this issue by downloading the two missing libraries from and copying them both (, ) to .", "Correct me if I am wrong, but that solution is not answering the original issue: I was having the same issue when trying to run the minimal cpp frontend example here: I grabbed libtorch linked on the main pytorch site: You are right, had neither nor in . However, copying those files to that location did not solve the issue for me. I was able to get the minimal example by giving cmake the for Pytorch's torch library (compiled from source): I am going to try compiling the DCGAN example for the cpp frontend in this same way. I assume that this is not going to work, that we will ultimately need to point to for something like data loading, etc. I am on Mac 10.14.3\nI was facing almost same problems here and managed to fix it by downloading and from . Just in case anyone needs this, please do NOT download the newest post there! Instead find the file instead of file. ! I didn't find the difference between and at first, but succeeded once downloaded the ones.\nAny idea whats going on with these missing MKL libraries for the mac c++ libtorch? Somehow this is still an issue dating back since 2018, and even after pytorch 1.3, I just tried it out again. The workaround done by above works for me, which involves manually copying the dylib files from old (pre v1.0) intel binaries.\nI fixed all such issues for v1.3.0 in or so I thought. But I only fixed them for pip and conda packages, not libtorch. I'm looking into it right now and will fix libtorch too.\nwill kick off new binary builds as the fix went in. will close the issue after the builds are uploaded and live.\nthis is fixed now with fixed binaries that are re-uploaded."], "neg": []}
{"query": "def batch_norm(input, running_mean, running_var, weight=None, bias=None, training=False, momentum=0.1, eps=1e-5): <ins> size = list(input.size()) if reduce(mul, size[2:], size[0]) == 1: raise ValueError('Expected more than 1 value per channel, got input size {}'.format(size)) </ins> f = torch._C._functions.BatchNorm(running_mean, running_var, training, momentum, eps, torch.backends.cudnn.enabled) return f(input, weight, bias)", "pos": ["Right now it makes everything .\nTo be clear, for a tensor, the problem is only when all of N, H, and W are 1. (So BatchNorm2d on batch size of 1 is OK as long as you don't have a 1x1 image). What's the desired behavior? The only reasonable behavior I can think of is: Raise an exception when the dimensions of which you are normalizing are one or Output zero (+ optional affine transform) I'm not sure the outputing zero is a good idea. I can't think of a case where that's what you want."], "neg": []}
{"query": "sum  ); } <del> THArgCheckWithCleanup((sum > 0), THCleanup(THDoubleTensor_free(cum_dist);), 2, </del> <ins> THArgCheckWithCleanup((sum > 0), THCleanup(THDoubleTensor_free(cum_dist); if (start_dim == 1) THTensor_(resize1d)(prob_dist, n_categories);), 2, </ins> \"invalid multinomial distribution (sum of probabilities <= 0)\"); /* normalize cumulative probability distribution so that last val is 1 i.e. doesn't assume original prob_dist row sums to one */", "pos": ["I was reading the documentation of method in the . I tried this in IPython environment. Of course, error occurs because 5 is larger than the size of weights and I didn't set . The error message is like this: But after this, the dimensions of weights changed. If I run as following without runtime exception, then the dimensions of weights do not change. This is caused by not cleaning up after exception raised in this . Maybe the solution is adding cleaning up code in the .\nThe problem emerged in two other situations as well, here are the examples: I'm sending a pull request with the fix. Thanks for reporting the problem. :)"], "neg": []}
{"query": "assert data_tensor.size(0) == target_tensor.size(0) self.data_tensor = data_tensor self.target_tensor = target_tensor <del> if self.data_tensor.dim() == 1: self.data_tensor = self.data_tensor.view(-1, 1) if self.target_tensor.dim() == 1: self.target_tensor = self.target_tensor.view(-1, 1) </del> def __getitem__(self, index): return self.data_tensor[index], self.target_tensor[index]", "pos": ["TensorDataset adds casts 1D target tensor to 2D which is not acceptable by NLL loss layer.\nFixed in ."], "neg": []}
{"query": "def grid_sampler(input, grid, padding_mode): <del> if cudnn.is_acceptable(input.data) and padding_mode == 'zeros' and input.dim() == 4: </del> <ins> if (cudnn.is_acceptable(input.data) and padding_mode == 'zeros' and input.dim() == 4 and input.size(1) <= 1024):  # as of cudnn 7102, will not work for larger than 1024 </ins> return torch.cudnn_grid_sampler(input, grid) else: return GridSampler.apply(input, grid, padding_mode)", "pos": ["PyTorch version: 0.4.0 Is debug build: No CUDA used to build PyTorch: 9.0.176 OS: Ubuntu 16.04.3 LTS GCC version: (Ubuntu 5.4.0-6ubuntu1~16.04.5) 5.4.0 CMake version: version 3.9.4 Python version: 3.6 Is CUDA available: Yes CUDA runtime version: 9.0.176 GPU models and configuration: GPU 0: TITAN X (Pascal) GPU 1: TITAN X (Pascal) Nvidia driver version: 384.98 cuDNN version: Probably one of the following: /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux-gnu/libcudnnstaticv7.a /usr/local/lib/python2.7/dist- /usr/local/lib/python3.5/dist- Versions of relevant libraries: [pip3] numpy (1.14.1) [pip3] numpydoc (0.6.0) [pip3] torch (0.4.0) [pip3] torchvision (0.2.0) [conda] cuda90 1.0 h6433d270 pytorch [conda] magma-cuda90 2.3.0 1 soumith [conda] pytorch 0.4.0 py36cuda9.0.176cudnn7.1.21 [cuda90] pytorch [conda] torchvision 0.2.1 py361 pytorch\n() returns 7102\nPossibly related:\nI'm not sure why, but I can't reproduce this on master. I can reproduce on 0.4 though.\nNever mind, this error happens when CUDNN is updated regardless of pytorch version."], "neg": []}
{"query": "#include <tuple> #include <TH/THMath.h> <del> #include \"THP.h\" #include \"copy_utils.h\" #include \"DynamicTypes.h\" </del> <ins> #include \"torch/csrc/THP.h\" #include \"torch/csrc/copy_utils.h\" #include \"torch/csrc/DynamicTypes.h\" </ins> <del> #include \"generic/Tensor.cpp\" #include <TH/THGenerateAllTypes.h> #include \"generic/Tensor.cpp\" #include <TH/THGenerateHalfType.h> </del> <ins> //generic_include TH torch/csrc/generic/Tensor.cpp </ins>", "pos": ["Trying to run on github actions via AWS instance gives me: I will of course set up now, but it shouldn't crash like that. May be it should use as a fallback? pt-2.0 cc\nIs this a multitenant setup? Also curious which distro are you running, I was under the impression that most linux distros set home by default\nthis is AWS image. one default user. I think it's user, but it wants you to run as root, which is weird. I have manually now set to overcome this.\nWe should update triton to use pathlib or something"], "neg": []}
{"query": "if [[ $WITH_CUDA -eq 1 ]]; then GLOO_FLAGS=\"-DUSE_CUDA=1 -DNCCL_ROOT_DIR=$INSTALL_DIR\" fi <ins> CWRAP_FILES=\" $BASE_DIR/torch/csrc/Declarations.cwrap; $BASE_DIR/torch/lib/ATen/Local.cwrap; $BASE_DIR/torch/lib/THNN/generic/THNN.h; $BASE_DIR/torch/lib/THCUNN/generic/THCUNN.h\" </ins> # Used to build an individual library, e.g. build TH function build() {", "pos": ["There are previous issues addressing this, but I want to write this up with a plan in a new issue. We should make differentiable instead of having a wrapper class. The Variable class adds conceptual overhead without sufficient benefit. It makes it harder for new users to adopt PyTorch by adding a second API which is similar but not exactly the same as Tensor. There are few prerequisites: on s need to be at least as fast as operations on () need scalar support at the Python level () The constructor will be deprecated and become a no-op. The accessor will be deprecated and return a view that has requiresgrad=False. The constructors will take an optional keyword argument There will be a number of internal changes, but we shouldn't break most existing code: performance reasons, we will no longer build up the graph of operations by default when unless the operation uses stochastic variables. We'll use a context manager or global variable to enable graph creation. new differentiable Tensor will need support all the operations currently on Tensor, including in-place and operations. -place operations on leaf Variables which requiregrad are not currently supported. In the new API, the tensor's value will be updated and it will remain a leaf node. (This is important for optimizers, which update leaf Variables in-place). There are probably a few other subtle differences between Tensor and Variable APIs. I'll update this issue as we address them. See also: 23/738 # Greg will be fixing the tests, Tensor vs. Variable requiresgrad 2/1837 # Just testbroadcast from TestTorch OK OK # error at top level OK OK OK OK OK OK OK OK 10/120 # missing sspaddmm, dsmm/mm, addmm 1/257 TestTorch.testbroadcast 2OK\nProperly handling in-place might not be as easy as it seems. Sometimes you actually need to move the Variable in the graph. Consider this case:\nOne of the cases of differences between variable and tensor is computation of var/std:\nIMO Variable gives user a clear difference between graph node and data. .data attribute and Variable constructor provide convenient and explicit ways to take control of the graph (or graphs).\nThis is done. (Still some clean-up to do)\nIs the conda PyTorch out of date? I'm getting the following error: .\nYes, we haven't released 0.4, which has the Variable Tensor merge, yet"], "neg": []}
{"query": "# Don't store the actual traceback to avoid memory cycle _queued_calls.append((callable, traceback.format_stack())) <ins> _lazy_call(_check_capability) </ins> class DeferredCudaCallError(Exception): pass", "pos": ["There's quite a few user-reports of installing the CUDA 7.5 binary of pytorch and then running on GTX 1080 / 1070, and there will be reports of installing 7.5 / 8.0 binaries and running on V100. Raising a user-warning in such situations and letting them know will fix the problem.\nThe version of CUDA used to compile the binary should be available thanks to this PR:\nI think is a more robust way of getting the CUDA used to compile PyTorch, as currently if there are multiple CUDA installed in the machine, the heuristics used in might not match those used by CMake FindCUDA. I think that we should modify the to return the cuda version used during compilation. I have the impression that it will not require changing much because doesn't require cuda to execute. What do you think?"], "neg": []}
{"query": "__global__ void renormRowsL1(T* dist, long rows, long cols) { extern __shared__  unsigned char my_smem[]; T *smem = reinterpret_cast<T *>(my_smem); <ins> T zero = ScalarConvert<int, T>::to(0); T val; </ins> for (int64_t row = blockIdx.x; row < rows; row += gridDim.x) { T sum = ScalarConvert<int, T>::to(0); for (int64_t col = threadIdx.x; col < cols; col += blockDim.x) { <del> sum = THCNumerics<T>::add(sum, dist[row * cols + col]); </del> <ins> val = dist[row * cols + col]; assert(THCNumerics<T>::ge(val, zero)); sum = THCNumerics<T>::add(sum, val); </ins> } <del> sum = reduceBlock(smem, blockDim.x, sum, ReduceAdd<T, T>(), ScalarConvert<int, T>::to(0)); </del> <ins> sum = reduceBlock(smem, blockDim.x, sum, ReduceAdd<T, T>(), zero); </ins> if (threadIdx.x == 0) { <ins> assert(THCNumerics<T>::gt(sum, zero)); </ins> smem[0] = sum; } __syncthreads();", "pos": ["Looking at the values returned from tn.multinomial(1) and counting them, it seems like it returns the values of the multinomial distribution taken from t, i.e. that it has ignored the signs of the values. However, it seem like always returns a tensor filled with zeros.\nlooking into this"], "neg": []}
{"query": "if declaration['inplace']: env['return_value'] = 'self' <del> env['return_name'] = 'self' </del> env['result'] = 'static_cast<Variable&>(self)' <ins> env['trace_outputs'] = '{ self }' </ins> else: env['return_value'] = '{}(std::move(ret))'.format(declaration['return_type']) <del> env['return_name'] = 'ret' </del> env['result'] = 'std::get<0>(ret)' if len(declaration['returns']) > 1 else 'ret' <ins> if len(declaration['returns']) > 1: # NB: This won't work if we get heterogenous outputs outs = ['std::get<{}>(ret)'.format(i) for i, v in enumerate(declaration['returns']) if v['type'] == 'Tensor'] else: outs = ['ret'] env['trace_outputs'] = CodeTemplate(\"{ ${outs} }\").substitute(outs=outs) </ins> if any(arg['simple_type'] in {'Generator', 'Storage'} for arg in arguments): env['record_trace'] = []", "pos": ["when executing:\nI'm seeing this as well as of last night. Its preceded by many warnings about : But the first error is the one about . I went back, and I get a good install as late as commit"], "neg": []}
{"query": "} cudaError_t CudaStorageWeakRefAllocator::free(void* ptr) { <del> PyGILState_STATE gstate = PyGILState_Ensure(); PyObject_SetAttrString(object.get(), \"cdata\", Py_None); object = nullptr; PyGILState_Release(gstate); </del> <ins> { AutoGIL gil; PyObject_SetAttrString(object.get(), \"cdata\", Py_None); object = nullptr; } </ins> cudaError_t err = allocator->free(allocatorContext, ptr); delete this; return err;", "pos": ["When both PyTorch and Pandas are imported, calling results in a segfault. The issue was originally reported by It looks like calling returns a null pointer here: Examining in GDB: I suspect it's both PyTorch and Pandas doing threading magic and somehow getting in each other's way. I haven't yet been able to provide a minimal example that triggers this behaviour and does not use my package; it's likely that I'm doing something odd that causes this. More documentation here: Possibly related issues: and Python version: 3.6 PyTorch version: v0.2.0 Pandas version: 0.20.1"], "neg": []}
{"query": "thisdir = path.dirname(__file__) libpaths = ['', path.join(thisdir, '../../lib')] if sys.platform.startswith('linux'): <del> libnames = ['libcudnn.so.5.1.5', 'libcudnn.so.5.1.3', 'libcudnn.so.5.0.5'] </del> <ins> libnames = ['libcudnn.so.5.1.5', 'libcudnn.so.5.1.3', 'libcudnn.so.5.0.5', 'libcudnn.so.5.1.10'] </ins> elif sys.platform == 'darwin': libnames = ['libcudnn.5.dylib'] else:", "pos": ["Kind of similar to Sometimes our Eigen build fails this way: It doesn't seem to reliably repro. cc\nI got some sort of similar-ish error persistently on a test branch:\nNot that persistently; a PR stacked on the failing one succeeded. So there is something nondeterministic going on here.\nanother instance: CI on\nLooks like an NVCC bug. The workground is protecting eigen headers with Can we extract some basic code example and compile it against NVCC?\nAnother one in\nCan you try with 10.2?\nMingbo, can you put general information on the issue you are seeing here in this issue?\nlatest word on this is from mingbo: \"from what I can tell, it's a flaky one. might depend on which VS version. I hit the problem before and problem fixed (at least seem so) after upgrading VS 2019 latest version\" and \"the same error showed up once last week, and re-run the job finished without problem.\"\nHi. Did anyone have this issue persistently, ie that a rebuild did not \"fix\" it? We've had this issue in another (closed source) project. Exposing cuda to less includes (such as boost) seems to help but is not predictable.\nIt's never been persistent for us, but this is mostly from our Windows CI builds where we blow everything away and rebuild from scratch, which may help.\nYet another occurrence in the forum: It seems to be fixed by , which implies there may be a race condition in reading the include files or generating the intermediate source files.\nI was able to reproduce the issue with a more verbose output with And then I inspected the variables carefully and found out that the VC env is activated twice. We should really avoid that.\nYou are SO COOL!!!\nI tried to build from scratch several times with this PR and could not reproduce the issue anymore. Let's assume it's fixed.\nLooks like it was not resolved.\nAnother occurrence: This time it seems to claim the error is around the usage of .\nThis sounds like a good old multiple-instances-of-nvcc-write-to-the-same-temp-file One can invoke to see all temporary filenames it creates. I guess on windows they are more likely to collide with each other than on linux. One solution to the problem would be to invoke nvcc sequentially.", "Another is to tweak environment (there must be something similar to TMPDIR on Windows, mustn't it?)\nYes, they are called TMP and TEMP.\nIs there any example of the latter method on Unix systems?\nI've written a utility . Hopefully it can be used to resolve this problem if that's the case.\nNew occurrence after the previous fix:\nseems related to\nIf this is the case, maybe we could apply to fix that.\nunfortunately, this introduces another thrust problem (e.g. ): D:/programs64/cuda101/includethrust/detail/complex/catrig.h(607): error: no instance of function template \"thrust::atanh\" matches the argument list argument types are: (double) with problematic include chain -thrust/complex.h -thrust/detail/complex/catrig.h (has cplusplus conditions) -thrust/detail/complex/c99math.h where \"using ::atanh\" is omitted for MSVC. Patching c99math.h (or even thrustdetailconfigcppdialect.h without using /Zc:_cplusplus) seems to work for me, but I'm not sure what's the best way to fix this within pytorch tree and whether c++ dialect is the only problem here.\nThanks for the investigation. BTW, could you please share with us how did you patch c99math.h?\nDisregard that, still got same unreproducible C2993 error with /Zc:_cplusplus. Earlier, I had \"build\" directory reproducibly failing (at ), and hardcoding c++17 in thrustdetailconfigcppdialect.h fixed that. But looks like it is just a symptom of some inconsistence in environment/flags (across nvcc stages maybe?) just \"using ::atanh\"\nYeah, maybe. I am considering to add a wrapper to retry the build a few times.\nI can now stably reproduce a related error by passing in .\nBy commenting out some part of the code, I find the code that causes the previous error. I don't know why it is executed in the device code.\nWell, I figured out the reason. It seems that and are implicitly marked as and . This behavior can be disabled by adding .\nNice work!\nSeems to be a nvcc bug\nThanks for reminding. I have commented there.\nThis is indeed an nvcc bug.", "There is no known workaround at the moment, but the next release of the CUDA toolkit will contain a fix. Ref thrust/thrust.\nMet this problem:\nJust retry a few times.\nThe CUDA TK 11.0 RC is now available: This issue has been fixed in that version of nvcc.\nThanks for the info. Actually, our release binaries rely on the cudatoolkit package in anaconda cloud. So we won't be able to publish the binaries with CUDA 11 now. I'm just curious whether that is maintained by Nvidia? Another question, any plan to backport the changes to exisiting CUDA versions (e.g. cuda 10.1, cuda 10.2)?\nI don't have that information unfortunately -- it'd be best to contact the maintainer of that package and see when they plan to update. I'm not aware of any plans to backport this fix.\nI have also obtained this error with CUDA 10.1/cuDNN 7.6.4 on Win Server 2019.\nthis is a cuda bug. Please see above discussion and upgrade to cuda 11.\nI fixed the error by installing sscache for use with cuda 10.1 and cuda 10.2, as done here.\nit's an intermittent error and the probability of occurrence varies depending on your system. So it's not really fixed, but you may have found a way to reduce the occurrence in your system, which may be sufficient in your case :)\nAnother instance, this time inside protobuf: Running with I found the syntax error is introduced by which is the cuda compiler front end. operates on the output from the compiler's pre-processor and that output barely even changed to trigger the build error, only line number metadata and some extra whitespace was changed in the pre-processed output. Here is the diff of the two inputs, one of which failed and the other didn't: And the corresponding change in the output: This failure reproduced reliably on the CI machine. Running nvcc manually multiple times, it always fails or passes given the same inputs. I think it's a deterministic failure, but is annoyingly sensitive to any changes in the input file whatsoever.\nclosing due to age\nWait, do we think this is fixed?", "It's not clear to me it is...\nI think this was related to CUDA 10.2 builds on windows which we don't actually support anymore so it is probably safe to keep closed."], "neg": []}
{"query": "# This will segfault if things have been erroneously released out.backward(torch.randn(out.size())) <ins> def test_norm_subgradient(self): def run_test(input_size, norm_deg): input = Variable(torch.zeros(*input_size), requires_grad=True) out = input.norm(norm_deg) out.backward() self.assertEqual(input.grad.data.abs().sum(), 0) run_test((10,), 2) run_test((10, 10), 2) run_test((10,), 3) </ins> def index_variable(shape, max_indices): if not isinstance(shape, tuple):", "pos": ["If a norm is zero, its gradient returns nan: Obviously just happening because the gradient divides by the norm, but the (sub)gradient here should probably be zero, or at least not nan, since that will propagate to make all updates nan. Probably low priority, as it's not going to be an issue in 99% of cases, but we're doing a few things with (exact) line searches where this caused a nan to appear, breaking everything downstream.\nI'm encountering exactly the same issue! Spent hours on debugging, just to find PyTorch has a bug on such basic thing.\n+1 just found this bug too\n+1 for this bug. Temporarily changing my code to something like the following for the sake of debugging. x = Variable((1), requires_grad=True) y = x + 1e-16 y.norm().backward() print x.grad\nThe thing is that in the 2 norm, there is a square root, which has a gradient of at 0. The gradient gives you because you then multiply 0 and an infinity during the backward pass.\nFor a scalar, norm 2 is basically abs. But x.abs().backward() gives you 0 gradient. In this sense, it's not coherent.\nI found this error, too\nAlban fixed this behavior in\nHi, the norm function can give use the 0 gradients now. However, the following code still has the nan gradient problem\nHo The square root has no gradient at 0. This is expected behavior.\nHi, but the sub-gradient of the square root should be zero? Also, y = ( x * x ) should equal to x.norm(), why they have different gradient ( 0 and nan )?\nI think was right. The left-side derivative of sqrt(x) at x=0 is undefined, so it doesn't even have a subgradient at x=0.\nsquare root has no subgradient at 0. You could define a gradient by continuity but then it would be ... Given that pytorch is using autograd, and (equivalent to your ) are completely different: The first one is a single function that is convex and defined on R, it has a subgradient of 0 at 0. The second one is composed of two function, the first function is the square function which is differentiable and outputs values in .", "The second function is the square root that is not convex and even though it is defined on , it is only differentiable on and it's gradient in 0 in undefined. Given that, even though and will return the same value, their gradients may differ at points where it is not differentiable, this is because automatic differentiation looks at each step of the computation one by one and even though in some cases a subgradient exist (because we look at multiple operations as a single function), it is not always the case and the gradient remains undefined.\nI think math is math. Any root's gradient at zero is either inf or undefined. This issue shall be handled by the users' themselves by adding a small value(as did), but an error(warning) message may be helpful since it is pretty hard to debug. Say: Infinite/Undefined gradient is detected at XfunctionX at line Y. Exit.\nAgree, norm is not differentiable at 0 , the bandaid that Alban put there in is wrong (even in the limit sense the gradient at 0 should be 1 not 0), but it should not have been there at all. Norm is norm, if someone want to add epsilons to their norms (like batchrnorm, e.g.) they are welcome to do so in the user code. What would numpy do?\nI've also run into a number of problems related to the change introduced in . Is there a reason the subgradient is set to 0, rather than the 1? (The limit as norm-0?) As a minimal example: Produces\nWell any value between [-1, 1] is a valid subgradient for the 2-norm. More genereally, any vector in the 1 ball for the dual norm is a valid subgradient. This means that 0 is always going to be a subgradient, while 1 will not be for all p. Anyway, the theory says that any of them could be taken and subgradient descent will work. I'm sure that depending on the application, one will be better than the other. For example, the relu function will also give a 0 subgradient at 0, you could have given 1. The main point here was to remove nans that make your network give nan for everything which is not convenient."], "neg": []}
{"query": "<ins> #include \"torch/csrc/utils/object_ptr.h\" #include <Python.h> template<> void THPPointer<PyObject>::free() { if (ptr) Py_DECREF(ptr); } template class THPPointer<PyObject>; </ins>", "pos": ["Go to Look at the html. It contains the following: ! This problem manifests in google search. Searching for 1.8.1 docs brings up the \"1.8.0 docs\": ! We can probably fix this by doing a quick sed on the website but we should develop a solution for future releases. I think this has happened before in the past and am not sure what causes it. cc"], "neg": []}
{"query": "def forward(input, weight, hidden): if batch_first: <del> input.transpose(0, 1) </del> <ins> input = input.transpose(0, 1) </ins> nexth, output = func(input, hidden, weight) if batch_first: <del> output.transpose(0, 1) </del> <ins> output = output.transpose(0, 1) </ins> return output, nexth", "pos": ["The last line here fails on CPU or when CUDNN is otherwise unavailable: This is because accidentally assumes 's in-place semantics rather than the functional semantics of ( gets it right): I can push a PR that fixes this, or one of the devs can put it in the next bugfix PR:\nIf you have a ready solution for your issues feel free to send the PR straight away. Contributions are welcome!\nThanks"], "neg": []}
{"query": "def fn_base(optimizer, weight, bias): optimizer.zero_grad() <del> loss = (weight.mv(input) + bias).pow(2).sum() </del> <ins> i = input_cuda if weight.is_cuda else input loss = (weight.mv(i) + bias).pow(2).sum() </ins> loss.backward() return loss", "pos": ["Hi, I encountered this bug: The code skeleton is like: It seems the loaded are , and I've tried a workaround to move to , but it still has the same bug.\nCould you provide a full script to reproduce the problem?\nmaybe you can try like this, () (beta1).add_(1 - beta1, ())\nSorry, I missed the reply email. I am afraid that I am unable to provide a reproducer now. It is a work I am doing for the OpenNMT-py , trying to use for doing lr update. And I encoutered this problem when testing the case. So I factored out the code skeleton about this problem above. I've tried several methods, including tricks like what suggests, but it still happens. Per my analysis, it is because the previous training was done on gpu, so when saving the , the stored states(tensors) are of version. During resuming, when we load the saved optimizer, loads this version to cpu(the model() can be moved to gpu easily, but seems lacking this ability?) , so this problem emerges.\nTry moving optimizer state to the GPU memory manually after loading it from the checkpoint. I agree that having an method for this operation would be nice.\nthanks. My work was suspended due to other problems, when resumed, I will try your method.\nI'm afraid solution won't work. It will make the error go away, but your optimizer will no longer be training the model. You should recreate optimizers after casting modules to a different type or device, and you can use to restore the state from a previous copy. This currently doesn't work, but we should fix it (by copying from the data from the state dict, instead of using the tensors directly - this allows for cross-device or cross-type updates).\n, yep, your method is what I currently use, it works. But I will wait for upstream to fix this problem though. Thanks for your great works!\nAh, my bad. I forgot to update the line where the optimizer is recreated. But otherwise, the following should do the job, right?\nah, right. That should work\nExcept that you should use instead of\nI had a similar problem. When I save the optimizer state from a GPU other than GPU 0 and then load the state it still loads everything to GPU 0.", "Specifying in didn't work either. 's solution solves this though.\nHi guys, I have a very similar problem as the one in this thread, here's my code: And then once I resume, I got KeyErrors on my optimizer: Do you guys know how to fix this issue? BTW, I have 8 GPUs used, I'm guessing if this issue was because of that?\nwere you able to solve your issue with multiple-GPUs?\nThanks for asking. Yeah, I'm able to, turns out the issue is because I used an early version of PyTorch, after I updated the version, it works like a charm!\nJust a comment, this problem is caused by makes all state tensor are moved into GPU0 so by moving the state of an optimizer to specific GPU will fix this problem.\nHi how do you know the specific GPU location of different tensors in advance?\nWould a feature where all () calls always makes use of an automatically generated CPU version be feasible ? And then at resume the () would make use of the \"current\" device being used (or any better strategy). At the moment it seems we need a lot of boilerplater code to ensure saving and loading are consistents across devices for models/optimizer/scheduler/etc.\nI have met similar problem, I recreated Adam optimizer without () after reloading model, () and DataParallel(model) according to solution.\nthanks, it work!\nHi, as you say that: every time moving model to other device, we should build optimizer, but, if we move the model to other device and move back, should we build the optimizer again? here is an example code: does optimizer run as expected? also, if doing , should we rebuild optimizer ?\nIs there a problem if you switch the order to something like this? Meaning moving the model to 'cuda' but only loading it's state dict from checkpoint after loading the optimizer's state dict first?\nThe problem can be concluded that the optimizer's state will be loaded to the device as same as the model. You must load the model to GPU at first, and then load the optimizer's state. So that both the model and the optimizer's state are loaded in GPU.", "Instead of moving optimizer to cuda after loading it in cpu, you could load the checkpoint directly in cuda:\nI've independently rediscovered that this works :) Should read until the end of the thread next time\nI find my codes still have the problem. I tried my best to range the modules as the examples shown in the above. can anyone give me some hints? I also tried this\nFor me, doesn't solve the issue, but the one by works. Any idea why? I create the optimizer, load the state (with map location to CUDA), pass to the train loop, where the mode is pushed to device (though I assume it's not needed if the loaded model is pushed to cuda by maplocation). Specifically, I would get stuck at call, throwing the device mismatch error.\nStill it doesn't work for me. Could you please help me?"], "neg": []}
{"query": "self._data_ptrs = [] return <ins> # If any parameters alias, we fall back to the slower, copying code path. This is # a sufficient check, because overlapping parameter buffers that don't completely # alias would break the assumptions of the uniqueness check in # Module.named_parameters(). unique_data_ptrs = set(p.data_ptr() for l in self.all_weights for p in l) if len(unique_data_ptrs) != sum(len(l) for l in self.all_weights): self._data_ptrs = [] return </ins> with torch.cuda.device_of(any_param): # This is quite ugly, but it allows us to reuse the cuDNN code without larger # modifications. It's really a low-level API that doesn't belong in here, but", "pos": ["While the fast path for cuDNN RNNs introduced in includes a check intended to fall back to copying weights if weights are moved or aliased, this only works if the aliasing happens after the weights are first compacted. But the most natural place to tie weights is inside a module's method, which runs before , the typical time when is first invoked. Here's an MWE. The three runs (CPU, weight tying before , and weight tying after ) should all give the same results, and they do before , but they have not since:"], "neg": []}
{"query": "(4.53, 3.83, -6.64, 2.06, -2.47, 4.70))).t() b = torch.Tensor(((8.58, 8.26, 8.48, -5.28, 5.72, 8.93), (9.35, -4.43, -0.70, -0.26, -7.36, -2.52))).t() <del> _test(a, b, expectedNorm) </del> <ins> _test_overdetermined(a, b, expectedNorm) </ins> # test underdetermined expectedNorm = 0", "pos": ["b = (600,1) A = (600,2) X, = (b,A) X.size() #torch.Size([2, 1]) Y, = (b.cuda(),A.cuda()) Y.size() #torch.Size([600, 1])\nthat is interesting. says that the first rows returned ( here) include the solution of the least-squares problem. I've verified that the first 2 rows of X and Y are the same so you can probably use that for now while we figure out what the rest of the output is."], "neg": []}
{"query": "ensure_attr('ClassNLLCriterion', 'weights') ensure_attr('ParallelCriterion', 'repeatTarget') ensure_attr('MultiMarginCriterion', 'weights') <ins> ensure_attr('SpatialConvolution', 'bias', 'finput', 'fgradInput', 'gradWeight', 'gradBias') </ins> attr_map('ReLU', {'val': 'value'}) attr_map('Threshold', {'val': 'value'}) attr_map('Unsqueeze', {'pos': 'dim'})", "pos": ["Subclassing torch.Tensor with fails due to a on the latest nightly builds, whereas it previously worked without issue. Run this snippet of code: And you'll recieve an error message like this: Changing the to results in the exact same error message. There should be no error message as it should be working like it was previously in Torch 1.8.1 and below. I think that the issue started on Torch version: , and I can confirm that it still exists as of , . The issue started on the nightly builds at least 2 days ago, if that helps resolve things. It's been causing some of Captum's CircleCI tests to fail, and the tests show the first occurrence of the error on . cc\nThis is due to either or\nI think the problem is I'm not supposed to delete because that means the super() call doesn't work. Instead, I should just verify that the passed in class is a subclass, then this should work."], "neg": []}
{"query": ") @unittest.skipIf(BACKEND != 'gloo', \"Only Gloo backend supports CUDA allReduce\") <ins> @skip_if_no_cuda_distributed </ins> def test_all_reduce_sum_cuda(self): group, group_id, rank = self._init_global_test() self._test_all_reduce_helper(", "pos": ["is causing a crash when we try to build pytorch. Crash looks like this: (all times in US/West) Our 7AM, 8AM builds and earlier were all fine but 9AM, 10AM, 11AM builds all crashed with that same traceback. The new import was in which I think went in between 8AM and 9AM today See command below: See here if needed for : <!-- A clear and concise description of what you expected to happen. --Please copy and paste the output from our (or fill out the checklist below manually). You can get the script and run it with: PyTorch Version (e.g., 1.0): building from master OS (e.g., Linux): linux How you installed PyTorch (, , source): source Build command you used (if compiling from source): see above under 'reproduce' Python version: 3.6\nmaybe easier to read for build command used:\nplease try in the CircleCi environment?\nin our build script seems to have fixed this. Thanks !"], "neg": []}
{"query": "# test is no more than 4 higher than the 10th available at the # start. This attempts to catch file descriptor leaks, but allows # one-off initialization that may use up a file descriptor <del> available_fds = self._get_next_fds(10) self.test_case.assertLessEqual( available_fds[-1] - self.next_fds[-1], 5) </del> <ins> # TODO: Disabled because this check is too flaky # available_fds = self._get_next_fds(10) # self.test_case.assertLessEqual( #     available_fds[-1] - self.next_fds[-1], 5) </ins> self.test_case.assertFalse(self.has_shm_files()) return False", "pos": ["causes a host-device synchronization on CUDA. This is because it must first calculate the count of nonzero elements, and then copy this count to host. There are cases where I can easily calculate the count myself (from existing host data), or already know the count, which would make this unnecessary. For some given of shape : Instead of: Do this trickery: This avoids the host-device synchronization, but on the other hand, is more complicated, and a bit slower. I'm not sure if there are any other ways? Btw, I'm using to measure the speed. But actually I'm not sure if this really properly covers the problem of host-device synchronization with . Does it? E.g. when comparing this trickery code to , I just see that is slower. No response_\ndupe of i think\nAh yes, I think you are right."], "neg": []}
{"query": "export CMAKE_PREFIX_PATH=${PYTORCH_ENV_DIR}/miniconda3/ # Test PyTorch <ins> if [[ \"${JOB_BASE_NAME}\" == *cuda9.2* ]]; then # Eigen gives \"explicit specialization of class must precede its first use\" error # when compiling with Xcode 9.1 toolchain, so we have to use Xcode 8.2 toolchain instead. export DEVELOPER_DIR=/Library/Developer/CommandLineTools else export DEVELOPER_DIR=/Applications/Xcode9.app/Contents/Developer fi </ins> export MACOSX_DEPLOYMENT_TARGET=10.9 export CXX=clang++ export CC=clang", "pos": ["Mac OS X builds failing this way. This appears to be correlated with worker-macos-high-sierra-2 and worker-macos-high-sierra-16."], "neg": []}
{"query": "output (Tensor): output Tensor \"\"\" <ins> if mode != 'bilinear': raise NotImplementedError(\"nn.functional.grid_sample got unsupported mode: '{}'\".format(mode)) </ins> return vision.grid_sampler(input, grid, padding_mode)", "pos": ["The only valid mode is 'bilinear', but one can pass anything to it and it will not complain:\nApparently, the is not passed to the backend function for . I think it should be removed.\nYes. I think the idea is that we'd like to support other modes at some point."], "neg": []}
{"query": "python setup.py install cd test/ echo \"Ninja version: $(ninja --version)\" <del> sh run_test.sh </del> <ins> sh run_test.sh -- -v </ins> echo \"BUILD PASSED\"", "pos": ["If a test segfaults, it's hard to determine which test it is because the output looks like: ! I can work on this if we think this is a good idea.\nLast time I asked about this, he didn't want to because it adds a lot of scrolling. I also did some work to move the CI tests to so that we can output XML status files, but (1) this is unlikely to help in case of segfault, and (2) you can't do this straightforwardly, because we run some test files multiple times with different environment variables, and the XML exporter isn't really set up to handle this case. For the record, I am personally OK with just slapping on the tests and calling it a day.\nin the conbuild sounds good to me too\nclosed via"], "neg": []}
{"query": "are zeros. A sparse tensor is represented as a pair of dense tensors: a tensor <del> of values and a tensor of indices.  A sparse tensor can be constructed </del> <ins> of values and a 2D tensor of indices.  A sparse tensor can be constructed </ins> by providing these two tensors, as well as the size of the sparse tensor <del> (which cannot be inferred from these tensors!) </del> <ins> (which cannot be inferred from these tensors!)  Suppose we want to define a sparse tensor with the entry 3 at location (0, 2), entry 4 at location (1, 0), and entry 5 at location (1, 2).  We would then write: </ins> <del> >>> i = torch.LongTensor([[0, 1], [2, 0]]) >>> v = torch.FloatTensor([3, 4]) </del> <ins> >>> i = torch.LongTensor([[0, 1, 1], [2, 0, 2]]) >>> v = torch.FloatTensor([3, 4, 5]) </ins> >>> torch.sparse.FloatTensor(i, v, torch.Size([2,3])).to_dense() 0  0  3 <del> 4  0  0 </del> <ins> 4  0  5 [torch.FloatTensor of size 2x3] Note that the input to LongTensor is NOT a list of index tuples.  If you want to write your indices this way, you should transpose before passing them to the sparse constructor: >>> i = torch.LongTensor([[0, 2], [1, 0], [1, 2]]) >>> v = torch.FloatTensor([3,      4,      5    ]) >>> torch.sparse.FloatTensor(i.t(), v, torch.Size([2,3])).to_dense() 0  0  3 4  0  5 </ins> [torch.FloatTensor of size 2x3] You can also construct hybrid sparse tensors, where only the first n", "pos": ["The indices don't align with the values. In the example given on the website, by this logic, seems like it's indexing the columns first and that the indices are in reverse order to their values. Seems like corresponds to 4 and corresponds to 3. i = torch.LongTensor([[0, 1], [2, 0]]) v = torch.FloatTensor([3, 4]) (i, v, torch.Size([2,3])).todense() 0 0 3 4 0 0 [torch.FloatTensor of size 2x3] If I change the indices to and , I get i = torch.LongTensor([[0, 1], [0,1]]) v = torch.FloatTensor([3, 4]) (i, v, torch.Size([2,3])).todense() 3 0 0 0 4 0 [torch.FloatTensor of size 2x3] seems like unexpected behavior. Also if I try - should be intuitive - putting a value in position i = torch.LongTensor([[0, 1], [2, 0], [1,1]]) v = torch.FloatTensor([3, 4, 5]) (i, v, torch.Size([2,3])).to_dense() Traceback (most recent call last): File \"<stdin\", line 1, in <moduleRuntimeError: invalid argument 2: number of dimensions must be nDimI + nDimV at /pytorch/torch/lib/THS/generic/THSTensor.c:169 Help please?\nThis is a doc problem, thank you. The docs are misleading here, because it is not that is a coordinate, but 0 from the first tensor, and 2 from the second tensor. I'll try to rewrite these docs so they are clearer.\nBasically, our constructor expects to be (transpose of what you gave it).\nWe should consider changing that"], "neg": []}
{"query": "1.2469  0.0064 -1.6250 [torch.FloatTensor of size 3x3] <del> >>> torch.tril(a, k=1) </del> <ins> >>> torch.tril(a, diagonal=1) </ins> 1.3225  1.7304  0.0000 -0.3052 -0.3111 -0.1809 1.2469  0.0064 -1.6250 [torch.FloatTensor of size 3x3] <del> >>> torch.tril(a, k=-1) </del> <ins> >>> torch.tril(a, diagonal=-1) </ins> 0.0000  0.0000  0.0000 -0.3052  0.0000  0.0000", "pos": ["The states that the argument controls which diagonal to consider. However, this argument is actually named in pytorch 0.2.0."], "neg": []}
{"query": "// should have saved the grad accumulator. Even if the Variable no longer // alive, the accumulator should be kept alive by the references in the graph). if (requires_grad && !grad_fn && weak_grad_fn.expired() && grad_accumulator.expired()) <del> throw std::logic_error(\"No grad accumulator for a saved leaf!\"); </del> <ins> throw std::logic_error(\"No grad accumulator for a saved leaf!\"); </ins> new_var->grad_accumulator = grad_accumulator; return new_var;", "pos": ["there is a bug. the code is from The code for installing cuda10.2 should be corrected to The code for installing cuda10.2 should be corrected to cc\nThank you for reporting the typo", "This is how I implement the decoder of a sequence to sequence model I'm not sure about the new autograd mechanics but this worked in the previous version. If I didn't make the unrolling codes a function it will work. It will also work with CPU. I compiled from source with Python 2.7, CUDA 8.0 and Cudnn 6\nThanks for the minimal reproduction! It broke OpenNMT and a big internal model but we weren't looking forward to isolating what parts of those had the issue.\nThat's great, I'll take a look today. Thanks!\nI've pushed a fix to .\nHow does the inference work in this case? I mean, once you train it how do you test it in the test phase?"], "neg": []}
{"query": "self += modules def __getitem__(self, idx): <del> if not (-len(self) <= idx < len(self)): raise IndexError('index {} is out of range'.format(idx)) if idx < 0: idx += len(self) return self._modules[str(idx)] </del> <ins> if isinstance(idx, slice): return ModuleList(list(self._modules.values())[idx]) else: if not (-len(self) <= idx < len(self)): raise IndexError('index {} is out of range'.format(idx)) if idx < 0: idx += len(self) return self._modules[str(idx)] </ins> def __setitem__(self, idx, module): return setattr(self, str(idx), module)", "pos": ["it'd be great to pick out elements of of a sequential with array slicing:\nI agree this idea!\nHowever,\nthis has come up again (last in ). Shall we reconsider?\nI attempted getting a to work. Is this fine? If this is fine, I can issue a PR too. Let me know what you think. I believe this modification needs to be done in the class.\ncc: This issue has been open for quite sometime."], "neg": []}
{"query": "#include <initializer_list> #include <iostream> #include <functional> <ins> #include <cstddef> </ins> #ifdef _MSC_VER #ifdef Type", "pos": ["Hello, I'm trying to build pytorch (I tried both v3_buildfix and master branch) on a raspberry pi 3 but i'm getting that error: Note: I'm building without CUDA and DISTRIBUTED support. OS: Ubuntu 16.04 (Mate) gcc: arm-linux-gnueabihf-gcc (Ubuntu/Linaro 5.4.0-6ubuntu1~16.04.5) 5.4.0 cmake: v3.5.1\nCan you try building with and let me know if that fixes it?\nThanks, it compiles now ;)"], "neg": []}
{"query": "call %CONDA_PARENT_DIR%Miniconda3Scriptsactivate.bat %CONDA_PARENT_DIR%Miniconda3 if \"%INSTALL_FRESH_CONDA%\"==\"1\" ( <del> call conda install -y -q python=%PYTHON_VERSION% numpy cffi pyyaml boto3 libuv </del> <ins> call conda install -y -q python=%PYTHON_VERSION% numpy\"<1.23\" cffi pyyaml boto3 libuv </ins> if errorlevel 1 exit /b if not errorlevel 0 exit /b call conda install -y -q -c conda-forge cmake=3.22.3", "pos": ["Numpy shouldn't require a pin with PyTorch. This issue is to ensure that should get reverted and all issues addressed (issues look like ) CI\nNever mind, we apparently should have been pinning it."], "neg": []}
{"query": "l = nn.Linear(10, 5).float().cuda() i = Variable(torch.randn(20, 10).float().cuda(1)) l.cuda(1) <del> expected_out = l(i).data l.cuda(0) out = dp.data_parallel(l, i, (0, 1)) self.assertEqual(out.get_device(), 0) self.assertEqual(out.data, expected_out) </del> <ins> expected_out = l(i) loss = expected_out.sum() loss.backward() expected_grads = [] for param in l.parameters(): expected_grads.append(param.grad.clone()) dev_ids_list = [(0, 1), (1, 0)] for dev_id in dev_ids_list: with torch.cuda.device(dev_id[0]): l.cuda() l.zero_grad() out = dp.data_parallel(l, i, dev_id) loss = out.sum() loss.backward() self.assertEqual(out.get_device(), dev_id[0]) self.assertEqual(out.data, expected_out.data) for expected, param in zip(expected_grads, l.parameters()): self.assertEqual(param.grad.data, expected.data) </ins> # Check for None device_ids <ins> l = l.cuda() </ins> out = dp.data_parallel(l, i) @unittest.skipIf(not TEST_MULTIGPU, \"multi-GPU not supported\")", "pos": ["The loss will never decrease when I train a model using multi-gpus but without gpu0. This is an example I modified from the : If the result after one epoch is If the result after one epoch is\nshould always contain . If you want to specify which GPUs to use, you can set the environment variable , either by running your script like this: or set it programmatically in your script like this: So that deviceids 0, 1 will map to GPUs 1, 2, respectively. (Note that for the second method, you should put that line before any cuda relevant function call.)\nThank you.\n<!-- drci-comment-start --:pagefacingup: Preview :pagefacingup: Preview :question: Need help or want to give feedback on the CI? Visit our Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. <!-- drci-comment-end --"], "neg": []}
{"query": "- name: adaptive_max_pool3d(Tensor self, IntList[3] output_size) cname: VolumetricAdaptiveMaxPooling <del> - name: avg_pool2d(Tensor self, IntList[2] kernel_size, IntList[2] stride={}, IntList[2] padding=0, bool ceil_mode=false, bool count_include_pad=false) </del> <ins> - name: avg_pool2d(Tensor self, IntList[2] kernel_size, IntList[2] stride={}, IntList[2] padding=0, bool ceil_mode=false, bool count_include_pad=true) </ins> cname: SpatialAveragePooling default_init: stride: kernel_size <del> - name: avg_pool3d(Tensor self, IntList[3] kernel_size, IntList[3] stride={}, IntList[3] padding=0, bool ceil_mode=false, bool count_include_pad=false) </del> <ins> - name: avg_pool3d(Tensor self, IntList[3] kernel_size, IntList[3] stride={}, IntList[3] padding=0, bool ceil_mode=false, bool count_include_pad=true) </ins> cname: VolumetricAveragePooling default_init: stride: kernel_size", "pos": [", thank you! The documentation for avgpool2d/3d says Indeed, the default was just that up to PyTorch 0.2. In 0.3-0.4 the default is False for 2d/3d and True 1d and the doc says it was True for all. The change in default happened when replaced the previous mechanism of calling THNN. Now I was about to submit a PR reverting the default to True, but the subtle twist is that TF uses False by default. avgpool2d is used in the wild, e.g. by ResNet. Should I change the docstring or should I change the default in Do you have an opinion ( maybe)? Best regards Thomas\nSo after submitting this, I saw that the module interface uses countincludepad=True all along, so the right option seems to be to have the same in the functional interface. I'll submit a PR.\nwe should indeed default functional interface to True as well. Thanks for the PR"], "neg": []}
{"query": "self.assertEqual(reference[None, 2:5, None, None], reference.unsqueeze(0)[:, 2:5].unsqueeze(2).unsqueeze(2)) # indexing with step <del> reference = self._consecutive((10, 10, 10)) </del> <ins> reference = consec((10, 10, 10)) </ins> self.assertEqual(reference[1:5:2], torch.stack([reference[1], reference[3]], 0)) self.assertEqual(reference[1:6:2], torch.stack([reference[1], reference[3], reference[5]], 0)) self.assertEqual(reference[1:9:4], torch.stack([reference[1], reference[5]], 0))", "pos": ["Arguably, indexing a Tensor by a LongTensor with no dimension should return a LongTensor with no dimension. Consider the following case: This breaks indexing by indices when the list of indices is empty, which can be hard to detect. In my use-case, I use such a masking to select rows of a tensor corresponding to \"valid\" entries in a vector; I had not considered that it would not work when there are no valid entries.\nThanks! I'll take a look."], "neg": []}
{"query": "} #define RUN_T(INDEX_T)                                                   <del> TensorInfo<real, INDEX_T> inputInfo =                                 getTensorInfo<THCTensor, INDEX_T>(state, input);                  TensorInfo<real, INDEX_T> topKInfo =                                  getTensorInfo<THCTensor, INDEX_T>(state, topK);                   </del> <ins> TensorInfo<real, INDEX_T> inputInfo =                                  getTensorInfo<THCTensor, INDEX_T>(state, input);                     TensorInfo<real, INDEX_T> topKInfo =                                   getTensorInfo<THCTensor, INDEX_T>(state, topK);                      </ins> TensorInfo<int64_t, INDEX_T> indicesInfo =                             getTensorInfo<THCudaLongTensor, INDEX_T>(state, indices);            ", "pos": ["Repro: Expect two small tensors of small integers to be printed. Instead, only the first case works. In the second case, you get a series of failed assertions starting with this: I wonder if it is related to\nis this still an outstanding issue? I cannot reproduce this with an install from the master branch.\nThe size which causes failure doesn't seem deterministic. Using a pretty recent build (a few days old), I now have to increase the \"9\" to a larger number to make it fail. Sometimes it will fail at 10, sometimes at 11. Here's an updated case that seems to fail reliably at some size, for me:"], "neg": []}
{"query": "{\"_write_file\", (PyCFunction)THPStorage_(writeFile), METH_O, NULL}, {\"_new_with_file\", (PyCFunction)THPStorage_(newWithFile), METH_O | METH_STATIC, NULL}, #ifndef THC_GENERIC_FILE <ins> {\"from_buffer\", (PyCFunction)THPStorage_(fromBuffer), METH_VARARGS | METH_KEYWORDS | METH_STATIC, NULL}, </ins> {\"_share\", (PyCFunction)THPStorage_(_share), METH_NOARGS, NULL}, {\"_new_shared\", (PyCFunction)THPStorage_(_newShared), METH_VARARGS | METH_STATIC, NULL}, #endif", "pos": ["<!-- A clear and concise description of what the bug is. --The target name is used by a test executable in mkldnn. Steps to reproduce the behavior: BUILDTEST=1 build <!-- If you have a code sample, error messages, stack traces, please provide it here as well --<!-- A clear and concise description of what you expected to happen. --CMake configure passes. Please copy and paste the output from our (or fill out the checklist below manually). You can get the script and run it with: PyTorch Version (e.g., 1.0): v1.5.0-rc2 OS (e.g., Linux): Windows How you installed PyTorch (, , source): source Build command you used (if compiling from source): python build Python version: 3.x CUDA/cuDNN version: 9.2 / 10.1 / 10.2 / NA GPU models and configuration: Tesla M60 Any other relevant information: <!-- Add any other context about the problem here. --We should not build tests for mkldnn.\nI have encountered the same error, how do I resolve it?\nIt is actually mkldnn that is out of sync. So running the following script should help.\nThanks a lot for this I posted this on the PyTorch Discussions but to no avail, and now I've finally understood the reason of this error. Thanks a lot. Surprisingly, for me the build works out if I build from the master repo, but changing to a different version always fails."], "neg": []}
{"query": ".format(torch.typename(value), name)) modules[name] = value else: <del> object.__setattr__(self, name, value) </del> <ins> buffers = self.__dict__.get('_buffers') if buffers is not None and name in buffers: if value is not None and not torch.is_tensor(value): raise TypeError(\"cannot assign '{}' as buffer '{}' \" \"(torch.Tensor or None expected)\" .format(torch.typename(value), name)) buffers[name] = value else: object.__setattr__(self, name, value) </ins> def __delattr__(self, name): if name in self._parameters: del self._parameters[name] <ins> elif name in self._buffers: del self._buffers[name] </ins> elif name in self._modules: del self._modules[name] else:", "pos": ["They do one pair of these calls for each slicing operation.\nDoes this actually have a performance impact? These calls tend to be very quick\nYes, it's a major reason why small unrolled LSTMs (which involve in addition to two GEMMs: a chunk into four gates, for 4 pairs of calls, five unary pointwise ops, for 5 more, and five binary pointwise ops, for another 5) are much slower than they are in Theano and TensorFlow.\nAlso, there are excessive cudaGetDevice calls for each pointwise operation (9 cudaGetDevice) even without chunking, as in the following repro code: As for small unrolled LSTMs it does show up.\nThis should no longer be an issue, now that is implemented as a native ATen function. I think we're still using the Python version for tensors, but this will be resolved soon when we stop using tensors altogether."], "neg": []}
{"query": "namespace torch { namespace autograd { <ins> extern const char* ERR_BACKWARD_TWICE; </ins> struct Variable : std::enable_shared_from_this<Variable> { struct SavedVariable {", "pos": ["Usually calling backwards twice gives this message: But conv2d just segfaults. $ pt A B Segmentation fault (core dumped)"], "neg": []}
{"query": "INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR}) ADD_LIBRARY(THPP SHARED ${all_cpp}) <del> SET_PROPERTY(TARGET THPP PROPERTY CXX_STANDARD 11) </del> SET_TARGET_PROPERTIES(THPP PROPERTIES VERSION 1 SOVERSION 1) <ins> if(NOT ${CMAKE_VERSION} VERSION_LESS \"3.1\") SET_PROPERTY(TARGET THPP PROPERTY CXX_STANDARD 11) endif(NOT ${CMAKE_VERSION} VERSION_LESS \"3.1\") </ins> TARGET_LINK_LIBRARIES(THPP PRIVATE ${TH_LIBRARIES}) INSTALL(TARGETS THPP", "pos": ["Currently, CMake 3.0 or higher is required for THPP. This is annoying for building from source on Ubuntu 14.04 systems, as default CMake is 2.8, and updating it requires either compiling from source or adding some random repos:\nit should work fine if you change it to 2.8. I'm testing this out in a sec, and will push it into master\nNope, it does not for me."], "neg": []}
{"query": "<ins> #pragma once #include <Python.h> template<class T> class THPPointer { public: THPPointer(): ptr(nullptr) {}; THPPointer(T *ptr): ptr(ptr) {}; THPPointer(THPPointer &&p) { free(); ptr = p.ptr; p.ptr = nullptr; }; ~THPPointer() { free(); }; T * get() { return ptr; } const T * get() const { return ptr; } T * release() { T *tmp = ptr; ptr = NULL; return tmp; } operator T*() { return ptr; } THPPointer& operator =(T *new_ptr) { free(); ptr = new_ptr; return *this; } THPPointer& operator =(THPPointer &&p) { free(); ptr = p.ptr; p.ptr = nullptr; return *this; } T * operator ->() { return ptr; } operator bool() { return ptr != nullptr; } private: void free(); T *ptr = nullptr; }; typedef THPPointer<PyObject> THPObjectPtr; </ins>", "pos": ["Go to Look at the html. It contains the following: ! This problem manifests in google search. Searching for 1.8.1 docs brings up the \"1.8.0 docs\": ! We can probably fix this by doing a quick sed on the website but we should develop a solution for future releases. I think this has happened before in the past and am not sure what causes it. cc"], "neg": []}
{"query": "add_docstr(torch._C.tril, \"\"\" <del> tril(input, k=0, out=None) -> Tensor </del> <ins> tril(input, diagonal=0, out=None) -> Tensor </ins> Returns the lower triangular part of the matrix (2D Tensor) :attr:`input`, the other elements of the result Tensor :attr:`out` are set to 0.", "pos": ["The states that the argument controls which diagonal to consider. However, this argument is actually named in pytorch 0.2.0."], "neg": []}
{"query": "fi # sccache will fail for CUDA builds if all cores are used for compiling <del> # gcc 7.2 with sccache seems to have intermittent OOM issue if all cores are used if ([[ \"$BUILD_ENVIRONMENT\" == *cuda* ]] || [[ \"$BUILD_ENVIRONMENT\" == *gcc7.2* ]]) && which sccache > /dev/null; then </del> <ins> # gcc 7 with sccache seems to have intermittent OOM issue if all cores are used if ([[ \"$BUILD_ENVIRONMENT\" == *cuda* ]] || [[ \"$BUILD_ENVIRONMENT\" == *gcc7* ]]) && which sccache > /dev/null; then </ins> export MAX_JOBS=`expr $(nproc) - 1` fi", "pos": ["We're not testing against CUDA 9.2 and we should. C.f. We should test against GCC 7 (because this is how Anaconda is building the packages), AND we should use test with this in OS X, because CUDA 9.2 now supports the default xcode shipped on OS X.\nBuilding PyTorch with CUDA 9.2 for 7.0 compute compatibility devices produces quite worrisome warnings: warning : Instruction 'vote' without '.sync' may produce unpredictable results on sm70 and later architectures warning : Instruction 'vote' without '.sync' is deprecated since PTX ISA version 6.0 and will be discontinued in a future PTX ISA version warning : Instruction 'shfl' without '.sync' is deprecated since PTX ISA version 6.0 and will be discontinued in a future PTX ISA version warning : Instruction 'shfl' without '.sync' may produce unpredictable results on sm70 and later architectures\nCUDA 9.2 CI on macOS is done in"], "neg": []}
{"query": "from torch.autograd import gradcheck from torch.autograd.function import once_differentiable <del> from common import TestCase, run_tests </del> <ins> from common import TestCase, run_tests, skipIfNoLapack </ins> from torch.autograd._functions import * from torch.autograd import Variable, Function", "pos": ["Command to install the latest nightly binary: installs : However, PyTorch itself reports : I guess this version tag might be defined during the build process from e.g. and then forwarded from , but also source builds should use based on the tag. Unsure, if this is just a reporting issue or if we are building the nightly binaries with and executing them with , so creating this issue to dig a bit further into it. CC cc\nThis is really weird, I've put up measure in place to validate against it... Will investigate today (and moving to pytorch repo)\ncan you please check that 2.2.0 reports (and is build) against the correct version of nccl library?\nConfirmed for nightlies : Same issue for RC:\nLooking at the build log Looks like we are compiling with: USESYSTEMNCCL\nOn Windows I see following exception: We compiling with should we return None rather then exception ?\nIn this case we are indeed using 2.18.1: Do you know why we are using the system-wide NCCL installation instead of the one? Should we delete the system NCCL and use by default?\nLooks like its set here: Looks like it was by: Let me propose the PR to unset this flag : USESYSTEM_NCCL\nI guess we are using the system NCCL installation to be able to during the runtime. If we would use the module I assume we would link NCCL into the PyTorch binaries. If so, we should make sure to update the NCCL version whenever is updated. This could break our dependency on the NCCL pip package.\nYes. Also when we update this nccl here:\nWe used following path  What would be good path for nccl 2.19.3 and cuda 12.1 ? As per this issue: Will be using 12.0 ncck 2.19.3\nInstalling system NCCL into the nightly build docker container feels great, but we need to figure out, how to keep it in sync with PyTorch. I wonder if one can, for example, just check nccl version from this requirements file and build from source?\nyes probably we should build and upload nccl for our supported cuda versions, same way as we build docker containers, by following details here:\nVerified with final rc"], "neg": []}
{"query": "// Each block handles one distribution // First pass, find the total sum of the distribution AccT sum = accZero; <ins> T val; </ins> for (int cat = threadIdx.x; cat < categories; cat += blockDim.x) { <del> sum = THCNumerics<AccT>::add( sum, ScalarConvert<T, AccT>::to(dist[curDist * categories + cat])); </del> <ins> val = dist[curDist * categories + cat]; assert(THCNumerics<T>::ge(val, zero)); sum = THCNumerics<AccT>::add(sum, ScalarConvert<T, AccT>::to(val)); </ins> } // threadIdx.x == 0 has the sum value from this", "pos": ["Looking at the values returned from tn.multinomial(1) and counting them, it seems like it returns the values of the multinomial distribution taken from t, i.e. that it has ignored the signs of the values. However, it seem like always returns a tensor filled with zeros.\nlooking into this"], "neg": []}
{"query": "self.assertEqual(hy.data[0][0][0], 10) self.assertEqual(hy.data[1][0][0], output_val) <del> @unittest.skipIf(not TEST_CUDNN, \"needs cudnn\") </del> <ins> @unittest.skipIf(not (TEST_CUDNN and TEST_CUDNN_VERSION >= 5103), \"needs cudnn >= 5.1\") </ins> def test_RNN_dropout_state(self): import sys if sys.version_info[0] == 2:", "pos": ["cuDNN backend for RNN stabilizes at large mem usage. The following code stabilizes for me at 14+GiB reported by nvidia-smi. The same RNN without cuDNN uses <2.5GiB.\nOops... any idea what's going on with the garbage collector in this case?\nare you running this on a P100 GPU?\nYes, I observe similar behavior on Titan X Pascal\nhmmm, do cudnn RNNs have weird memory-oversubscribing properties? If it wasn't for CuDNN RNN, then it's not unreasonable. we use a custom CUDA allocator that will oversubscribe if needed, but will free the memory blocks it's holding onto if it hits OOM / failed allocations.\nTried the same RNN in torch and am getting ~2.6GiB on nvidia-smi.\nWith cuDNN I'm getting a constant usage of 2241MB on my machine using PyTorch (M40).\nare you using cudnnv5, or v6?\nv5\nno, cudnn does not do memory oversubscribing, and already tried the same network in lua torch with cudnn, where it gets expected memory usage. Also, he is getting OOMs on slightly bigger networks, where apparently caching allocator should have tried to free previous blocks. Interesting that can not repro this.\nJust to make sure did you set when you ran? I left it as False in the repro.\nI can repro in the docker image built from the current version on P100, and I go OOM after 8 batches with the same image on 1080, which generally should not happen - if there is enough memory to go through 1 batch, memory usage should not increase afterwards.\nNo, I just copied your script and ran it. Weird thing is that it goes OOM after 18 batches once I enable . Without it memory remains stable.\nA lot of cudaMalloc allocations go not through THCudaMalloc, and never get freed:\nSo not going through THCudaMalloc is actually expected for storages allocated with THCStorage_(newWithSize), but what's not expected is to every time cudaMalloc bytes of rnn workspace, instead of reusing earlier allocations. (~784 MB allocations are for workspace, ~156MB allocations happen somewhere in criterion).\nMy guess is that the bug is at the Python level.", "Either: Something is holding on to a Tensor longer than it should Something is creating a circular reference which prevents quick clean-up\nThe issue is that in creates cyclic references. It's made worse by the fact that it holds on to more memory than it should (like the workspace). You can see this by adding a as the first line in your for loop. With the explicit call to the GC every iteration, the memory usage doesn't explode.\nhow can these cyclic references be fixed?\nNot sure yet, but I will fix it tomorrow with"], "neg": []}
{"query": "modules = list(self._modules.keys()) buffers = list(self._buffers.keys()) keys = module_attrs + attrs + parameters + modules + buffers <ins> # Eliminate attrs that are not legal Python variable names keys = [key for key in keys if not key[0].isdigit()] </ins> return sorted(keys)", "pos": ["Hello, In the case where these classes sequentially append the received elements with , the modules become inaccessible since a variable name starting with a number is forbidden in Python. I attached a patch that names the modules as instead of just numbers.\nUse syntax for Sequential and ModueList.\npossible but the attributes are there and they are obviously shown in autocompletion suggestions in IPython for example and when you try to access them you got an error because their naming does not conform to Python rules.\nPerhaps we should exclude them from"], "neg": []}
{"query": "<del> FROM nvidia/cuda:8.0-devel-ubuntu16.04 </del> <ins> FROM nvidia/cuda:8.0-cudnn6-devel-ubuntu16.04 </ins> RUN echo \"deb http://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1604/x86_64 /\" > /etc/apt/sources.list.d/nvidia-ml.list <del> ENV CUDNN_VERSION 6.0.20 </del> RUN apt-get update && apt-get install -y --no-install-recommends  build-essential  cmake  git  curl  <ins> vim  </ins> ca-certificates  libjpeg-dev  <del> libpng-dev  libcudnn6=$CUDNN_VERSION-1+cuda8.0  libcudnn6-dev=$CUDNN_VERSION-1+cuda8.0 &&  </del> <ins> libpng-dev && </ins> rm -rf /var/lib/apt/lists/* RUN curl -o ~/miniconda.sh -O  https://repo.continuum.io/miniconda/Miniconda3-4.2.12-Linux-x86_64.sh  && ", "pos": ["Torchvision is considered a basic requirement of the tutorials. Perhaps it makes sense to include it in the docker build.\nIt's easy to install it by command ;)\nIt's easy to install torch itself using a similar command? :). I was thinking a docker image should have torchvision included.\nRuntime docker file already has torchvision."], "neg": []}
{"query": "[[17, 7], [18, 2], [19, 7], [20, 1]], ]) res = dense._sparse_mask(x) <del> exp_v = self.ValueTensor([[7, 9], [14, 1], [14, 1], [3, 3], [20, 1]]) </del> <ins> exp_v = self.ValueTensor([[7, 9], [14, 1], [3, 3], [20, 1]]) </ins> expected = self.SparseTensor(i, exp_v, torch.Size([5, 4, 2])) self.assertEqual(res, expected)", "pos": ["and I talked about the intended semantics of sparsemask when given a sparse tensor as an argument, and we came to the following conclusions: with an uncoalesced sparse tensor does not make sense, full stop. Unlike values, where your update might be linear and work correctly on an uncoalesced sparse tensor, there isn't any good reason why you want to pluck out values from the dense tensor multiple times, it is basically always a bug. the (eventual) documentation for sparsemask, we will recommend calling coalesce before doing any operations involving sparsemask (or nonlinear updates to indices/values). should add a test to sparsemask which checks if the flag of the input tensor is true, and raise an error if it is not. This is a cheap safety check which will prevent people from forgetting to call coalesce (as in (2)). There is a performance consequence to calling coalesce every iteration of training. If the gradient is producing coalesced vectors, we can alleviate this problem by adding a preprocessing step to coalesce() which tests if the input is already coalesced, and bugs out in that case. However, I think this is out of scope for this particular ticket for now."], "neg": []}
{"query": "pytorch-linux-xenial-cuda9-cudnn7-py2-test pytorch-linux-xenial-cuda9-cudnn7-py3-build pytorch-linux-xenial-cuda9-cudnn7-py3-test <ins> pytorch-linux-xenial-cuda9.2-cudnn7-py3-gcc7-build pytorch-linux-xenial-cuda9.2-cudnn7-py3-gcc7-test </ins> pytorch-linux-xenial-py3-clang5-asan-build pytorch-linux-xenial-py3-clang5-asan-test pytorch-linux-trusty-py2.7.9-build", "pos": ["We're not testing against CUDA 9.2 and we should. C.f. We should test against GCC 7 (because this is how Anaconda is building the packages), AND we should use test with this in OS X, because CUDA 9.2 now supports the default xcode shipped on OS X.\nBuilding PyTorch with CUDA 9.2 for 7.0 compute compatibility devices produces quite worrisome warnings: warning : Instruction 'vote' without '.sync' may produce unpredictable results on sm70 and later architectures warning : Instruction 'vote' without '.sync' is deprecated since PTX ISA version 6.0 and will be discontinued in a future PTX ISA version warning : Instruction 'shfl' without '.sync' is deprecated since PTX ISA version 6.0 and will be discontinued in a future PTX ISA version warning : Instruction 'shfl' without '.sync' may produce unpredictable results on sm70 and later architectures\nCUDA 9.2 CI on macOS is done in"], "neg": []}
{"query": "x_clone ^= y self.assertEqual(x_clone, xor_result) <ins> def test_apply(self): x = torch.range(1, 5) res = x.clone().apply_(lambda k: k + k) self.assertEqual(res, x * 2) self.assertRaises(RuntimeError, lambda: x.apply_(lambda k: \"str\")) </ins> if __name__ == '__main__': unittest.main()", "pos": ["On Python 3.5.2 errors out with . map and map2 seem to be work just fine."], "neg": []}
{"query": "def test_broadcast_batched_matmul(self): TestTorch._test_broadcast_batched_matmul(self, lambda t: t.cuda()) <ins> def test_index(self): TestTorch._test_index(self, lambda t: t.cuda()) </ins> def test_advancedindex(self): TestTorch._test_advancedindex(self, lambda t: t.cuda())", "pos": ["Arguably, indexing a Tensor by a LongTensor with no dimension should return a LongTensor with no dimension. Consider the following case: This breaks indexing by indices when the list of indices is empty, which can be hard to detect. In my use-case, I use such a masking to select rows of a tensor corresponding to \"valid\" entries in a vector; I had not considered that it would not work when there are no valid entries.\nThanks! I'll take a look."], "neg": []}
{"query": "if probs is not None: self.probs = probs / probs.sum(-1, keepdim=True) else: <del> self.logits = logits - log_sum_exp(logits) </del> <ins> self.logits = logits - _log_sum_exp(logits) </ins> self._param = self.probs if probs is not None else self.logits self._num_events = self._param.size()[-1] batch_shape = self._param.size()[:-1] if self._param.ndimension() > 1 else torch.Size()", "pos": ["Leaving this issue here for higher visibility. Originally reported by here: Also mentioned by Piggybacking off of this issue, has strange behavior (reported here: )\nI have fixed this locally. Could I send a PR?\nYes, go ahead!"], "neg": []}
{"query": "y._backward_hooks['test'] = error b.backward(torch.ones(5, 5)) <ins> def test_requires_grad_inplace(self): a = Variable(torch.randn(5, 5)) b = Variable(torch.randn(5, 5), requires_grad=True) a += b self.assertTrue(a.requires_grad) # non-leaf Variable a = Variable(torch.randn(5, 5)) + 0 b = Variable(torch.randn(5, 5), requires_grad=True) a += b self.assertTrue(a.requires_grad) </ins> def test_previous_functions(self): x = Variable(torch.randn(5, 5), requires_grad=True) y = Variable(torch.randn(5, 5), requires_grad=True)", "pos": ["import torch as th from torch.autograd import Variable x = Variable(th.Tensor([1 , 2 , 3]) , requiresgrad = False) y = Variable(th.Tensor([100]) , requiresgrad = True) x = x.cuda(0) y = y.cuda(0) x[0] = y print(x.requiresgrad) If ((0) (0)) is executed then the output is False, True otherwise. I think the result that x.requiresgrad = True is more natural.\nI further checked some in-place operator in GPU mode(say, x = Variable(...).cuda(0) ), and find out that after the operation is done, the requiresgrad of the Variable doesn't change. For example: import torch as th from torch.autograd import Variable x = Variable(torch.Tensor([1]) , requiresgrad = False) y = Variable(torch.Tensor([2]) , requiresgrad = True) x = x.cuda(0) y = y.cuda(0) x.add(y) print(x.requires_grad) the output is False\nIt doesn't have to do with Variable holding a CUDA tensor, but rather with the fact that it's no longer a leaf. I've sent a fix in .\nThanks for the report!"], "neg": []}
{"query": "SET_PROPERTY(TARGET ATen PROPERTY CXX_STANDARD 11) endif(NOT ${CMAKE_VERSION} VERSION_LESS \"3.1\") <del> FIND_PACKAGE(BLAS) </del> IF(BLAS_FOUND) <del> SET(USE_BLAS 1) </del> IF ($ENV{TH_BINARY_BUILD}) MESSAGE(STATUS \"TH_BINARY_BUILD detected. Enabling special linkage.\") TARGET_LINK_LIBRARIES(ATen \"${BLAS_LIBRARIES};${BLAS_LIBRARIES};${BLAS_LIBRARIES}\") ELSE ($ENV{TH_BINARY_BUILD}) TARGET_LINK_LIBRARIES(ATen ${BLAS_LIBRARIES}) ENDIF ($ENV{TH_BINARY_BUILD}) <del> IF(BLAS_INFO STREQUAL \"mkl\") ADD_DEFINITIONS(-DTH_BLAS_MKL) ENDIF() </del> ENDIF(BLAS_FOUND) IF(LAPACK_FOUND)", "pos": ["currently compiled master does not use any blas for any ops and uses our lame fallback C code. I'm looking into it. Basically USE_BLAS is not being set in"], "neg": []}
{"query": "Args: input: Variable :math:`(N, C)` where `C = number of classes` target: Variable :math:`(N)` where each value is `0 <= targets[i] <= C-1` <del> weight (Variable, optional): a manual rescaling weight given to each class. If given, has to be a Variable of size \"nclasses\" </del> <ins> weight (Tensor, optional): a manual rescaling weight given to each class. If given, has to be a Tensor of size \"nclasses\" </ins> size_average (bool, optional): By default, the losses are averaged over observations for each minibatch. However, if the field sizeAverage is set to False, the losses are instead summed", "pos": ["The of nn.functional.crossentropy seems wrong. The parameter weight should be a tensor, rather than a Variable."], "neg": []}
{"query": "fn.w_desc = init_weight_descriptor(fn, fn.weight_buf) w = fn.weight_buf <del> if tuple(hx.size()) != hidden_size: raise RuntimeError('Expected hidden size {}, got {}'.format( hidden_size, tuple(hx.size()))) </del> if cx is not None and tuple(cx.size()) != hidden_size: raise RuntimeError('Expected cell size {}, got {}'.format( hidden_size, tuple(cx.size())))", "pos": ["I run RNN sucessfully with the code. and I run embedding example. Then, got RuntimeError. I set to use nothing seems wrong. And the error info seems not the dim problem? The error is so strange.\nWhat exactly do you mean by \"embedding example\"? Could you provide a code snippet that causes the RuntimeError?\nTo better reproduce the error, I use the official code example of The origin runs well: Due to use , I have to make batch first. And then , it raises error.\nThank you for the example! The say that setting means that the first dim of input is the batch, ie, input has size . however is unchanged, so the following should work: I do agree the error message could be better, though"], "neg": []}
{"query": "return storage <ins> def rebuild_storage_empty(cls): return cls() </ins> def reduce_storage(storage): from . import get_sharing_strategy if storage.is_cuda:", "pos": ["PyTorch version: 0.4.0a0+ Repro: Result:\nEmm. I fail to reproduce this bug in Python 3.5.0 and Mac OS 10.13.3. What's your system environment?\nThe bug is probably specific to the sharing mode. Mac OS uses by default."], "neg": []}
{"query": "pass def run(self): <ins> # Check if you remembered to check out submodules def check_file(f): if not os.path.exists(f): print(\"Could not find {}\".format(f)) print(\"Did you run 'git submodule update --init'?\") sys.exit(1) check_file(os.path.join(lib_path, \"gloo\", \"CMakeLists.txt\")) check_file(os.path.join(lib_path, \"nanopb\", \"CMakeLists.txt\")) check_file(os.path.join(lib_path, \"pybind11\", \"CMakeLists.txt\")) </ins> libs = [] if WITH_NCCL and not WITH_SYSTEM_NCCL: libs += ['nccl']", "pos": ["It seems like running has a dependency on first running . This command also seems to also error out when installing via a wheel. This creates a problem for tools like  which rely on the output of to examine dependencies. Is there anyway to allow this command to not error out when installing via a wheel? Full stack trace: I understand the preferred package manager for this project is conda but being officially endorsed from , pipenv has a fairly large audience:\nthanks do you know when's the next planned release?\nHopefully within a month"], "neg": []}
{"query": "torch.stack([reference[:, 2, 1], reference[:, 2, 3], reference[:, 2, 5]], 1)) lst = [list(range(i, i + 10)) for i in range(0, 100, 10)] <del> tensor = torch.DoubleTensor(lst) </del> <ins> tensor = conv_fn(torch.DoubleTensor(lst)) </ins> for _i in range(100): idx1_start = random.randrange(10) idx1_end = idx1_start + random.randrange(1, 10 - idx1_start + 1)", "pos": ["Arguably, indexing a Tensor by a LongTensor with no dimension should return a LongTensor with no dimension. Consider the following case: This breaks indexing by indices when the list of indices is empty, which can be hard to detect. In my use-case, I use such a masking to select rows of a tensor corresponding to \"valid\" entries in a vector; I had not considered that it would not work when there are no valid entries.\nThanks! I'll take a look."], "neg": []}
{"query": "<del> From PyTorch: Copyright (c) 2016-     Facebook, Inc            (Adam Paszke) Copyright (c) 2014-     Facebook, Inc            (Soumith Chintala) Copyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert) Copyright (c) 2012-2014 Deepmind Technologies    (Koray Kavukcuoglu) Copyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu) Copyright (c) 2011-2013 NYU                      (Clement Farabet) Copyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou, Iain Melvin, Jason Weston) Copyright (c) 2006      Idiap Research Institute (Samy Bengio) Copyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert, Samy Bengio, Johnny Mariethoz) From Caffe2: Copyright (c) 2016-present, Facebook Inc. All rights reserved. All contributions by Facebook: Copyright (c) 2016 Facebook Inc. All contributions by Google: Copyright (c) 2015 Google Inc. All rights reserved. All contributions by Yangqing Jia: Copyright (c) 2015 Yangqing Jia All rights reserved. All contributions from Caffe: Copyright(c) 2013, 2014, 2015, the respective contributors All rights reserved. All other contributions: Copyright(c) 2015, 2016 the respective contributors All rights reserved. Caffe2 uses a copyright model similar to Caffe: each contributor holds copyright over their contributions to Caffe2. The project versioning records all such contribution and copyright details. If a contributor wants to further mark their specific copyright on a particular contribution, they should indicate their copyright solely in the commit message of the change when it is committed. </del> ======================================================================= Software under third_party =======================================================================", "pos": ["The three-clause BSD license in file LICENSE says in clause 1 \"Redistributions of source code must retain the above copyright notice...\" However, there is no longer a copyright notice in the file; it appears to have been removed in commit (see ).\ncc"], "neg": []}
{"query": "\"torch/csrc/Generator.cpp\", \"torch/csrc/Size.cpp\", \"torch/csrc/Exceptions.cpp\", <del> \"torch/csrc/Tensor.cpp\", </del> \"torch/csrc/Storage.cpp\", \"torch/csrc/DynamicTypes.cpp\", \"torch/csrc/byte_order.cpp\",", "pos": ["Trying to run on github actions via AWS instance gives me: I will of course set up now, but it shouldn't crash like that. May be it should use as a fallback? pt-2.0 cc\nIs this a multitenant setup? Also curious which distro are you running, I was under the impression that most linux distros set home by default\nthis is AWS image. one default user. I think it's user, but it wants you to run as root, which is weird. I have manually now set to overcome this.\nWe should update triton to use pathlib or something"], "neg": []}
{"query": "is_master_only=True, requires=[\"binary_linux_manywheel_3_7m_cu102_devtoolset7_build\"], extra_props={ <del> \"resource_class\": \"gpu.medium\", </del> <ins> \"resource_class\": \"gpu.nvidia.small\", </ins> \"use_cuda_docker_runtime\": miniutils.quote((str(1))), }, ),", "pos": ["As it still using deprecated resource class, see See example of the failure in\nFollowup: we need to plan better --we knew it was gonna happen and none of us fixed it early enough. Action item: go through tasks about deprecation in sprint planning"], "neg": []}
{"query": "{ Py_VISIT(self->data); Py_VISIT(self->backward_hooks); <ins> // We don't want to traverse the grad_fn, even if the Variable owns it and the // shared pointer's use count is 1. This is because we would need to treat // the grad_fn as part of the Python state and hold the GIL sometimes when // grad_fn's shared_ptr is copied, otherwise a race condition with the Python // GC could occur. Holding the GIL when the shared_ptr is copied adds // undesirable complexity/overhead. // // When hooks, a Variable, and its grad_fn are involved in a Python reference // cycle, because we're not traversing the grad_fn, the reference cycle will // in fact leak. // // See https://gist.github.com/zou3519/7ac92b84dd7d206dcc6eae55fee8372c // for more details about the race condition involving traversing the grad_fn // and the python GC. </ins> if (self->cdata.defined()) { <del> // Only visit this if we actually own it (no one else use the shared pointer) auto& grad_fn = self->cdata.grad_fn(); if (grad_fn.use_count() == 1) { if (auto fn = dynamic_cast<PyFunction*>(grad_fn.get())) { Py_VISIT(fn->obj); } else { // visit hooks in C++ implemented autograd functions for (auto& hook : grad_fn->pre_hooks) { if (auto pyhook = dynamic_cast<PyFunctionPreHook*>(hook.get())) { Py_VISIT(pyhook->dict); } } for (auto& hook : grad_fn->post_hooks) { if (auto pyhook = dynamic_cast<PyFunctionPostHook*>(hook.get())) { Py_VISIT(pyhook->dict); } } } } </del> for (auto& hook : self->cdata.hooks()) { if (auto pyhook = dynamic_cast<PyFunctionPreHook*>(hook.get())) { Py_VISIT(pyhook->dict);", "pos": ["My code runs well for a while , it occurs randomly..... Traceback (most recent call last): File \"\", line 236, in &lt;moduletrainvgg(\"\",dset='luna16') File \"\", line 54, in trainvgg lossvalue.backward() File \"/home/uih/anaconda2/lib/python2.7/site-\", line 169, in backward torch.autograd.backward(self, gradient, retaingraph, creategraph, retainvariables) File \"/home/uih/anaconda2/lib/python2.7/site-\", line 99, in backward variables, gradvariables, retaingraph) RuntimeError: add: Assertion \"pos = 0 &amp;&amp; pos &lt; ()\" failed. Any idea what's happend? Thanks in advance~\nYes, it looks like we have a bug somewhere in the code. Can you please try to isolate a small snippet that still triggers the same error so we can look into it? Thank you!\nIt's hard for me to isolate, the error occurs when I do the train and validation in turn, if I just do either the train or val, everthing runs well.\nThe problem is that we won\u2019t be able to fix it, until we actually see under what conditions does the error happen\nI got the same error in v0.3.0 today. when I trained my model with single GPU, it worked perfectly;with multi-gpu the error occured. RuntimeError: add: Assertion failed.\nWe've also been struggling with this issues for a long time. Unfortunately, finding which part of our code specifically is triggering the assert (which used to be a segfault before ) has proven extremely hard and, as of now, unsuccessful. In fact, we never submitted a bug report as we couldn't create a self-contained piece of code to trigger it. I'll try to summarize all the information we've gathered: We know for a fact that the issue appeared sometime after this commit: . All our code works perfectly fine on or v0.2. The assert is always triggered during the backward pass, but the exact moment when it happens seems to be completely random: when training over Imagenet we've seen this after 50 iterations as well as after 5000.", "This makes it super-hard to understand whether a particular piece of code is triggering the assert or not, as you can never be sure how many iterations are enough to exclude that the assert will ever trigger! The issue seems to be somehow related to the use of multiple GPUs (via ): we've never observed it while training with a single GPU. The issue seems also to be related to the use of user-defined classes: we've never seen it when training networks composed of \"standard\" Pytorch functions / modules only... ... however, simply running forward / backward passes with our custom functions doesn't seem to be enough to trigger the assertion. In our experience, working with \"complex\" computational graphs increases the likelihood of encountering this issue. Recently we've been able to re-create the issue using and a modified version of , where we replace some of the layers with our custom-made ones. Unfortunately I can't share the code for now (our paper describing the new layers is still not out), but I'll come back to you as soon as I can! Some additional information: all our experiments are on Ubuntu 16.04 using CUDA 8, CUDNN 7 and TITAN X and Xp GPUs.\nalmost the same condition as us\nEncounter the same condition after serveral imagenet epochs on a NasNet after update to 3.0. Pytorch 2.0 is working well.\nI am seeing this too -- I have two runs of the same model running with different hyperparameters, one is running smoothly, this one quit with the exact same error avgloss = self.trainepoch(trainingdata,validationdata,batchsize,saveprefix,epoch) File \"\", line 759, in trainepoch gradients = loss.backward() File \"/u/dcontrac/anaconda2/lib/python2.7/site-\", line 167, in backward torch.autograd.backward(self, gradient, retaingraph, creategraph, retainvariables) File \"/u/dcontrac/anaconda2/lib/python2.7/site-\", line 99, in backward variables, gradvariables, retaingraph) RuntimeError: add: Assertion failed. Running PyTorch v0.3 and training on 4 K-80 GPUs. My other run with different parameters is also running on the exact same setup but its still running.", "Both runs differ only in shuffle of data and a different batch size (the one that failed had a slightly larger batch size)\n: I'm not using any user define autograd.Function but I see the error too, so maybe thats not the primary place with a bug. Using nn.DataParallel though.\nI'll try again to train a network using standard layers only, maybe I'll also encounter the error in the no-custom-functions case if I let it run a little bit longer. If this is the case, then I guess we have some strong indication that (or multi-GPU computation in general) could indeed be the source of the error!\n: I dont think the duration has got anything to do with it --- One model run has been running without a problem for the last 18 hours, the other one had quit last night pretty early on (Crashed after the backward pass of the 8th batch). I have just restarted it and watching it.\nI agree that the bug is probably not related to training length. My guess is that each backward pass has a certain probability of randomly failing with this assert, so running for longer just increases the overall chances of encountering the bug.\nThanks for the information. We'd really like to fix this bug, and so there are a few things that would help us out a lot (in order of difficulty): you run your training process with gdb attached with a breakpoint, so that we can get a backtrace when the assertion fails? Also, if you can go one frame up and print the name of the function (run in gdb) that would also be helpful. We basically want to know if it's either a C++ function or a Python function. soon as you can publish your latest layers, uploading a script that reproduces (even if it takes a while to reproduce); we would be happy to run the script to try to reproduce, which would be very helpful for us. We'd love a bisection, but given that it can sometimes take 18hrs+ to repro, it doesn't sound like that would be easy to do.\nThis is the backtrace: For the function name I'll probably have to recompile pytorch in debug mode, since in gdb returns . Edit - I managed to get the name of the function by catching the exception raised by the assert: . Edit 2 - I did a second run and the name of the function is still the same.", "Actually I just realized that what we really need is Can you please try printing that (might require a debug build as well)?\nPrinting instead of I get , which is the name of our custom function's backward pass. More importantly, our code and paper are now public so I can share this with you: As I mentioned in a previous comment, this code is just the Pytorch ImageNet + ResNet examples, except that standard BN is replaced with our custom layer. When running this on our machines with 4 GPUs we always encounter the assert within the first epoch. Please feel free to ask if you need help running the code.\nI also encountered this assert. I have two machines, one with 2, and other one with 4 GPUs. Both machines are Ubuntu 16.04, python 2.7, cuda 9.0, cudnn 7, and pytorch v0.3.0. I never encountered this assert on my 2 GPUs machine. This assert always appeared on my 4 GPUs machine.\nsame as here. I have python 2.7, cuda 8.0, cudnn 7 and pytorch v0.3.0 installed from binaries. The assertion was triggered when I was training networks on 4 GPUs. It happened at a random time during training.\nSame here, with multiple pixel-wise prediction models with just CNNs. Seems to occur only with multiple GPUs, has not happened so far with single GPU. python 2.7, CUDA 8.0, CUDNN 7, PyTorch v0.3.0.\ncan you take a look at this asap\nedit: resolved, I ended up adding a compilation flag for compute capacity 3.5 to pytorchbugreport's building scripts (to work with my gpus) I'm getting the following after trying to run pytorchbugreport. Have you seen this before? !\nyou have to change to cover your GPU, which is compute35,code=sm_35 i think (if it is K40)\nI ran your code for six epochs on a machine with 4 gpus on pytorch 0.3.0 but haven't been able to trigger the assert. I'll try letting it run for longer and see if anything pops up\nFor us the assert almost always triggers during the first epoch.", "There might be some important difference in our setups. This is the exact sequence of steps we have used to reproduce the error across several machines: from a clean Ubuntu 16.04 installation with CUDA 8.0, CUDNN 7.0, pip, virtualenv and the compilers. and activate a new python2 virtualenv. necessary python packages . the pytorch repository and checkout and install with our repository and compile our native module with the training as described before. Our machines all have either Titan X or Titan Xp GPUs, so a different architecture from the one you are using, I don't know if this could have an influence on the error.\nMy machine has 4 GTX 1080 Ti which triggers the error.\nwe (finally) found a different repro that triggers this issue internally on one of our machines, a fix will be issued soon.\nQuick update: we're still looking into this, but so far, I've been able to get this bug to trigger with two machines using python 2.7 but never with python 3. If you're in need of a quick fix using python 3 might help.\nI also encountered this propblem, with pytorch 0.3.0, python2.7, CUDA8.0, Titan X GPUs. My session always crashed around 8 epochs. And I degrade pytorch to 0.2.0. now the training seems work well. Someone in emergency can try this.\nThank you , I'll try to switch to python 3 for now!\nlet me know if the bug comes back in python 3. As far as I can tell the python version shouldn't matter in the triggering of this bug... but it's never happened to me with python 3 yet.\nCrashed in Python2.7 RuntimeError: add: Assertion failed Absolutely random crash using multi gpu.\nThis has been fixed in HEAD but there has not been a release of 0.3 with the fix yet.\nOh, Thanks for your reply. I will try the HEAD version. Thank you .\ni have encountered this error , but change the pytorch to the latest version(0.4) will solve it.\nEncountered same error when I add dropout layer to a Resnet which I modified. Trained on multi-GPU. Does latest conda release fix the problem?\nit is not yet on latest conda release.", "to get the fix at the moment you will have to install pytorch from source:\nThe same bug ! !\ndid you install pytorch from source?\nIs there a simple way to block this error without changingthe version of python or pyotrch?\nYou could disable the python reference cycle . This would probably lead to memory leaks though.\nIt should be safe if you do this once all the graphs are deleted, and there will be no leaks unless you create cycles in your data structures\nAfter 2 days of struggle, I finally changed back to 0.2.0 and expects the release of binary 0.4.0. Anyway, thank you for your advices. :)\nI encountered the same error with version 0.3.0.post4. What was the proper fix here?\nupgrade to version 0.3.1 or newer."], "neg": []}
{"query": "class Dropout(Module): <del> r\"\"\"Randomly zeroes some of the elements of the input tensor. </del> <ins> r\"\"\"During training, randomly zeroes some of the elements of the input tensor with probability *p* using samples from a bernoulli distribution. </ins> The elements to zero are randomized on every forward call. <ins> This has proven to be an effective technique for regularization and preventing the co-adaptation of neurons as described in the paper `Improving neural networks by preventing co-adaptation of feature detectors`_ . Furthermore, the outputs are scaled by a factor of *1/(1-p)* during training. This means that during evaluation the module simply computes an identity function. </ins> Args: p: probability of an element to be zeroed. Default: 0.5 inplace: If set to True, will do this operation in-place. Default: false", "pos": ["\"Randomly zeroes some of the elements of the input tensor. The elements to zero are randomized on every forward call.\" This is incorrect; the function also scales up by 1/(1-p), which the implementation correctly does."], "neg": []}
{"query": "del l1[i] return l1 <del> def _backward(self, method, input, gradOutput, scale): </del> <ins> def _backward(self, method, input, gradOutput, scale=1): </ins> isTable = isinstance(input, list) wasTable = isinstance(self.gradInput, list) if isTable: for i, module in enumerate(self.modules): <del> currentGradInput = getattr(module, method)(input, gradOutput[i], scale) </del> <ins> if method == 'updateGradInput': currentGradInput = module.updateGradInput(input, gradOutput[i]) elif method == 'backward': currentGradInput = module.backward(input, gradOutput[i], scale) </ins> if not isinstance(currentGradInput, list): raise RuntimeError(\"currentGradInput is not a table!\")", "pos": ["Subclassing torch.Tensor with fails due to a on the latest nightly builds, whereas it previously worked without issue. Run this snippet of code: And you'll recieve an error message like this: Changing the to results in the exact same error message. There should be no error message as it should be working like it was previously in Torch 1.8.1 and below. I think that the issue started on Torch version: , and I can confirm that it still exists as of , . The issue started on the nightly builds at least 2 days ago, if that helps resolve things. It's been causing some of Captum's CircleCI tests to fail, and the tests show the first occurrence of the error on . cc\nThis is due to either or\nI think the problem is I'm not supposed to delete because that means the super() call doesn't work. Instead, I should just verify that the passed in class is a subclass, then this should work."], "neg": []}
{"query": "-DCMAKE_CXX_FLAGS=\"$C_FLAGS $CPP_FLAGS\" make install cp \"lib/libnccl.so.1\" \"${INSTALL_DIR}/lib/libnccl.so.1\" <del> ln -s \"${INSTALL_DIR}/lib/libnccl.so.1\" \"${INSTALL_DIR}/lib/libnccl.so\" </del> <ins> if [ ! -f \"${INSTALL_DIR}/lib/libnccl.so\" ]; then ln -s \"${INSTALL_DIR}/lib/libnccl.so.1\" \"${INSTALL_DIR}/lib/libnccl.so\" fi </ins> cd ../.. }", "pos": ["there is a bug. the code is from The code for installing cuda10.2 should be corrected to The code for installing cuda10.2 should be corrected to cc\nThank you for reporting the typo", "This is how I implement the decoder of a sequence to sequence model I'm not sure about the new autograd mechanics but this worked in the previous version. If I didn't make the unrolling codes a function it will work. It will also work with CPU. I compiled from source with Python 2.7, CUDA 8.0 and Cudnn 6\nThanks for the minimal reproduction! It broke OpenNMT and a big internal model but we weren't looking forward to isolating what parts of those had the issue.\nThat's great, I'll take a look today. Thanks!\nI've pushed a fix to .\nHow does the inference work in this case? I mean, once you train it how do you test it in the test phase?"], "neg": []}
{"query": "For summation index :math:`j` given by `dim` and other indices :math:`i`, the result is <del> :math:`text{logsumexp}(x)_{i} = log sum_j exp(x_ij).` </del> <ins> .. math:: text{logsumexp}(x)_{i} = log sum_j exp(x_{ij}) </ins> If :attr:`keepdim` is ``True``, the output tensor is of the same size as :attr:`input` except in the dimension :attr:`dim` where it is of size 1.", "pos": ["Leaving this issue here for higher visibility. Originally reported by here: Also mentioned by Piggybacking off of this issue, has strange behavior (reported here: )\nI have fixed this locally. Could I send a PR?\nYes, go ahead!"], "neg": []}
{"query": "return g.op(\"Undefined\") <ins> def Constant(g, value): return g.op(\"Constant\", value_t=value) </ins> def add(g, self, other, alpha): if _scalar(alpha) != 1: return _unimplemented(\"add\", \"alpha != 1\")", "pos": ["Hi! OS: Linux PyTorch version: 0.4.0a0+ How you installed PyTorch (conda, pip, source): source Python version: 3.6Context around what you are trying to do I am trying to convert layer with torch.indexselect() operation to onnx A script to reproduce the bug. Please try to provide as minimal of a test case as possible.Error messages and/or stack traces of the bug What is the right way to export torch.indexselect() operation?\nIt kind of sounds like we forgot to implement constant export, which is embarrassing and should be fixed.\nAre there any chances it will be fixed soon?\nI took a look, and actually, it's just a spurious warning, the model seems to export successfully into"], "neg": []}
{"query": "return THC_float2half(0); #else THError(\"Cublas_Hdot requires CUDA 8.0+\"); <del> return THC_half2float(0); </del> <ins> return THC_float2half(0); </ins> #endif } #endif", "pos": ["Hi, I was trying to update from source but I get an error \"no suitable constructor exists to convert from \"int\" to \"__half\"\" when trying to build I am working with python 2.7 cuda 7.5 and GCC 4.8.3 in openSUSE. Any suggest? Here is the output error when building from source:\nATen Excluded: set(['bernoulli', 'bernoulli']) [ 0%] Built target atenfilesaregenerated [ 0%] Building NVCC (Device) object [ 1%] Building NVCC (Device) object [ 1%] Building NVCC (Device) object [ 1%] Building NVCC (Device) object [ 1%] Building NVCC (Device) object [ 2%] Building NVCC (Device) object [ 2%] Building NVCC (Device) object [ 2%] Building NVCC (Device) object (77): error: no suitable constructor exists to convert from \"int\" to \"half\" (77): error: no suitable constructor exists to convert from \"float\" to \"half\" 2 errors detected in the compilation of \"/tmp/tmpxft0000587a00000000-\". CMake Error at (message): Error generating file recipe for target '' failed make[2]: [] Error 1 make[2]: Waiting for unfinished jobs.... CMakeFiles/Makefile2:194: recipe for target '' failed make[1]: [] Error 2 Makefile:129: recipe for target 'all' failed make: [all] Error 2\nNVIDIA cuDNN v6.x or above\nsame problem here. gcc (Ubuntu 4.8.4), python 2.7, cuda 7.5, cuDNN 6.0.21\nSame problem here. CUDA: v7.5 CUDNN: v6.0"], "neg": []}
{"query": "} }; <ins> // Similar to shared_from_this. There's a problem that the Python object // and its cdata depend on each other being alive, so we can't keep // shared_ptrs as members, but we'd like to be able to manage the lifetime of // the objects using shared_ptrs in the C++ graph. The only way to get a new // shared_ptr that references them is through THPFunction_asFunction. When // called for the first time it will allocate a new shared_ptr and save a // weak_ptr in cdata_ptr attr. Later, when we try to take another reference, // we'll try to lock cdata_ptr and return its value if successful. Otherwise it // means that all shared_ptrs returned previously have been freed, so we can // create a new one. This ensures that this object is managed by at most one // shared_ptr control block at any time - a guarantee we depend on in other places // (e.g. we use weak_ptrs in SavedVariable because we know it won't go out of scope). </ins> std::shared_ptr<PyFunction> THPFunction_asFunction(THPFunction* self) { if (!self) { return std::shared_ptr<PyFunction>(); } Py_INCREF((PyObject*)self); <del> return std::shared_ptr<PyFunction>(&self->cdata, Decref()); </del> <ins> auto ptr = self->cdata_ptr.lock(); if (ptr) return ptr; ptr = std::shared_ptr<PyFunction>(&self->cdata, Decref()); self->cdata_ptr = ptr; return ptr; </ins> }", "pos": ["there is a bug. the code is from The code for installing cuda10.2 should be corrected to The code for installing cuda10.2 should be corrected to cc\nThank you for reporting the typo", "This is how I implement the decoder of a sequence to sequence model I'm not sure about the new autograd mechanics but this worked in the previous version. If I didn't make the unrolling codes a function it will work. It will also work with CPU. I compiled from source with Python 2.7, CUDA 8.0 and Cudnn 6\nThanks for the minimal reproduction! It broke OpenNMT and a big internal model but we weren't looking forward to isolating what parts of those had the issue.\nThat's great, I'll take a look today. Thanks!\nI've pushed a fix to .\nHow does the inference work in this case? I mean, once you train it how do you test it in the test phase?"], "neg": []}
{"query": "x.__repr__() str(x) <ins> x = torch.DoubleTensor([1e-324, 1e-323, 1e-322, 1e307, 1e308, 1e309]) x.__repr__() str(x), </ins> def test_unsqueeze(self): x = torch.randn(2, 3, 4) y = x.unsqueeze(1)", "pos": ["on print(torch.DoubleTensor(1)) I got ZeroDivisionError: float division by zero but torch.DoubleTensor(0)) or torch.DoubleTensor(2)) work just fine. I work in Jupyter notebook pytorch 0.1.12\nNote that returns undefined memory. Could you find the specific content of the tensor that gave this error, by printing ? I'm not sure if this bug has been fixed in the latest release already.\nthe error is pretty interesting. I restarted Jupyter notebook. # many freaking signatures: print(torch.FloatTensor(1)) print(torch.IntTensor(1)) print(torch.DoubleTensor(1)) # no error right after initial run. #print(torch.DoubleTensor(0)) print(torch.DoubleTensor(2)) print(torch.ByteTensor(1)) print(torch.CharTensor(1)) print(torch.ShortTensor(1)) print(torch.LongTensor(1)) steps to reproduce the error: 1). run, no errors 2) repeat, no error 3) uncomment ..DoubleTensor(0), run, got an error (see below): 4) comment ..DouleTensor(0), run still have this error 2.+24 [torch.FloatTensor of size 1] 7.+08 [torch.IntTensor of size 1] ZeroDivisionError Traceback (most recent call last) &lt;ipython-input-5-3629c250c505in &lt;module() 2 print(torch.FloatTensor(1)) 3 print(torch.IntTensor(1)) ----4 print(torch.DoubleTensor(1)) # ZeroDivisionError: float division by zero 5 #print(torch.DoubleTensor(0)) 6 print(torch.DoubleTensor(2)) C:UsersleogAppDataLocalcondacondaenvspy35libsite-packagestorch in str(self) 126 # characters to replace unicode characters with.", "127 if sys.versioninfo (3,): --128 return (self) 129 else: 130 if hasattr(, 'encoding'): C:UsersleogAppDataLocalcondacondaenvspy35libsite-packagestorch in str(self) 284 return '[{} with no dimension]n'.format(torch.typename(self)) 285 elif self.ndimension() == 1: --286 strt = vectorstr(self) 287 elif self.ndimension() == 2: 288 strt = matrixstr(self) C:UsersleogAppDataLocalcondacondaenvspy35libsite-packagestorch in vectorstr(self) 270 if () &lt; PRINTOPTS.threshold: 271 return (strt + --272 'n'.join(ident + (val / scale) for val in self) + 273 'n') 274 else: C:UsersleogAppDataLocalcondacondaenvspy35libsite-packagestorch in &lt;genexpr(.0) 270 if () &lt; PRINT_OPTS.threshold: 271 return (strt + --272 'n'.join(ident + (val / scale) for val in self) + 273 'n') 274 else: ZeroDivisionError: float division by zero\nOk, I got a repro. Here is the failing case: Fails with the error you mentioned. This seems to be an edge case though, as both and works. The issue is in , where becomes zero because is a too large negative number. Maybe we should clip to be positive?"], "neg": []}
{"query": "res2 += i * j self.assertEqual(res1, res2) <ins> # Test 0-strided for tname, _prec in types.items(): v1 = torch.randn(1).type(tname).expand(100) v2 = torch.randn(100).type(tname) res1 = torch.dot(v1, v2) res2 = 0 for i, j in zip(v1, v2): res2 += i * j self.assertEqual(res1, res2) def test_ger(self): types = { 'torch.DoubleTensor': 1e-8, 'torch.FloatTensor': 1e-4, } for tname, _prec in types.items(): v1 = torch.randn(100).type(tname) v2 = torch.randn(100).type(tname) res1 = torch.ger(v1, v2) res2 = torch.zeros(100, 100).type(tname) for i in range(100): for j in range(100): res2[i, j] = v1[i] * v2[j] self.assertEqual(res1, res2) # Test 0-strided for tname, _prec in types.items(): v1 = torch.randn(1).type(tname).expand(100) v2 = torch.randn(100).type(tname) res1 = torch.ger(v1, v2) res2 = torch.zeros(100, 100).type(tname) for i in range(100): for j in range(100): res2[i, j] = v1[i] * v2[j] self.assertEqual(res1, res2) def test_addmv(self): types = { 'torch.DoubleTensor': 1e-8, 'torch.FloatTensor': 1e-4, } for tname, _prec in types.items(): t = torch.randn(10).type(tname) m = torch.randn(10, 100).type(tname) v = torch.randn(100).type(tname) res1 = torch.addmv(t, m, v) res2 = torch.zeros(10).type(tname) res2 += t for i in range(10): for j in range(100): res2[i] += m[i, j] * v[j] self.assertEqual(res1, res2) # Test 0-strided for tname, _prec in types.items(): t = torch.randn(1).type(tname).expand(10) m = torch.randn(10, 1).type(tname).expand(10, 100) v = torch.randn(100).type(tname) res1 = torch.addmv(t, m, v) res2 = torch.zeros(10).type(tname) res2 += t for i in range(10): for j in range(100): res2[i] += m[i, j] * v[j] self.assertEqual(res1, res2) def test_addmm(self): types = { 'torch.DoubleTensor': 1e-8, 'torch.FloatTensor': 1e-4, } for tname, _prec in types.items(): M = torch.randn(10, 25).type(tname) m1 = torch.randn(10, 50).type(tname) m2 = torch.randn(50, 25).type(tname) res1 = torch.addmm(M, m1, m2) res2 = torch.zeros(10, 25).type(tname) res2 += M for i in range(10): for j in range(25): for k in range(50): res2[i, j] += m1[i, k] * m2[k, j] self.assertEqual(res1, res2) # Test 0-strided for tname, _prec in types.items(): M = torch.randn(10, 1).type(tname).expand(10, 25) m1 = torch.randn(10, 1).type(tname).expand(10, 50) m2 = torch.randn(50, 25).type(tname) res1 = torch.addmm(M, m1, m2) res2 = torch.zeros(10, 25).type(tname) res2 += M for i in range(10): for j in range(25): for k in range(50): res2[i, j] += m1[i, k] * m2[k, j] self.assertEqual(res1, res2) </ins> def _testMath(self, torchfn, mathfn): size = (10, 5) # contiguous", "pos": ["It seems like autodiff of matrix terms of .mv() doesn't behave as I'd expect. But if we use .mm(), things work fine: python A = Variable((3,2), requiresgrad=True) x = Variable((2), requiresgrad=True) (A.mm(x[:,None])).sum().backward() print(A.grad) print(x.grad)\nHi, As a temporary fix, you can use . The problem is the behavior between the backward and the backward. The sum returns a strided gradient which is not supported by MKL/BLAS implementations of (used in the backward). not sure where this was introduced and how to fix this properly. Should the blas wrapper clone the input if it is strided? Or should it just raise an error (meaning that we need to add some s in some places in the code like the sum backward)?\nIt should clone the input. Still, this problem should have been caught. We must be missing some return code checks from BLAS\nAh apparently these functions have no return codes and the docs clearly state that vector strides can't be 0... We have to fix those bindings"], "neg": []}
{"query": "from ._utils import _type, _cuda, _range from functools import reduce from itertools import chain <ins> import sys </ins> import math", "pos": ["It seems that the following link that can be reached from here (check the screenshot) is broken: Page: Broken link: Screenshot: ! cc\nthis should be transferred to , but I don't have permissions to do that. Fix is available at pytorch/hub."], "neg": []}
{"query": "set(BLAS_SET_BY_USER ${BLAS_SET_BY_USER} CACHE STRING \"Marks whether BLAS was manually set by user or auto-detected\") endif() <ins> # These lines are an attempt to make find_package(cuda) pick up # libcuda.dylib, and not cuda.framework.  It doesn't work all # the time, but it seems to help for some users. # TODO: replace this with a more robust fix if(APPLE) set(CMAKE_FIND_FRAMEWORK LAST) set(CMAKE_FIND_APPBUNDLE LAST) endif() </ins> # ---[ Options. # Note to developers: if you add an option below, make sure you also add it to # cmake/Summary.cmake so that the summary prints out the option values.", "pos": ["error occurs when linking the cuda library ld: can't map file, errno=22 file '/usr/local/cuda/lib/stubs/cuda.framework' for architecture x8664 Provide a short description. I have tracked down in findlibrary(CUDACUDALIB cuda PATHS ${CUDATOOLKITROOTDIR} PATHSUFFIXES lib lib64 lib/stubs lib64/stubs lib/x64) finds the cuda.framework instead of Have confirm it was broken by this commit () Any idea why it is linking the cuda.framework? CUDA used to build PyTorch: 9.2 OS: Mac OSX 10.13.4 GCC version: Could not collect CMake version: version 3.11.1 Python version: 3.6 Is CUDA available: Yes CUDA runtime version: 9.2.64 GPU models and configuration: Could not collect Nvidia driver version: Could not collect cuDNN version: Probably one of the following: /usr/local/cuda/lib/libcudnn_static.a\ncc\nFrom cmake doc , CMake will use a -framework A, and a -F<fullPathto link the framework to the target, however it is not. I have tried separating ::cuda from Caffe2PUBLICCUDADEPENDENCY_LIBS, still not linking properly. Removing ::cuda passes the build though.\ncan you post your cmake output that contains the list of CUDA-related libraries that cmake found? Is this statically linking CUDA? What's the actual linking command that is failing (probably need to pass VERBOSE=1 to the make command to see this)? None of us will be able to repro on our end as we do not have macs with GPUs :( , but we can try to answer questions and give pointers as to where to look.\nthanks for reporting this and helping reproduce. Can you try and add this to the root around line 162? You'll then want to run to clear the directory and try again.", "[ 94%] Linking CXX shared library cd /Users/hongbinliu/pytorch/build/ && /usr/local/miniconda3/bin/cmake -E cmakelinkscript --verbose=1 /Library/Developer/CommandLineTools/usr/bin/clang++ -msse4.1 -msse4.2 --std=c++11 -Wno-deprecated -fvisibility-inlines-hidden -Wno-deprecated-declarations -DONNXNAMESPACE=onnxc2 -O2 -fPIC -Wno-narrowing -Wno-invalid-partial-specialization -Wall -Wextra -Wno-missing-field-initializers -Wno-type-limits -Wno-unused-parameter -Wno-unknown-warning-option -Wno-unknown-pragmas -DHAVEAVXCPUDEFINITION -DHAVEAVX2CPUDEFINITION -O3 -isysroot -mmacosx-version-min=10.10 -dynamiclib -Wl,-headerpadmaxinstallnames -compatibilityversion 1.0.0 -currentversion 1.0.0 -o -installname -Wl,-rpath,/usr/local/miniconda3/lib -Wl,-rpath,/usr/local/cuda/lib -Wl,-rpath,/Users/hongbinliu/pytorch/build/lib /usr/local/cuda/lib/libcudartstatic.a -Wl,-rpath,/usr/local/cuda/lib /usr/local/cuda/lib/libcudartstatic.a -Wl,-rpath,/usr/local/cuda/lib /usr/local/cuda/lib/stubs/cuda.framework /usr/local/cuda/lib/libcublasdevice.a ld: can't map file, errno=22 file '/usr/local/cuda/lib/stubs/cuda.framework' for architecture x8664 clang: error: linker command failed with exit code 1 (use -v to see invocation) make[2]: [] Error 1 make[1]: [] Error 2 make: * [all] Error 2\nThx, adding that did work. However, when I run training with cuda, it throws an RuntimeError: cannot initialize CUDA without ATen_cuda library. I guess this is another issue anyway.\nSame issue.", "<img width=\"545\" alt=\"downloads\" src=\"https://user-\"Also, I found a small problem when copying the dylibs to torch/lib, it is replacing the symbolic link with the actual file (screenshot above). Replacing cp -r to cp -R does the trick, please let me know if we want a pr.\nHave tracked down the problem was in , always return null because in Registry.h line 77 the (key) is always 0; ObjectPtrType Create(const SrcType&amp; key, Args... args) { ` Suspect CUDAHooksRegistry() is creating new instance, because of namespace? Is it working in Linux? Any help will be appreciated.\nI think Mac OS X with CUDA is broken at the moment. We think the problem has something to do with not getting a dependency on (or, I guess, , now)\nHi ezyang, thanks for your hint. I just tried before . It gives me \"Key already registered\". I will leave it to you guys, and I can help with testing, please @ me.\nThanks for giving it a try. Sounds like is being loaded correctly, so my initial guess was wrong. Would it be possible to print the keys of the map to see what is in it when you get the key already registered error message? And maybe also, you noted that there was never anything at the key \"CUDAHooks\" when you didn't have the dlopen line. Is the map always empty in that case, or is there something registered, perhaps at a different name? (I made this bug when I developed the patch and figured it out by seeing that there was another CUDAHooks thing but named slightly differently). I'd like to debug this myself but I can't get PyTorch to build with Mac OS X and CUDA at the moment. (See also the bugs I've recently filed)\nYes, the map was always empty when calling CUDAHooksRegistry(). Loading the dylib works as expected that it is reusing the same static Registry and registers the key. However calling CUDAHooksRegistry() doesn't get the same static Registry instance, instead it creates a new one. Could it be relating to how mac handles dynamic loading? e.g different memory space?\nSuccessfully reproed.", "And yes, it looks like the dynamic dependencies are setup correctly:\nBTW, fix for the cuda.framework error didn't work for me. Where exactly did you end up placing the block in the OTOH, this patch worked for me: (It's totally the wrong way to solve the problem but it might help someone)\nand I diagnosed the problem, fix coming.\nThat is great!\nGo try . You'll need to do a clean build from scratch. Also, you'll still need a cuda.framework fix of some sort or another. Working on a more robust version of this fix.\nThanks ezyang, it is working."], "neg": []}
{"query": "if (tensor->isCuda() != isCuda) { throw invalid_tensor(isCuda ? \"CUDA\" : \"CPU\", tensor->isCuda() ? \"CUDA\" : \"CPU\"); } <ins> if (tensor->type() != type) { throw invalid_tensor(thpp::toString(type), thpp::toString(tensor->type())); } </ins> } }", "pos": ["the error only happens in very specific conditions. (change the view of input tensor and use weight converted from numpy array.) The output of functional conv2d is far off in scale (e12 or so.)\nthis happens because generates double precision tensor by default, whereas torch generates single precision. We need to add guards for tensor types in THNN and THCUNN.\nIt's a problem with function, that only . Should be a simple fix.\nWhoops! Yeah it just ignores the \"thpp::Type type\" parameter\nwhat type is expected here?\na is a FloatTensor but weights is a DoubleTensor. Change: to: (or change a to (...).double())\ncool thanks!"], "neg": []}
{"query": "Py_INCREF(input_var); output_var = input_var; input_var_.creator = THPFunction_asFunction(self); <ins> input_var_.requires_grad = self->cdata.requires_grad; </ins> } else { // If the leaf Variable has been returned, we have to move it after the // current function to ensure the gradient is computed correctly.", "pos": ["import torch as th from torch.autograd import Variable x = Variable(th.Tensor([1 , 2 , 3]) , requiresgrad = False) y = Variable(th.Tensor([100]) , requiresgrad = True) x = x.cuda(0) y = y.cuda(0) x[0] = y print(x.requiresgrad) If ((0) (0)) is executed then the output is False, True otherwise. I think the result that x.requiresgrad = True is more natural.\nI further checked some in-place operator in GPU mode(say, x = Variable(...).cuda(0) ), and find out that after the operation is done, the requiresgrad of the Variable doesn't change. For example: import torch as th from torch.autograd import Variable x = Variable(torch.Tensor([1]) , requiresgrad = False) y = Variable(torch.Tensor([2]) , requiresgrad = True) x = x.cuda(0) y = y.cuda(0) x.add(y) print(x.requires_grad) the output is False\nIt doesn't have to do with Variable holding a CUDA tensor, but rather with the fact that it's no longer a leaf. I've sent a fix in .\nThanks for the report!"], "neg": []}
{"query": "return torch._C._cuda_getDeviceName(device) <ins> def get_device_capability(device): \"\"\"Gets the cuda capability of a device. Arguments: device (int): device for which to return the name. This function is a no-op if this argument is negative. Returns: tuple(int, int): the major and minor cuda capability of the device \"\"\" if device >= 0: return torch._C._cuda_getDeviceCapability(device) </ins> @contextlib.contextmanager def stream(stream): \"\"\"Context-manager that selects a given stream.", "pos": ["There's quite a few user-reports of installing the CUDA 7.5 binary of pytorch and then running on GTX 1080 / 1070, and there will be reports of installing 7.5 / 8.0 binaries and running on V100. Raising a user-warning in such situations and letting them know will fix the problem.\nThe version of CUDA used to compile the binary should be available thanks to this PR:\nI think is a more robust way of getting the CUDA used to compile PyTorch, as currently if there are multiple CUDA installed in the machine, the heuristics used in might not match those used by CMake FindCUDA. I think that we should modify the to return the cuda version used during compilation. I have the impression that it will not require changing much because doesn't require cuda to execute. What do you think?"], "neg": []}
{"query": "try: _loadlib() except Exception: <del> warnings.warn('cuDNN library not found. Check your LD_LIBRARY_PATH') </del> <ins> warnings.warn('cuDNN library not found. Check your {libpath}'.format( libpath={ 'darwin': 'DYLD_LIBRARY_PATH', 'win32': 'PATH' }.get(sys.platform, 'LD_LIBRARY_PATH'))) </ins> return False return True", "pos": ["The official website does not provide the torchvison version installation method. Here is the error message when torchvison 0.3.0 is used. torchvision/C.cpython-36m-x8664-linux- undefined symbol: ZN2at7getTypeERKNS6TensorE That should be caused by the mismatch version between pytorch-nightly and torchvision 0.3.0.\nIt should be that matches with , but they are not provided currently.\nI tried installing it with pip as per , but now I get when importing . I did a similar thing on June 25th. It worked then. Any known fix for this?\ndo we document what to do here?\nI finally solved this by building from source with Before running that make sure to upgrade and to use . Note: this will not work if you have on your system when building since you'd get version mismatch between and . Tested only on .\nyou build pytorch from source or just torchvision? or both of them.\nI installed from conda with: I then cloned the repo and ran inside the repo. I guess this would also work if you do , but I haven't tried it.\nJust a reminder (what I encountered just now): one should uninstall previous ` from source, or they can mix up and mess up everything.\nThere are nightlies for torchvision under Windows now.\nI have problems using Pytorch with my Geforce RTX 3060, (sm86). I tried to install with help of : Create a new venv, based on python 3.7.10. Activate it. conda install pytorch-nightly cudatoolkit=10.0 -c pytorch (I also did it without the \"=10.0\") After this, Pytorch detects the sm86 with no errors, problem sm86 solved with pytorch-nightly Now, I have a problem with torchvision. tried installing torchvision with File \"C:repovisiontorchvisionmodels\", line 195, in Inception3 AttributeError: module '' has no attribute 'unused' tried installing torchvision version 0.3.0, from , but an error rose: tried installing torchvision at origin/master, from , but another error rose: Here is the list of packages installed:"], "neg": []}
{"query": "transposedIndices.emplace_back(); } } <del> return std::make_tuple<>(self.permute(dims), std::move(transposedIndices)); </del> <ins> return std::make_tuple(self.permute(dims), std::move(transposedIndices)); </ins> } static std::vector<int64_t> computeLinearStride(const Tensor & tensor) {", "pos": ["In PyTorch master 0.4 , variables cannot be indexed. This used to work in PyTorch 0.2 :\nThis is breaking some of Pyro's inference algorithms on PyTorch master:"], "neg": []}
{"query": "finally: torch.backends.cudnn.enabled = prev <ins> def test_rnn_args_check(self): input_size = 3 hidden_size = 5 num_layers = 2 batch_size = 4 seq_len = 6 num_directions = 1 def test(input_shape, hidden_shape, mode): for input, hidden in get_inputs(input_shape, hidden_shape, mode): model = getattr(nn, mode)(input_size, hidden_size, num_layers) self.assertRaises(RuntimeError, lambda: model(input, hidden)) correct_input_shape = (seq_len, batch_size, input_size) correct_hidden_shape = (num_layers * num_directions, batch_size, hidden_size) def update_tuple(tup, dim, delta): new_tup = list(tup) new_tup[dim] = delta return tuple(new_tup) def get_inputs(input_shape, hidden_shape, mode): '''returns list( tuple(input, hidden) ) where input, hidden are inputs to a model''' input = Variable(torch.randn(input_shape)) hidden = Variable(torch.randn(hidden_shape)) if mode is not 'LSTM': return [(input, hidden)] if hidden_shape == correct_hidden_shape: return [(input, (hidden, hidden))] good_hidden = Variable(torch.randn(correct_hidden_shape)) return [ (input, (hidden, good_hidden)), (input, (good_hidden, hidden)), ] rnn_modes = ['RNN', 'GRU', 'LSTM'] for mode in rnn_modes: # Incorrect input batch size input_shape = update_tuple(correct_input_shape, 1, -1) hidden_shape = correct_hidden_shape test(input_shape, hidden_shape, mode) # Incorrect hidden batch size input_shape = correct_input_shape hidden_shape = update_tuple(correct_hidden_shape, 1, -1) test(input_shape, hidden_shape, mode) # Incorrect input size input_shape = update_tuple(correct_input_shape, 2, -1) hidden_shape = correct_hidden_shape test(input_shape, hidden_shape, mode) # Incorrect hidden size input_shape = correct_input_shape hidden_shape = update_tuple(correct_hidden_shape, 2, -1) test(input_shape, hidden_shape, mode) # Incorrect hidden[0] input_shape = correct_input_shape hidden_shape = update_tuple(correct_hidden_shape, 0, -1) test(input_shape, hidden_shape, mode) </ins> def test_rnn_initial_hidden_state(self): rnn_modes = ['RNN', 'GRU', 'LSTM'] for mode in rnn_modes:", "pos": ["I run RNN sucessfully with the code. and I run embedding example. Then, got RuntimeError. I set to use nothing seems wrong. And the error info seems not the dim problem? The error is so strange.\nWhat exactly do you mean by \"embedding example\"? Could you provide a code snippet that causes the RuntimeError?\nTo better reproduce the error, I use the official code example of The origin runs well: Due to use , I have to make batch first. And then , it raises error.\nThank you for the example! The say that setting means that the first dim of input is the batch, ie, input has size . however is unchanged, so the following should work: I do agree the error message could be better, though"], "neg": []}
{"query": "make_variable(std::move(std::get<3>(tensors)))); } <del> std::vector<Variable> VariableType::as_variable(TensorList tl) const { </del> <ins> static std::vector<Variable> as_variable(TensorList tl) { </ins> std::vector<Variable> variables; for (auto& t : tl) { variables.emplace_back(make_variable(std::move(t)));", "pos": ["See Docker run fails with , which is due to the fact that is specified in docker container: Which results in: CI cc\nCheck that adding to manywheels image results in a failure:\nTentatively marking hi-pri since nightly testing are failing\nThis is resolved, here is the successful run:"], "neg": []}
{"query": "<ins> #include <stdint.h> #include <stddef.h> enum THPByteOrder { THP_LITTLE_ENDIAN = 0, THP_BIG_ENDIAN = 1 }; THPByteOrder THP_nativeByteOrder(); void THP_decodeInt16Buffer(int16_t* dst, const uint8_t* src, THPByteOrder order, size_t len); void THP_decodeInt32Buffer(int32_t* dst, const uint8_t* src, THPByteOrder order, size_t len); void THP_decodeInt64Buffer(int64_t* dst, const uint8_t* src, THPByteOrder order, size_t len); void THP_decodeFloatBuffer(float* dst, const uint8_t* src, THPByteOrder order, size_t len); void THP_decodeDoubleBuffer(double* dst, const uint8_t* src, THPByteOrder order, size_t len); </ins>", "pos": ["<!-- A clear and concise description of what the bug is. --The target name is used by a test executable in mkldnn. Steps to reproduce the behavior: BUILDTEST=1 build <!-- If you have a code sample, error messages, stack traces, please provide it here as well --<!-- A clear and concise description of what you expected to happen. --CMake configure passes. Please copy and paste the output from our (or fill out the checklist below manually). You can get the script and run it with: PyTorch Version (e.g., 1.0): v1.5.0-rc2 OS (e.g., Linux): Windows How you installed PyTorch (, , source): source Build command you used (if compiling from source): python build Python version: 3.x CUDA/cuDNN version: 9.2 / 10.1 / 10.2 / NA GPU models and configuration: Tesla M60 Any other relevant information: <!-- Add any other context about the problem here. --We should not build tests for mkldnn.\nI have encountered the same error, how do I resolve it?\nIt is actually mkldnn that is out of sync. So running the following script should help.\nThanks a lot for this I posted this on the PyTorch Discussions but to no avail, and now I've finally understood the reason of this error. Thanks a lot. Surprisingly, for me the build works out if I build from the master repo, but changing to a different version always fails."], "neg": []}
{"query": "input=(torch.randn(2, 3, 5, 5, 5) * 1000)), OldModuleTest(nn.VolumetricMaxPooling, (2, 2, 2, 2, 2, 2), <del> input_size=(2, 3, 5, 5, 5), </del> <ins> input=(torch.randn(2, 3, 5, 5, 5) * 1000), </ins> desc='stride'), OldModuleTest(nn.VolumetricMaxPooling, (2, 2, 2, 2, 2, 2, 1, 1, 1), <del> input_size=(2, 3, 5, 5, 5), </del> <ins> input=(torch.randn(2, 3, 5, 5, 5) * 1000), </ins> desc='stride_padding'), OldModuleTest(nn.VolumetricReplicationPadding, (1, 2, 3, 4, 5, 6),", "pos": ["Repro: stack: says he's already investigating a similar problem in internal models. cc\nwe think may be root cause\nI think this might be correct? I am seeing CPU devices explicitly created in the model, and in dynamo. I do not believe this is a bug, but we need much better tooling around detecting devices.\nIn that case we need to add float16 support to the cpu backend (ugh)\nThe node being copied CPU=CUDA is: so the only CPU op involved is -- which we are actually generating ourselves. This patch might help, by doing the conversion on GPU: Though after applying it we get a new error:\nAfter applying some partial fixes in pytorch/torchdynamo, I tried the minifier -- but it didn't seem to work. cc I was running I get a lot of printouts of: But then it finishes I get:\nIf I change where the cpu device is set to \"cuda\" in the model source - and remove an assert on our end, I can get the same exact error. That might make for an easier minification and repro story. Let me see if that minimizes on my end.\nPlease rebase to this PR,\npytorch/torchdynamo gets the same error without needing to do dtype conversion, potentially easier for debugging?\nMinified version (requires pytorch/torchdynamo)\nref pytorch/torchdynamo"], "neg": []}
{"query": "THCAssertSameGPU(THCTensor_(checkGPU)(state, 2, self_, src_)); THArgCheck(src_->nDimension == 2, 1, \"expected a matrix\"); <del> THCTensor *src = src_; if (self_ == src_) src = THCTensor_(newContiguous)(state, src_); </del> <ins> if (self_ != src_) THCTensor_(resizeAs)(state, self_, src_); </ins> <del> int64_t stride0 = src->stride[0]; int64_t stride1 = src->stride[1]; real *start = THCTensor_(data)(state, src); </del> <ins> int64_t stride0 = self_->stride[0]; int64_t stride1 = self_->stride[1]; real *start = THCTensor_(data)(state, self_); </ins> TensorTriOp<real, 0> op(start, stride0, stride1, k); if (self_ == src_) { <del> if (!THC_pointwiseApply1(state, src, op)) { </del> <ins> if (!THC_pointwiseApply1(state, src_, op)) { </ins> THArgCheck(false, 2, CUTORCH_DIM_WARNING); } } else { <del> THCTensor_(resizeAs)(state, self_, src); </del> <ins> THCTensor_(resizeAs)(state, self_, src_); </ins> <del> if (!THC_pointwiseApply2(state, self_, src, op)) { </del> <ins> if (!THC_pointwiseApply2(state, self_, src_, op)) { </ins> THArgCheck(false, 2, CUTORCH_DIM_WARNING); } } <del> if (self_ == src_) THCTensor_(freeCopyTo)(state, src, src_); </del> THCudaCheck(cudaGetLastError()); }", "pos": ["On recent PyTorch master (built from source; 0.4.0a0+) on a P100, CUDA 9:\nI think the results on CPU meets the defination of triu (upper triangular part of the matrix). It is really hard to handle with a tensor from expand operation with multi-thread not only on GPU but also on CPU. The unexpected result will also happen on CPU if we apply openmp to it. I think it is necessary to eleminate the difference between the implementation and defination.\nIt doesn\u2019t matter if it\u2019s cpu or not. triu doesn\u2019t have an underscore which means it doesn\u2019t operate in-place and should always give the same result (in new memory)"], "neg": []}
{"query": "line_search_fn = group['line_search_fn'] history_size = group['history_size'] <del> state = self.state['global_state'] </del> <ins> # NOTE: LBFGS has only global state, but we register it as state for # the first param, because this helps with casting in load_state_dict state = self.state[self._params[0]] </ins> state.setdefault('func_evals', 0) state.setdefault('n_iter', 0)", "pos": ["Hi, I encountered this bug: The code skeleton is like: It seems the loaded are , and I've tried a workaround to move to , but it still has the same bug.\nCould you provide a full script to reproduce the problem?\nmaybe you can try like this, () (beta1).add_(1 - beta1, ())\nSorry, I missed the reply email. I am afraid that I am unable to provide a reproducer now. It is a work I am doing for the OpenNMT-py , trying to use for doing lr update. And I encoutered this problem when testing the case. So I factored out the code skeleton about this problem above. I've tried several methods, including tricks like what suggests, but it still happens. Per my analysis, it is because the previous training was done on gpu, so when saving the , the stored states(tensors) are of version. During resuming, when we load the saved optimizer, loads this version to cpu(the model() can be moved to gpu easily, but seems lacking this ability?) , so this problem emerges.\nTry moving optimizer state to the GPU memory manually after loading it from the checkpoint. I agree that having an method for this operation would be nice.\nthanks. My work was suspended due to other problems, when resumed, I will try your method.\nI'm afraid solution won't work. It will make the error go away, but your optimizer will no longer be training the model. You should recreate optimizers after casting modules to a different type or device, and you can use to restore the state from a previous copy. This currently doesn't work, but we should fix it (by copying from the data from the state dict, instead of using the tensors directly - this allows for cross-device or cross-type updates).\n, yep, your method is what I currently use, it works. But I will wait for upstream to fix this problem though. Thanks for your great works!\nAh, my bad. I forgot to update the line where the optimizer is recreated. But otherwise, the following should do the job, right?\nah, right. That should work\nExcept that you should use instead of\nI had a similar problem. When I save the optimizer state from a GPU other than GPU 0 and then load the state it still loads everything to GPU 0.", "Specifying in didn't work either. 's solution solves this though.\nHi guys, I have a very similar problem as the one in this thread, here's my code: And then once I resume, I got KeyErrors on my optimizer: Do you guys know how to fix this issue? BTW, I have 8 GPUs used, I'm guessing if this issue was because of that?\nwere you able to solve your issue with multiple-GPUs?\nThanks for asking. Yeah, I'm able to, turns out the issue is because I used an early version of PyTorch, after I updated the version, it works like a charm!\nJust a comment, this problem is caused by makes all state tensor are moved into GPU0 so by moving the state of an optimizer to specific GPU will fix this problem.\nHi how do you know the specific GPU location of different tensors in advance?\nWould a feature where all () calls always makes use of an automatically generated CPU version be feasible ? And then at resume the () would make use of the \"current\" device being used (or any better strategy). At the moment it seems we need a lot of boilerplater code to ensure saving and loading are consistents across devices for models/optimizer/scheduler/etc.\nI have met similar problem, I recreated Adam optimizer without () after reloading model, () and DataParallel(model) according to solution.\nthanks, it work!\nHi, as you say that: every time moving model to other device, we should build optimizer, but, if we move the model to other device and move back, should we build the optimizer again? here is an example code: does optimizer run as expected? also, if doing , should we rebuild optimizer ?\nIs there a problem if you switch the order to something like this? Meaning moving the model to 'cuda' but only loading it's state dict from checkpoint after loading the optimizer's state dict first?\nThe problem can be concluded that the optimizer's state will be loaded to the device as same as the model. You must load the model to GPU at first, and then load the optimizer's state. So that both the model and the optimizer's state are loaded in GPU.", "Instead of moving optimizer to cuda after loading it in cpu, you could load the checkpoint directly in cuda:\nI've independently rediscovered that this works :) Should read until the end of the thread next time\nI find my codes still have the problem. I tried my best to range the modules as the examples shown in the above. can anyone give me some hints? I also tried this\nFor me, doesn't solve the issue, but the one by works. Any idea why? I create the optimizer, load the state (with map location to CUDA), pass to the train loop, where the mode is pushed to device (though I assume it's not needed if the loaded model is pushed to cuda by maplocation). Specifically, I would get stuck at call, throwing the device mismatch error.\nStill it doesn't work for me. Could you please help me?"], "neg": []}
{"query": "inputs, module_kwargs = scatter_kwargs(inputs, module_kwargs, device_ids, dim) if len(device_ids) == 1: return module(*inputs[0], **module_kwargs[0]) <del> replicas = replicate(module, device_ids[:len(inputs)]) outputs = parallel_apply(replicas, inputs, module_kwargs, device_ids) </del> <ins> used_device_ids = device_ids[:len(inputs)] replicas = replicate(module, used_device_ids) outputs = parallel_apply(replicas, inputs, module_kwargs, used_device_ids) </ins> return gather(outputs, output_device, dim)", "pos": ["has broken DataParallel tests. Now that deviceids are explicitly sent to parallelapply, this assert gets triggered if inputs are not big enough to be on all devices (e.g. batch size of 20 on 8 GPUs gets chunked into 6*3+2, so that 8-th GPU is idle, and assert gets triggered).\nyour latest fix in doesn't fix this completely. Here's a failing stack-trace:\nHow is it possible that the tests didn't catch that in the PR?\nit's because our build box which only has 2 GPUs doesn't hit this case. It's failing on 4-GPU setting (maybe batch size becomes too small then)"], "neg": []}
{"query": "grad = grad.index(c); } <del> int64_t num_features = grad.size(-1); </del> <ins> int64_t num_features = grad_.size(-1); </ins> auto weight_size = std::array<int64_t, 2>{{ num_weights, num_features }}; <ins> auto& dense_type = grad.type(); auto& sparse_type = dense_type.toBackend(grad.is_cuda() ? kSparseCUDA : kSparseCPU); // check if all our grad come from padding_idx if (grad.numel() == 0) { return sparse_type.sparse_coo_tensor(indices_.type().tensor(), dense_type.tensor(), weight_size); } </ins> auto index = indices.view({1, -1}); auto values = grad.contiguous().view({-1, num_features}); <del> auto& sparse_type = grad.type().toBackend(grad.is_cuda() ? kSparseCUDA : kSparseCPU); </del> return sparse_type.sparse_coo_tensor(index, values, weight_size); }", "pos": ["I have a straightforward use case. Given a token (or word), lookup the embedding of the word, pass that through a linear layer and perform L2 regression. If the token corresponds to the of the embedding layer, throws an error saying OS: macOS High Sierra PyTorch version: 0.3.1 How you installed PyTorch (conda, pip, source): Tried both conda and pip Python version: 3.6.4 CUDA/cuDNN version: CPU GPU models and configuration: NA GCC version (if compiling from source): NA A script to reproduce the bug:\nI got the same trouble, and it throw this error after some epochs of training. Does anyone have some solutions?"], "neg": []}
{"query": "THCTensor *newValues = THCTensor_(newNarrow)(state, sorted, dimension, k, 1); THCudaLongTensor *newIndices = THCudaLongTensor_newNarrow(state, sorted_indices, dimension, k, 1); <ins> THCTensor_(free)(state, sorted); THCudaLongTensor_free(state, sorted_indices); </ins> if (!keepdim) { THCTensor_(squeeze1d)(state, newValues, newValues, dimension); THCudaLongTensor_squeeze1d(state, newIndices, newIndices, dimension);", "pos": ["has a memory leak on the GPU when using PyTorch version 0.3.1. The memory leak goes away if I downgrade to PyTorch 0.3.0. I was able to reproduce the memory leak on 3 separate machines using P100s, 1080TIs and Titan Xs.How you installed PyTorch (conda, pip, source): conda OS: Ubuntu 16.04 PyTorch version: 0.3.1.post3 Python version: 3.6.3 CUDA version: 8.0.61 GPU models and configuration: 4x P100s, 2x 1080 TIs, 4x Titan X (3 separate machines)\nthanks for the report. I can reproduce the bug on master. We'll get it fixed."], "neg": []}
{"query": "Best practices -------------- <ins> Device-agnostic code ^^^^^^^^^^^^^^^^^^^^ Due to the structure of PyTorch, you may need to explicitly write device-agnostic (CPU or GPU) code; an example may be creating a new tensor as the initial hidden state of a recurrent neural network. The first step is to determine whether the GPU should be used or not. A common pattern is to use Python's `argparse` module to read in user arguments, and have a flag that can be used to disable CUDA, in combination with `torch.cuda.is_available()`. In the following, `args.cuda` results in a flag that can be used to cast tensors and modules to CUDA if desired:: import argparse import torch parser = argparse.ArgumentParser(description='PyTorch Example') parser.add_argument('--disable-cuda', action='store_true', help='Disable CUDA') args = parser.parse_args() args.cuda = not args.disable_cuda and torch.cuda.is_available() If modules or tensors need to be sent to the GPU, `args.cuda` can be used as follows:: x = torch.Tensor(8, 42) net = Network() if args.cuda: x = x.cuda() net.cuda() When creating tensors, an alternative to the if statement is to have a default datatype defined, and cast all tensors using that. An example when using a dataloader would be as follows:: dtype = torch.cuda.FloatTensor for i, x in enumerate(train_loader): x = Variable(x.type(dtype)) When working with multiple GPUs on a system, you can use the `CUDA_VISIBLE_DEVICES` environment flag to manage which GPUs are available to PyTorch. To manually control which GPU a tensor is created on, the best practice is to use the `torch.cuda.device()` context manager:: print(\"Outside device is 0\")  # On device 0 (default in most scenarios) with torch.cuda.device(1): print(\"Inside device is 1\")  # On device 1 print(\"Outside device is still 0\")  # On device 0 If you have a tensor and would like to create a new tensor of the same type on the same device, then you can use the `.new()` function, which acts the same as a normal tensor constructor. Whilst the previously mentioned methods depend on the current GPU context, `new()` preserves the device of the original tensor. This is the recommended practice when creating modules in which new tensors/variables need to be created internally during the forward pass:: x_cpu = torch.FloatTensor(1) x_gpu = torch.cuda.FloatTensor(1) x_cpu_long = torch.LongTensor(1) y_cpu = x_cpu.new(8, 10, 10).fill_(0.3) y_gpu = x_gpu.new(x_gpu.size()).fill_(-5) y_cpu_long = x_cpu_long.new([[1, 2, 3]]) If you want to create a tensor of the same type and size of another tensor, and fill it with either ones or zeros, `torch.ones_like()` or `torch.zeros_like()` are provided as more convenient functions (which also preserve device):: x_cpu = torch.FloatTensor(1) x_gpu = torch.cuda.FloatTensor(1) y_cpu = torch.ones_like(x_cpu) y_gpu = torch.zeros_like(x_gpu) </ins> Use pinned memory buffers ^^^^^^^^^^^^^^^^^^^^^^^^^", "pos": ["A common issue - for both new and more seasoned users, including myself - is how to write nice device-agnostic code. There seems to be several potential solutions (using a global CUDA flag, always calling , using etc.), used by various people in various bits of code. Ideally, we should have \"best practices\" documented somewhere, like ; or at least some place where we can link to.\nAlternative proposal - we provide a method that deals with the most common use cases (tensors, variables and networks). We could have something similar to Chainer's , something like a , etc. We're not going to be \"write once, run everywhere\", but we can probably catch most use cases. Pros: It makes it much simpler for people writing CPU/1 GPU code (presumably a majority of users). Depending on the approach it'll probably be fine for multiple GPUs too. Cons: An abstraction of this form isn't foolproof, but I can't immediately think of any major issues. Setting everything to CUDA by default is probably going to break things, but a wrapper, being more manual, can be used selectively enough.\nI agree we should figure out a way to make writing device-agnostic code easier. I don't think fakecuda is a good idea, it's too magical and has global side effects. Here's a list of options I see, along with some comments. In general I'm not 100% happy with any of those, but I don't have any better ideas right now. Example: I think that sooner or later we should add arg to functions anyway, so it seems like a good enough extension. I think I'm fine with this option too, but I'm afraid you would either need a single global module captured inside functions, or use it everywhere inside. In general it can be simulated using the approach using partials (of course this is an inefficient implementation, but you get the idea): This is fine, except that you have to call it a lot and it makes the code super verbose. A similar alternative that people sometimes use now is this: This is really the best option, as it will give you a tensor of exactly the same type and on exactly the same device as the source. This is usually the desired behaviour, which is why we use it all over the place in the core library.", "On the other hand, it feels a bit unnatural, and is annoying when you do more complicated things, because you first construct the tensor and only then fill it with the data. and helps a lot with that\nThanks for more in-depth feedback on some options. For existing practices, highlighting , and are probably the way to go. In terms of what to do when there aren't existing tensors, what I see as the main target is the . Precisely, can we do and without using the if statement? The cast/dtype solution seems good for tensors, but anything we can do for modules?\nI've put together some of this and other stuff I know into \"best practices\" documentation in I think we still want to think about future possibilities, but there's plenty that's worth noting down now for users.\nIt would indeed be amazing if PyTorch could make it simpler to write device agnostic code. I really like your first suggestion. Just a quick note from my (applied) perspective: maybe it would be possible to split up the and the into two attributes? First of all, I think this would make it easier to handle cases where tensors are on different gpus and second of all, it would allow to write code like this: ((5, 5), , ) # Returns on same device as x Don't know how practical this is though.\none thing I like about suggestion is that it makes it explicit and obvious when something is happening on cuda, rather than requiring users to know that non-negative device numbers refer to cuda. Maybe this is a bad idea, but what about also taking an optional kw in every function that takes a but to have both and non- dtypes? So passing would give you a cuda float tensor on the current GPU, passing would give you a CPU float, and passing would give you a cuda tensor on GPU 5. Obviously we'd need error checking to check things like non- devices don't have non-negative s specified (which is kind of silly because if the device is specified it tells you everything), but it does seem to make the common case of not specifying a device more clear.\nSpeaking only as an end-use with no particular insight into how difficult it would be to code: These suggestions make me feel like end users will end up having to think more about device semantics, not less.", "I'd like to see the API move away from having the tensor type imply the storage location entirely. Have a context manager that sets cpu vs. cuda, default device, etc. Something like: Then once this exists, class wraps its calls to the with an appropriate context, as do Modules. I think that this results in everything Just Working(tm).\nI'm not convinced that making it easy for end users to avoid thinking about device semantics is the correct goal. The different devices have different performance characteristics, amount of memory/cache, incur data transfer costs/synchronizations, etc. My worry is that by papering over these issues, we make it really easy to write non-performant code and really difficult to figure out why. This discussion is similar to a discussion in the distributed systems community: is a good discussion of why trying to unify all interfaces might not be the best idea (there are admittedly additional complexities in distributed computing like partial failures). Anyway, that's why I like making it obvious when there are cuda storage / device transfers going on, because those are obvious places to look at performance.\nI'm not an expert on any of this, but about the only two things that bug me about using is when I have to cast are I don't have a prototypical tensor at hand (e.g. in a module it feels unnatural to use so it might be neat to have a \"canonical tensor\" in modules that reacts to .cuda() and whose is exposed as or so. I want a different type is just a bit much. Other than that, I think using it is cool, I think that there is not much advantage in over But this is only from my limited experience.\nRegarding the \"canonical tensor\" it's hard, because the module can have a lot of different parameters that have different types and are on different devices. I agree that chains are awful. I think (and similar for all other types) would be a convenient helper.\nOh yes, something like would be cool (or ). For the \"canonical tensor\", I imagine that, for basic applications, having something that \"does the right thing\" whether called before or after would be great. For more advanced applications, the user could create their own set of prototype tensors and just do , or so. But that is just pure imagination, as I don't have multiple GPUs.", ":slightlysmilingface:\nthere are several potential goals here, and they are somewhat conflicting. One is to make CPU/GPU code seamless for end users, with just one flag somewhere. On the other hand, that's likely to results in less-performant code and harder debugging. It seems like the conversation is going towards device-aware code, but making it a little less clunky in places. Sounds like is a good step forward, and was probably in the pipeline anyway?\na latecomer here: i agree that it's not a good idea to hide things behind the curtain from users. though, in many cases, it might be easier to to have a global behavior of ignoring, e.g., gpu (i.e., make .cuda() do nothing). could we at least give some kind of global switch allowing this behiavour?\nClosing as 0.4 has plenty of options for this"], "neg": []}
{"query": "os.mkdir(os.path.join(tmp_dir, 'barrier')) os.mkdir(os.path.join(tmp_dir, 'test_dir')) if backend == 'mpi': <del> mpiexec = 'mpiexec -n 3 --noprefix {}'.format(python) </del> <ins> # test mpiexec for --noprefix option devnull = open(os.devnull, 'w') noprefix_opt = '--noprefix' if subprocess.call( 'mpiexec -n 1 --noprefix bash -c \"\"', shell=True, stdout=devnull, stderr=subprocess.STDOUT) == 0 else '' mpiexec = 'mpiexec -n 3 {} {}'.format(noprefix_opt, python) </ins> return_code = run_test(mpiexec, test_module, test_directory, options) else:", "pos": ["I installed to get available, but the test run failed this way: Log: So it sounds like we need a specific MPI backend for the tests, any old one won't do. CC and\nHmm can you try removing ?\nThank you, It's work!\nI am sorry and , how to remove --noprefix,,? should I reinstall the mpi using --prefix,,? thank you\nIt needs to be removed from the script\nThank you so much adam, I have solved it Sent from Mail for Windows 10"], "neg": []}
{"query": "#include \"function.h\" <ins> #include <string> </ins> #include <THPP/THPP.h> #include \"variable.h\"", "pos": ["Version include <stringin some cpp files: define order in The pre-defined struct is not passed by clang in macOS, so we should reorder them.\nSimple fix"], "neg": []}
{"query": "('cross', (S, 3), ((S, 3),)), ('cross', (S, 3, S), ((S, 3, S), 1), 'dim'), ('inverse', (S, S), (), '', (), [skipIfNoLapack]), <ins> ('gesv', (S, S), ((S, S),), '', (), [skipIfNoLapack]), </ins> ('clone', (S, M, S), ()), ('eq', (S, S, S), ((S, S, S),)), ('ne', (S, S, S), ((S, S, S),)),", "pos": ["There seems to be regression in the dynamo export of the MONAI VNet model as regards . Export script: The versions I'm using is torch 2.5.0.dev20240625+cu121 Previous torch versions such as were able to export this model successfully. Perhaps dynamo was generating FX graph without the expm1 node previously? cc: PyTorch version: 2.5.0.dev20240625+cu121 Is debug build: False CUDA used to build PyTorch: 12.1 ROCM used to build PyTorch: N/A OS: Ubuntu 22.04.4 LTS (x8664) GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 Clang version: Could not collect CMake version: version 3.29.0 Libc version: glibc-2.35 Python version: 3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0] (64-bit runtime) Python platform: Linux-6.5.0-35-generic-x8664-with-glibc2.35 Is CUDA available: True CUDA runtime version: 12.4.131 CUDAMODULELOADING set to: LAZY GPU models and configuration: GPU 0: NVIDIA RTX A5000 Nvidia driver version: 535.171.04 cuDNN version: Probably one of the following: /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- HIP runtime version: N/A MIOpen runtime version: N/A Is XNNPACK available: True CPU: Architecture: x8664 CPU op-mode(s): 32-bit, 64-bit Address sizes: 46 bits physical, 48 bits virtual Byte Order: Little Endian CPU(s): 20 On-line CPU(s) list: 0-19 Vendor ID: GenuineIntel Model name: Intel(R) Core(TM) i9-9820X CPU @ 3.", "30GHz CPU family: 6 Model: 85 Thread(s) per core: 2 Core(s) per socket: 10 Socket(s): 1 Stepping: 4 CPU max MHz: 4200.0000 CPU min MHz: 1200.0000 BogoMIPS: 6599.98 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constanttsc art archperfmon pebs bts repgood nopl xtopology nonstoptsc cpuid aperfmperf pni pclmulqdq dtes64 monitor dscpl vmx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse41 sse42 x2apic movbe popcnt tscdeadlinetimer aes xsave avx f16c rdrand lahflm abm 3dnowprefetch cpuidfault epb catl3 cdpl3 invpcidsingle pti ssbd mba ibrs ibpb stibp tprshadow flexpriority ept vpid eptad fsgsbase tscadjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm mpx rdta avx512f avx512dq rdseed adx smap clflushopt clwb intelpt avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqmllc cqmoccupllc cqmmbmtotal cqmmbmlocal dtherm ida arat pln pts hwp hwpactwindow hwpepp hwppkgreq vnmi mdclear flushl1d archcapabilities Virtualization: VT-x L1d cache: 320 KiB (10 instances) L1i cache: 320 KiB (10 instances) L2 cache: 10 MiB (10 instances) L3 cache: 16.5 MiB (1 instance) NUMA node(s): 1 NUMA node0 CPU(s): 0-19 Vulnerability Gather data sampling: Mitigation;", "Microcode Vulnerability Itlb multihit: KVM: Mitigation: VMX disabled Vulnerability L1tf: Mitigation; PTE Inversion; VMX conditional cache flushes, SMT vulnerable Vulnerability Mds: Mitigation; Clear CPU buffers; SMT vulnerable Vulnerability Meltdown: Mitigation; PTI Vulnerability Mmio stale data: Mitigation; Clear CPU buffers; SMT vulnerable Vulnerability Retbleed: Mitigation; IBRS Vulnerability Spec rstack overflow: Not affected Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and _user pointer sanitization Vulnerability Spectre v2: Mitigation; IBRS; IBPB conditional; STIBP conditional; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Mitigation; Clear CPU buffers; SMT vulnerable Versions of relevant libraries: [pip3] numpy==1.24.4 [pip3] onnx==1.16.0 [pip3] onnxscript==0.1.0.dev20240625 [pip3] optree==0.11.0 [pip3] pytorch-quantization==2.1.2 [pip3] pytorch-triton==3.0.0+ [pip3] torch==2.5.0.dev20240625+cu121 [pip3] torch-tensorrt==2.3.0a0 [pip3] torchaudio==2.4.0.dev20240625+cu121 [pip3] torchdata==0.7.1a0 [pip3] torchtext==0.17.0a0 [pip3] torchvision==0.20.0.dev20240625+cu121 [conda] Could not collect\nI am not very sure, but this looks like the kind of ops that should get decomp'ed. do you know? could you help take a look as part of the op support work you are doing? Thanks!\nYes, it looks like there is a graph-changed in dynamo, which dynamo now decomposes elu for some reasons. Using can preserve elu and then successfully export though.", "Yes can run but we would need to be resolved first since the weights are still not exported properly.\nI can take a look at this.\nshould fix this.", "Here, I can't export trivial net with sort() using dynamoexport(). Legacy () works fine, dynamoexport errors out : Unsupported FX nodes: {'callfunction': ['']}. Pytorch nightly 06/03 cc\na missing op, could you take this?\nNot sure if related:\nHmmm. Ideally, ONNX should add a sort op. It doesn't have one. I wonder what the legacy exporter converts it to.\nshould fix this issue.", "Are there any plans to implement something like tensorflow's and ? exists but it doesn't have the backward step implemented.\nOne workaround for the moment would be to write as , which should support backprop since . Adding support for should be trivial, the gradient for it can be found in\nI'll add backward support for , which should be equivalent to . I'll try to do it later today.\nHow difficult would it be to add in batched versions (where only the final two dimensions are square) of GESV and inverse? Are there any plans for this?\nIf we have batched version of , then adapting the autograd part is very simple. There is an open issue already that discuss implementing batched gesv\nActually, there is an open PR for that, so it might be coming soon\nI'll rebase the batched version of gesv. It'll come in soon\nMay I ask why gesv only support . As I saw from , it support that A is a MxN matrix, and B is a NxK matrix, then solve a matrix X (NxK). If I want to use a non-square matrix A in gesv, would you mind to let me know how to implement in PyTorch?\nthis doesn't exactly answer your question but you can add zero-rows to the A matrix until it becomes square.\nthanks a lot, I find the solution, can solve the problem.\nAny plan to add gradient calculation for gels?\nI am also wondering if will have differentiation support. My current workaround is to use instead\nI actually implemented overdetermined case following tensorflow implementation. For backward pass I only did basic grad test. On Monday, March 18, 2019, Johnson Zhong <:\nThanks, seems to work fine!"], "neg": []}
{"query": "lr=1e-1) ) <ins> def test_adagrad_sparse(self): self._test_rosenbrock_sparse( lambda params: optim.Adagrad(params, lr=1e-1) ) </ins> def test_adamax(self): self._test_rosenbrock( lambda params: optim.Adamax(params, lr=1e-1),", "pos": ["and I talked about the intended semantics of sparsemask when given a sparse tensor as an argument, and we came to the following conclusions: with an uncoalesced sparse tensor does not make sense, full stop. Unlike values, where your update might be linear and work correctly on an uncoalesced sparse tensor, there isn't any good reason why you want to pluck out values from the dense tensor multiple times, it is basically always a bug. the (eventual) documentation for sparsemask, we will recommend calling coalesce before doing any operations involving sparsemask (or nonlinear updates to indices/values). should add a test to sparsemask which checks if the flag of the input tensor is true, and raise an error if it is not. This is a cheap safety check which will prevent people from forgetting to call coalesce (as in (2)). There is a performance consequence to calling coalesce every iteration of training. If the gradient is producing coalesced vectors, we can alleviate this problem by adding a preprocessing step to coalesce() which tests if the input is already coalesced, and bugs out in that case. However, I think this is out of scope for this particular ticket for now."], "neg": []}
{"query": "set /a NO_NNPACK=1 ) <ins> IF \"%DEBUG%\"==\"1\" ( set BUILD_TYPE=Debug ) ELSE ( set BUILD_TYPE=Release ) </ins> IF \"%CMAKE_GENERATOR%\"==\"\" ( set CMAKE_GENERATOR_COMMAND= set MAKE_COMMAND=msbuild INSTALL.vcxproj /p:Configuration=Release", "pos": ["The current master triggers the following error messages from Visual Studio 2015 (Debug build): 2------ Build started: Project: ATen, Configuration: Debug x64 ------ 2cl : Command line error D8016: '/Ox' and '/RTC1' command-line options are incompatible ... 14LINK : fatal error LNK1104: cannot open file '....srcATenDebug' It appears that they are triggered by *.c files in the ATen project, like \"srcTHvectorAVX.c\"\nWhy do you need Debug build?\nIsn't ATen meant to support both Debug and Release builds? FYI I'm involved with the project at which does support both build configurations, and we would like to be able to link our project with ATen.\nWell, you can remove the flags in as a workground.\nWe should patch this file to ignore this flag when building the library in debug mode. It shouldn't be too hard\nI catched similar errors , when I Tryed to build Debug Version I try to build Release and catched next errors e:pytorchatensrcatencpuvec256vec256int.h(184): note: see reference to function template instantiation 'at::vec256::anonymous-namespace'::Vec256<int16t::blend<1(at::vec256::anonymous-namespace'::Vec256<int16t)' being compiled 28e:pytorchatensrcatencpuvec256vec256int.h(147): error C3861: 'mm256extract_epi16': identifier not found What I should do?\nThere're two issues in your description. The first question is easy. You'll also have to specify the flag to 1. The second question is about your compiler. I guess you are using VS2015 or earlier versions of VS2017. Make sure the version in the output is 19.11 after your enter in CMD.\nYes I am using VS2015. My Compiler is 19.0.2. Must I use VS 2017 in any case?\nSorry, but the answer is yes. The AVX2 functions was in VS2017 and they are used in PyTorch. So VS2017 19.11 is a must.\nNo problem:) Thank you."], "neg": []}
{"query": "cudnnDestroyConvolutionDescriptor(desc); } void set(cudnnDataType_t dataType, int dim, int* pad, int* stride, int * upscale) { <ins> cudnnDataType_t mathType = dataType; if (dataType == CUDNN_DATA_HALF) mathType = CUDNN_DATA_FLOAT; </ins> CHECK(cudnnSetConvolutionNdDescriptor(desc, dim, pad, stride, upscale, <del> CUDNN_CROSS_CORRELATION, dataType)); </del> <ins> CUDNN_CROSS_CORRELATION, mathType)); </ins> } };", "pos": ["to reproduce Will result in:\nWhat's your GPU and CUDA version?\nTitanX Maxwell, cuda 8.0\nIn cudnn bindings convolution math type is inferred from input data type (and inferred to be equal to input data type): Math type for convolutions with half inputs and weights should be CUDNNDATAFLOAT because 1) computations with half math are not supported anywhere except Pascal and TX1 2) Even on Pascal, 6.1 cards have very low performance fp16 math 3) fp16 math comes with a set of accuracy problems\nThanks. Seems like batchnorm fp16 support is still missing. e.g: will result in\nAh, yes. fp16 support was removed from cudnn bn backend (because it still requires fp32 for parameters), and probably it is not properly routed to or not supported in nn backend. BTW, be careful with truncating bn params to 16 bits, in our experiments on wide resnets/cifar it cost ~ half a percent of accuracy, though there may be ways to overcome this.\nIs there an update regarding this? Could you describe how to truncate the bn params?\ndid you update PyTorch to a recent version? The snippet no longer raises an error for me\nThanks for the fast response. I installed the newest version via conda. Just retried the above code snippet, it works indeed. Shouldn't it be possible to \"half\" a entire model like this: For this i get: CUDNNSTATUSBAD_PARAM\nYour input size is wrong, models in torchvision expect specific input sizes. You can play around with the batch size, but number of channels and image size are fixed. When I change input size to (64,3,224,224) the snippet above does not return an error."], "neg": []}
{"query": "fn.w_desc = rnn.init_weight_descriptor(fn, fn.weight_buf) # Slice off views into weight_buf <del> params = rnn.get_parameters(fn, handle, fn.weight_buf) </del> all_weights = [[p.data for p in l] for l in self.all_weights] <ins> params = rnn.get_parameters(fn, handle, fn.weight_buf) </ins> # Copy weights and update their storage rnn._copyParams(all_weights, params)", "pos": ["While the fast path for cuDNN RNNs introduced in includes a check intended to fall back to copying weights if weights are moved or aliased, this only works if the aliasing happens after the weights are first compacted. But the most natural place to tie weights is inside a module's method, which runs before , the typical time when is first invoked. Here's an MWE. The three runs (CPU, weight tying before , and weight tying after ) should all give the same results, and they do before , but they have not since:"], "neg": []}
{"query": "return !was_null && requires_grad == var.requires_grad() && is_volatile == var.is_volatile(); } <del> Node* recordTraceHelper(std::string op, at::ArrayRef<Variable> inputs, at::ArrayRef<Variable> outputs); // These overloads are intended to simplify code generation inline Node* recordTrace(std::string op, std::initializer_list<Variable> inputs, const Variable& output) { return recordTraceHelper(op, inputs, {output}); } inline Node* recordTrace(std::string op, std::initializer_list<Variable> inputs, const std::tuple<Variable, Variable>& outputs) { return recordTraceHelper(op, inputs, {std::get<0>(outputs), std::get<1>(outputs)}); } inline Node* recordTrace(std::string op, std::initializer_list<Variable> inputs, const std::tuple<Variable, Variable, Variable>& outputs) { return recordTraceHelper(op, inputs, {std::get<0>(outputs), std::get<1>(outputs), std::get<2>(outputs)}); } inline Node* recordTrace(std::string op, at::TensorList inputs, const Variable& output) { // TODO: Eliminate the intermediate vector allocation return recordTraceHelper(op, variable_list(inputs.begin(), inputs.end()), {output}); } </del> <ins> Node* recordTrace(std::string op, at::ArrayRef<Variable> inputs, at::ArrayRef<Variable> outputs); </ins> }}} // namespace torch::jit::tracer", "pos": ["when executing:\nI'm seeing this as well as of last night. Its preceded by many warnings about : But the first error is the one about . I went back, and I get a good install as late as commit"], "neg": []}
{"query": "('topk', small_3d_unique, lambda t: [2, 1, True, True], 'dim_desc_sort'), ('trace', medium_2d, lambda t: [],), ('tril', medium_2d, lambda t: [],), <ins> ('tril', medium_2d_expanded, lambda t: [], 'zero_stride', types, True), </ins> ('tril', medium_2d, lambda t: [2], 'positive'), ('tril', medium_2d, lambda t: [-2], 'negative'), ('triu', medium_2d, lambda t: [],), <ins> ('triu', medium_2d_expanded, lambda t: [], 'zero_stride', types, True), </ins> ('triu', medium_2d, lambda t: [2], 'positive'), ('triu', medium_2d, lambda t: [-2], 'negative'), ('unsqueeze', new_t(2, 3, 4), lambda t: [2],),", "pos": ["On recent PyTorch master (built from source; 0.4.0a0+) on a P100, CUDA 9:\nI think the results on CPU meets the defination of triu (upper triangular part of the matrix). It is really hard to handle with a tensor from expand operation with multi-thread not only on GPU but also on CPU. The unexpected result will also happen on CPU if we apply openmp to it. I think it is necessary to eleminate the difference between the implementation and defination.\nIt doesn\u2019t matter if it\u2019s cpu or not. triu doesn\u2019t have an underscore which means it doesn\u2019t operate in-place and should always give the same result (in new memory)"], "neg": []}
{"query": "END_HANDLE_TH_ERRORS } <ins> #ifndef THC_GENERIC_FILE static PyObject * THPStorage_(fromBuffer)(PyObject *_unused, PyObject *args, PyObject *keywds) { HANDLE_TH_ERRORS PyObject *obj = NULL; const char* byte_order_str = NULL; Py_ssize_t count = -1, offset = 0; Py_buffer buffer; static char *kwlist[] = {\"buffer\", \"byte_order\", \"count\", \"offset\", NULL}; const char* argtypes; #if defined(TH_REAL_IS_BYTE) || defined(TH_REAL_IS_CHAR) argtypes = \"O|snn\"; #else argtypes = \"Os|nn\"; #endif if (!PyArg_ParseTupleAndKeywords(args, keywds, argtypes, kwlist, &obj, &byte_order_str, &count, &offset)) { return NULL; } #if !(defined(TH_REAL_IS_BYTE) || defined(TH_REAL_IS_CHAR)) THPByteOrder byte_order; if (strcmp(byte_order_str, \"native\") == 0) { byte_order = THP_nativeByteOrder(); } else if (strcmp(byte_order_str, \"big\") == 0) { byte_order = THP_BIG_ENDIAN; } else if (strcmp(byte_order_str, \"little\") == 0) { byte_order = THP_LITTLE_ENDIAN; } else { PyErr_Format(PyExc_ValueError, \"invalid byte_order '%s' (expected 'big', 'little', or 'native')\", byte_order_str); return NULL; } #endif if (PyObject_GetBuffer(obj, &buffer, PyBUF_SIMPLE) < 0) return NULL; if (offset < 0 || offset > buffer.len) { PyErr_Format(PyExc_ValueError, \"offset must be non-negative and no greater than buffer length (%ld)\", (long) buffer.len); return NULL; } if (count < 0) { if ((buffer.len - offset) % sizeof(real) != 0) { PyErr_Format(PyExc_ValueError, \"buffer size must be a multiple of element size\"); return NULL; } count = (buffer.len - offset) / sizeof(real); } if (offset + (count * (Py_ssize_t)sizeof(real)) > buffer.len) { PyErr_Format(PyExc_ValueError, \"buffer is smaller than requested size\"); return NULL; } uint8_t* src = (uint8_t*) buffer.buf; THStorage* storage = THStorage_(newWithSize)(count); #if defined(TH_REAL_IS_BYTE) || defined(TH_REAL_IS_CHAR) memcpy(storage->data, src + offset, count); #elif defined(TH_REAL_IS_SHORT) THP_decodeInt16Buffer(storage->data, src + offset, byte_order, count); #elif defined(TH_REAL_IS_INT) THP_decodeInt32Buffer(storage->data, src + offset, byte_order, count); #elif defined(TH_REAL_IS_LONG) THP_decodeInt64Buffer(storage->data, src + offset, byte_order, count); #elif defined(TH_REAL_IS_FLOAT) THP_decodeFloatBuffer(storage->data, src + offset, byte_order, count); #elif defined(TH_REAL_IS_DOUBLE) THP_decodeDoubleBuffer(storage->data, src + offset, byte_order, count); #else #error \"Unknown type\" #endif return (PyObject*)THPStorage_(newObject)(storage); END_HANDLE_TH_ERRORS } #endif </ins> PyObject * THPStorage_(writeFile)(THPStorage *self, PyObject *file) { HANDLE_TH_ERRORS", "pos": ["<!-- A clear and concise description of what the bug is. --The target name is used by a test executable in mkldnn. Steps to reproduce the behavior: BUILDTEST=1 build <!-- If you have a code sample, error messages, stack traces, please provide it here as well --<!-- A clear and concise description of what you expected to happen. --CMake configure passes. Please copy and paste the output from our (or fill out the checklist below manually). You can get the script and run it with: PyTorch Version (e.g., 1.0): v1.5.0-rc2 OS (e.g., Linux): Windows How you installed PyTorch (, , source): source Build command you used (if compiling from source): python build Python version: 3.x CUDA/cuDNN version: 9.2 / 10.1 / 10.2 / NA GPU models and configuration: Tesla M60 Any other relevant information: <!-- Add any other context about the problem here. --We should not build tests for mkldnn.\nI have encountered the same error, how do I resolve it?\nIt is actually mkldnn that is out of sync. So running the following script should help.\nThanks a lot for this I posted this on the PyTorch Discussions but to no avail, and now I've finally understood the reason of this error. Thanks a lot. Surprisingly, for me the build works out if I build from the master repo, but changing to a different version always fails."], "neg": []}
{"query": "The transposed convolution operator multiplies each input value element-wise by a learnable kernel, and sums over the outputs from all input feature planes. <del> **This module can be seen as the exact reverse of Conv3d**. It is sometimes (but incorrectly) refered to as a deconvolutional operation. </del> <ins> This module can be seen as the gradient of Conv3d with respect to its input. It is also known as a fractionally-strided convolution or a deconvolution (although it is not an actual deconvolution operation). </ins> | :attr:`stride` controls the stride for the cross-correlation. | If :attr:`padding` is non-zero, then the input is implicitly zero-padded on both sides <del> for :attr:`padding` number of points </del> <ins> for :attr:`padding` number of points. </ins> | If :attr:`output_padding` is non-zero, then the output is implicitly zero-padded on one side <del> for :attr:`output_padding` number of points | :attr:`groups` controls the connections between inputs and outputs. </del> <ins> for :attr:`output_padding` number of points. | :attr:`dilation` controls the spacing between the kernel points; also known as the \u00e0 trous algorithm. It is harder to describe, but this `link`_ has a nice visualization of what :attr:`dilation` does. | :attr:`groups` controls the connections between inputs and outputs. `in_channels` and `out_channels` must both be divisible by `groups`. </ins> |       At groups=1, all inputs are convolved to all outputs. |       At groups=2, the operation becomes equivalent to having two conv layers side by side, each seeing half the input channels, and producing half the output channels, and both subsequently concatenated. <ins> At groups=`in_channels`, each input channel is convolved with its own set of filters (of size `out_channels // in_channels`). </ins> The parameters :attr:`kernel_size`, :attr:`stride`, :attr:`padding`, :attr:`output_padding` can either be: <del> - a single ``int`` -- in which case the same value is used for the height and width dimension </del> <ins> - a single ``int`` -- in which case the same value is used for the depth, height and width dimensions </ins> - a ``tuple`` of three ints -- in which case, the first `int` is used for the depth dimension, the second `int` for the width dimension and the third `int` for the width dimension", "pos": ["Hi, I have installed two versions of CUDA. But I want to use one of it, can I specify it during setup process?\nTry running\nI have tried it, () still returns False. Thx for your reply!\nDo you have the cuda driver installed?\nOf course I do. It raise assertion error which shows \"The NVIDIA driver on your system is too old(found version 8000)\".\nthat's the answer to your problem. upgrade your driver.\nThx!", "When building from source, I got this error. My system is OSX 10.12.4, conda, CUDA 8.0. Any thoughts? Thanks!\ncan you try adding to the file and see if that fixes things? if so i'll commit it.\nactually, that file already has memory included. Are you following instructions from Especially: This part is important:\nworks nicely! Thanks a lot!", "The suggest that takes a as a parameter, but it actually must be wrapped in a Variable first.\nyou are looking at the Tensor docs. When operating on Variable, it is implied from the tutorials that all functions operating on Variables have to use Variables. We didn't write separate docs for Variable itself.", "something wrong with import torchvision Traceback (most recent call last): File \"\", line 3, in <moduleimport torchvision File \"\", line 4, in <moduleimport torchvision.transforms as transforms ImportError: No module named transforms\nyou have a file called for some reason. delete it or change your pythonpath to ignore it.\nHey, I have the same problem persisting when I import torchvision. I have no other file by the same name. ImportError Traceback (most recent call last) <ipython-input-6-3baf6837d8f1in <module() 8 from import Dataset 9 from import DataLoader ---10 from torchvision import transforms ImportError: No module named torchvision Any idea why this is happening?\nDo you have torchvision installed (it is a separate package)?\npip Install torchvision try this\nfrom data import BaseTransform,VOC_CLASSES as labelmap this is not working\npip install torchvision", "Hi, I need to pass input through one nn.Module, then argmax the output, and then pass that output to a second nn.Module. I want to backprop through the argmax back to the weights of the first module. That's impossible right now, though, it seems. When I loop through the gradient of the weights of my first module (with the following code): firstmodel.zerograd() secondmodel.zerograd() loss.backward() # retainvariables=True for param in secondmodel.parameters(): print( 'data: ', ) print( 'grad data: ', ) for param in firstmodel.parameters(): print( 'data: ', ) print( 'grad data: ', ) I get the following error, stating that those gradients can no longer be viewed/exist: -main()\nargmax is not differentiable (and even if you try to think about it as a function differentiable almost everywhere, its derivative is always 0, so you won't get anything meaningful).", "def inithidden(self): h0 = Variable(FloatTensor(self.numlayers, self.batchlen, self.hiddendim).zero(),  requiresgrad=True) c0 = Variable(FloatTensor(self.numlayers, self.batchlen, self.hiddendim).zero(),  requiresgrad=True) return (h0, c0) def forward(self, sequences): lstmout, = ((1, self.batchlen, self.inputdim), ) ... = nn.LSTM(inputdim, hiddendim, numlayers, dropout = dropout, batchfirst=True) = self.inithidden() output = model(traininput) produces the following error although I defined weight and input dimensions as in the documentation: File \"/usr/local/lib/python2.7/dist-\", line 202, in call result = self.forward(input, kwargs) File \"/usr/local/lib/python2.7/dist-\", line 91, in forward output, hidden = func(input, self.allweights, hx) File \"/usr/local/lib/python2.7/dist-\", line 327, in forward return func(input, fargs, fkwargs) File \"/usr/local/lib/python2.7/dist-\", line 227, in forward nexth, output = func(input, hidden, weight) File \"/usr/local/lib/python2.7/dist-\", line 67, in forward hy, output = inner(input, hidden[l], weight[l]) File \"/usr/local/lib/python2.7/dist-\", line 96, in forward hidden = inner(input[i], hidden, *weight) File \"/usr/local/lib/python2.7/dist-\", line 22, in LSTMCell gates = F.linear(input, wih, bih) + F.linear(hx, whh, bhh) File \"/usr/local/lib/python2.7/dist-\", line 744, in add return (other) File \"/usr/local/lib/python2.7/dist-\", line 293, in add return (other, False) File \"/usr/local/lib/python2.7/dist-\", line 287, in add return Add(inplace)(self, other) File \"/usr/local/lib/python2.7/dist-\", line 20, in forward return a.add(b) RuntimeError: inconsistent tensor size at /data/users/soumith/builder/wheel/pytorch-src/torch/lib/TH/generic/THTensorMath.c:827\nI can't reproduce your problem. Could you include a short runnable example that exhibits this behavior?", "You also don't need to manually init hidden anymore, the following would work as well...\nThanks! After removing problem solved.", "Hi, I would like to implement a sharing weight 'alexnet' with two input batches of images, and the fully connected layers are concatenated together like this: However, when I run it, I come across the following error: Does anyone know about it? Thank you very much.\nYou forgot to call on the output of . In the future please remember that we're using the issues for bug reports and feature requests only. If you have questions or need help, please post ."], "neg": []}
{"query": "flat_input = tuple(_iter_variables(input)) flat_output = super(NestedIOFunction, self)._do_forward(*flat_input) nested_output = self._nested_output <del> nested_variables = _map_tensor_fromiter(iter(flat_output))(self._nested_output) </del> <ins> nested_variables = _unflatten(flat_output, self._nested_output) </ins> return nested_variables def backward(self, *gradients): <del> nested_gradients = _map_tensor_fromiter(iter(gradients))(self._nested_output) </del> <ins> nested_gradients = _unflatten(gradients, self._nested_output) </ins> del self._nested_output result = self.backward_extended(*nested_gradients) del self._to_save_nested", "pos": ["cuDNN backend for RNN stabilizes at large mem usage. The following code stabilizes for me at 14+GiB reported by nvidia-smi. The same RNN without cuDNN uses <2.5GiB.\nOops... any idea what's going on with the garbage collector in this case?\nare you running this on a P100 GPU?\nYes, I observe similar behavior on Titan X Pascal\nhmmm, do cudnn RNNs have weird memory-oversubscribing properties? If it wasn't for CuDNN RNN, then it's not unreasonable. we use a custom CUDA allocator that will oversubscribe if needed, but will free the memory blocks it's holding onto if it hits OOM / failed allocations.\nTried the same RNN in torch and am getting ~2.6GiB on nvidia-smi.\nWith cuDNN I'm getting a constant usage of 2241MB on my machine using PyTorch (M40).\nare you using cudnnv5, or v6?\nv5\nno, cudnn does not do memory oversubscribing, and already tried the same network in lua torch with cudnn, where it gets expected memory usage. Also, he is getting OOMs on slightly bigger networks, where apparently caching allocator should have tried to free previous blocks. Interesting that can not repro this.\nJust to make sure did you set when you ran? I left it as False in the repro.\nI can repro in the docker image built from the current version on P100, and I go OOM after 8 batches with the same image on 1080, which generally should not happen - if there is enough memory to go through 1 batch, memory usage should not increase afterwards.\nNo, I just copied your script and ran it. Weird thing is that it goes OOM after 18 batches once I enable . Without it memory remains stable.\nA lot of cudaMalloc allocations go not through THCudaMalloc, and never get freed:\nSo not going through THCudaMalloc is actually expected for storages allocated with THCStorage_(newWithSize), but what's not expected is to every time cudaMalloc bytes of rnn workspace, instead of reusing earlier allocations. (~784 MB allocations are for workspace, ~156MB allocations happen somewhere in criterion).\nMy guess is that the bug is at the Python level.", "Either: Something is holding on to a Tensor longer than it should Something is creating a circular reference which prevents quick clean-up\nThe issue is that in creates cyclic references. It's made worse by the fact that it holds on to more memory than it should (like the workspace). You can see this by adding a as the first line in your for loop. With the explicit call to the GC every iteration, the memory usage doesn't explode.\nhow can these cyclic references be fixed?\nNot sure yet, but I will fix it tomorrow with"], "neg": []}
{"query": "strt = _tensor_str(self) size_str = 'x'.join(str(size) for size in self.size()) <del> device_str = '' if not self.is_cuda else ' (GPU {})'.format(self.get_device()) </del> <ins> device_str = '' if not self.is_cuda else  ' (GPU {})'.format(self.get_device()) </ins> strt += '[{} of size {}{}]n'.format(torch.typename(self), <del> size_str, device_str) </del> <ins> size_str, device_str) </ins> return 'n' + strt", "pos": ["It seems that the following link that can be reached from here (check the screenshot) is broken: Page: Broken link: Screenshot: ! cc\nthis should be transferred to , but I don't have permissions to do that. Fix is available at pytorch/hub."], "neg": []}
{"query": "<ins> #include \"byte_order.h\" static inline uint16_t decodeUInt16LE(const uint8_t *data) { return (data[0]<<0) | (data[1]<<8); } static inline uint16_t decodeUInt16BE(const uint8_t *data) { return (data[1]<<0) | (data[0]<<8); } static inline uint32_t decodeUInt32LE(const uint8_t *data) { return (data[0]<<0) | (data[1]<<8) | (data[2]<<16) | (data[3]<<24); } static inline uint32_t decodeUInt32BE(const uint8_t *data) { return (data[3]<<0) | (data[2]<<8) | (data[1]<<16) | (data[0]<<24); } static inline uint64_t decodeUInt64LE(const uint8_t *data) { return (((uint64_t)data[0])<< 0) | (((uint64_t)data[1])<< 8) | (((uint64_t)data[2])<<16) | (((uint64_t)data[3])<<24) | (((uint64_t)data[4])<<32) | (((uint64_t)data[5])<<40) | (((uint64_t)data[6])<<48) | (((uint64_t)data[7])<<56); } static inline uint64_t decodeUInt64BE(const uint8_t *data) { return (((uint64_t)data[7])<< 0) | (((uint64_t)data[7])<< 8) | (((uint64_t)data[5])<<16) | (((uint64_t)data[4])<<24) | (((uint64_t)data[3])<<32) | (((uint64_t)data[2])<<40) | (((uint64_t)data[1])<<48) | (((uint64_t)data[0])<<56); } THPByteOrder THP_nativeByteOrder() { uint32_t x = 1; return *(uint8_t*)&x ? THP_LITTLE_ENDIAN : THP_BIG_ENDIAN; } void THP_decodeInt16Buffer(int16_t* dst, const uint8_t* src, THPByteOrder order, size_t len) { for (size_t i = 0; i < len; i++) { dst[i] = (int16_t) (order == THP_BIG_ENDIAN ? decodeUInt16BE(src) : decodeUInt16LE(src)); src += sizeof(int16_t); } } void THP_decodeInt32Buffer(int32_t* dst, const uint8_t* src, THPByteOrder order, size_t len) { for (size_t i = 0; i < len; i++) { dst[i] = (int32_t) (order == THP_BIG_ENDIAN ? decodeUInt32BE(src) : decodeUInt32LE(src)); src += sizeof(int32_t); } } void THP_decodeInt64Buffer(int64_t* dst, const uint8_t* src, THPByteOrder order, size_t len) { for (size_t i = 0; i < len; i++) { dst[i] = (int64_t) (order == THP_BIG_ENDIAN ? decodeUInt64BE(src) : decodeUInt64LE(src)); src += sizeof(int64_t); } } void THP_decodeFloatBuffer(float* dst, const uint8_t* src, THPByteOrder order, size_t len) { for (size_t i = 0; i < len; i++) { union { uint32_t x; float f; }; x = (order == THP_BIG_ENDIAN ? decodeUInt32BE(src) : decodeUInt32LE(src)); dst[i] = f; src += sizeof(float); } } void THP_decodeDoubleBuffer(double* dst, const uint8_t* src, THPByteOrder order, size_t len) { for (size_t i = 0; i < len; i++) { union { uint64_t x; double d; }; x = (order == THP_BIG_ENDIAN ? decodeUInt64BE(src) : decodeUInt64LE(src)); dst[i] = d; src += sizeof(double); } } </ins>", "pos": ["<!-- A clear and concise description of what the bug is. --The target name is used by a test executable in mkldnn. Steps to reproduce the behavior: BUILDTEST=1 build <!-- If you have a code sample, error messages, stack traces, please provide it here as well --<!-- A clear and concise description of what you expected to happen. --CMake configure passes. Please copy and paste the output from our (or fill out the checklist below manually). You can get the script and run it with: PyTorch Version (e.g., 1.0): v1.5.0-rc2 OS (e.g., Linux): Windows How you installed PyTorch (, , source): source Build command you used (if compiling from source): python build Python version: 3.x CUDA/cuDNN version: 9.2 / 10.1 / 10.2 / NA GPU models and configuration: Tesla M60 Any other relevant information: <!-- Add any other context about the problem here. --We should not build tests for mkldnn.\nI have encountered the same error, how do I resolve it?\nIt is actually mkldnn that is out of sync. So running the following script should help.\nThanks a lot for this I posted this on the PyTorch Discussions but to no avail, and now I've finally understood the reason of this error. Thanks a lot. Surprisingly, for me the build works out if I build from the master repo, but changing to a different version always fails."], "neg": []}
{"query": "optimizer (Optimizer): Wrapped optimizer. milestones (list): List of epoch indices. Must be increasing. gamma (float): Multiplicative factor of learning rate decay. <del> Default: -0.1. </del> <ins> Default: 0.1. </ins> last_epoch (int): The index of last epoch. Default: -1. Example:", "pos": ["as shown in source code, this parameters is not used. Any reason that we keep it?\nthis can be cleaned up."], "neg": []}
{"query": "\"\"\"Fills the input Tensor or Variable with values according to the method described in \"Understanding the difficulty of training deep feedforward neural networks\" - Glorot, X. & Bengio, Y. (2010), using a uniform distribution. The resulting tensor will have values sampled from :math:`U(-a, a)` where <del> :math:`a = gain times sqrt{2 / (fan_in + fan_out)} times sqrt{3}`. Also known as Glorot initialisation. </del> <ins> :math:`a = gain times sqrt{2 / (fan_in + fan_out)} times sqrt{3}`. Also known as Glorot initialisation. </ins> Args: tensor: an n-dimensional torch.Tensor or autograd.Variable", "pos": ["why 1.21.6? because that's the oldest supported numpy on windows/python3.10: all cc\nHere is the test for pinning numpy to 1.21.6 The reason its was not pinned to this version is that conda can't properly resolve all the dependencies at this step here:\nIt looks like you are correct, I will put up a PR to pin lower then already pinned:\nHere is the PR:"], "neg": []}
{"query": "AT_PRIVATE_CASE_TYPE(at::ScalarType::Long, int64_t, __VA_ARGS__)         AT_PRIVATE_CASE_TYPE(at::ScalarType::Short, int16_t, __VA_ARGS__)        default:                                                                 <del> AT_ERROR(\"%s not implemented for '%s'\", (NAME), the_type.toString());  </del> <ins> AT_ERROR(#NAME, \" not implemented for '\", the_type.toString(), \"'\");  </ins> }                                                                          }()", "pos": ["is not available in the docs page, but the doc string has been in . This was mentioned earlier in", "Oh, bilinear, I missed that last time I looked at it. I'll add a docstring.", "returns nothing. Might be useful to have some linter in the build process to find Python functions missing", "Pytorch 0.4.0 gives a segmentation fault with dilated 2D convolutions over a specific size. import torch import as nn model = nn.Conv2d(inchannels=64, outchannels=64, kernelsize=5, dilation=2) model(([1, 64, 1166, 1166])) # This runs model(([1, 64, 1167, 1167])) # This crashes PyTorch version: 0.4.0 Is debug build: No CUDA used to build PyTorch: 9.1.85 OS: CentOS Linux 7 (Core) GCC version: (GCC) 4.8.5 (Red Hat 4.8.5-28) CMake version: version 2.8.12.2 Python version: 3.6 Is CUDA available: Yes CUDA runtime version: 8.0.61 GPU models and configuration: GPU 0: GeForce GTX 1080 Ti GPU 1: TITAN X (Pascal) GPU 2: Quadro K2200 Nvidia driver version: 390.59 cuDNN version: Probably one of the following: /usr/local/cuda- /usr/local/cuda- /usr/local/cuda-7.5/lib64/libcudnnstatic.a /usr/local/cuda- /usr/local/cuda-8.0/lib64/libcudnn_static.a\ndo you think this is the same as the segfault you showed me a week ago on convolution?\nah yes. int overflow. I only fixed conv3d because I didn\u2019t want to hurt perf. I\u2019ll fix all of them.", "This should be an easy one.. The function is implemented, just no sign of it in the docs."], "neg": []}
{"query": "char *bytes = (char *) data; int64_t remaining = sizeof(real) * size; while (remaining > 0) { <del> ssize_t result = write(fd, bytes, remaining); </del> <ins> // we write and read in 1GB blocks to avoid bugs on some OSes ssize_t result = write(fd, bytes, THMin(remaining, 1073741824)); </ins> if (result < 0) throw std::system_error(result, std::system_category()); bytes += result;", "pos": ["Running this on OSX: More details of failure in Ubuntu 14.04.1 in\nMaybe it's caused by a ? I just test this code and it doesn't work (on Mac OS X 10.12.3 with Python 3.6):\nThat thread links to this one , which says it\u2019s an OS bug and proposes a workaround patch that hasn\u2019t been merged. There\u2019s also a numpy issue thread, which links to this Torch7 commit that implements a workaround\nWe're doing a loop in case it writes less, but we always pass the full amount of remaining bytes. Putting a cap (std::min with INT_MAX) on the size of a single write might help.\nI can't reproduce it on CentOS. I think it might have been a report for a PyTorch version prior to fixing ."], "neg": []}
{"query": "c[1].fill_(20) self.assertEqual(c[1], c[3], 0) <ins> def test_from_buffer(self): a = bytearray([1, 2, 3, 4]) self.assertEqual(torch.ByteStorage.from_buffer(a).tolist(), [1, 2, 3, 4]) shorts = torch.ShortStorage.from_buffer(a, 'big') self.assertEqual(shorts.size(), 2) self.assertEqual(shorts.tolist(), [258, 772]) ints = torch.IntStorage.from_buffer(a, 'little') self.assertEqual(ints.size(), 1) self.assertEqual(ints[0], 67305985) f = bytearray([0x40, 0x10, 0x00, 0x00]) floats = torch.FloatStorage.from_buffer(f, 'big') self.assertEqual(floats.size(), 1) self.assertEqual(floats[0], 2.25) </ins> if __name__ == '__main__': unittest.main()", "pos": ["<!-- A clear and concise description of what the bug is. --The target name is used by a test executable in mkldnn. Steps to reproduce the behavior: BUILDTEST=1 build <!-- If you have a code sample, error messages, stack traces, please provide it here as well --<!-- A clear and concise description of what you expected to happen. --CMake configure passes. Please copy and paste the output from our (or fill out the checklist below manually). You can get the script and run it with: PyTorch Version (e.g., 1.0): v1.5.0-rc2 OS (e.g., Linux): Windows How you installed PyTorch (, , source): source Build command you used (if compiling from source): python build Python version: 3.x CUDA/cuDNN version: 9.2 / 10.1 / 10.2 / NA GPU models and configuration: Tesla M60 Any other relevant information: <!-- Add any other context about the problem here. --We should not build tests for mkldnn.\nI have encountered the same error, how do I resolve it?\nIt is actually mkldnn that is out of sync. So running the following script should help.\nThanks a lot for this I posted this on the PyTorch Discussions but to no avail, and now I've finally understood the reason of this error. Thanks a lot. Surprisingly, for me the build works out if I build from the master repo, but changing to a different version always fails."], "neg": []}
{"query": "# use faster C loader if available from yaml import CLoader as YamlLoader except ImportError: <del> from yaml import YamlLoader </del> <ins> from yaml import Loader as YamlLoader </ins> GENERATED_COMMENT = CodeTemplate(\"\"\"", "pos": ["Build error: According to , the should be modified as:\nI the same problem when installing pytorch from source. You mean the file is wrong?\nAlthough the PyYAML documentation site suggests that is possible, I could not perform it. That is a reason for the failing. This has been reported as well: I believe that the should have this instead: In fact when I did this: This indicates that the import didn't happen in PyYAML itself.\nyou guys are right, the case has to be modified to: . The original issue (my guess) is that your pyyaml had some compile error and the CLoader wasn't successfully installed."], "neg": []}
{"query": "TEST_CUDA = torch.cuda.is_available() TEST_MULTIGPU = TEST_CUDA and torch.cuda.device_count() >= 2 TEST_CUDNN = TEST_CUDA and torch.backends.cudnn.is_acceptable(torch.cuda.FloatTensor(1)) <ins> TEST_CUDNN_VERSION = TEST_CUDNN and torch.backends.cudnn.version() </ins> PRECISION = 1e-5 module_tests = [", "pos": ["cuDNN backend for RNN stabilizes at large mem usage. The following code stabilizes for me at 14+GiB reported by nvidia-smi. The same RNN without cuDNN uses <2.5GiB.\nOops... any idea what's going on with the garbage collector in this case?\nare you running this on a P100 GPU?\nYes, I observe similar behavior on Titan X Pascal\nhmmm, do cudnn RNNs have weird memory-oversubscribing properties? If it wasn't for CuDNN RNN, then it's not unreasonable. we use a custom CUDA allocator that will oversubscribe if needed, but will free the memory blocks it's holding onto if it hits OOM / failed allocations.\nTried the same RNN in torch and am getting ~2.6GiB on nvidia-smi.\nWith cuDNN I'm getting a constant usage of 2241MB on my machine using PyTorch (M40).\nare you using cudnnv5, or v6?\nv5\nno, cudnn does not do memory oversubscribing, and already tried the same network in lua torch with cudnn, where it gets expected memory usage. Also, he is getting OOMs on slightly bigger networks, where apparently caching allocator should have tried to free previous blocks. Interesting that can not repro this.\nJust to make sure did you set when you ran? I left it as False in the repro.\nI can repro in the docker image built from the current version on P100, and I go OOM after 8 batches with the same image on 1080, which generally should not happen - if there is enough memory to go through 1 batch, memory usage should not increase afterwards.\nNo, I just copied your script and ran it. Weird thing is that it goes OOM after 18 batches once I enable . Without it memory remains stable.\nA lot of cudaMalloc allocations go not through THCudaMalloc, and never get freed:\nSo not going through THCudaMalloc is actually expected for storages allocated with THCStorage_(newWithSize), but what's not expected is to every time cudaMalloc bytes of rnn workspace, instead of reusing earlier allocations. (~784 MB allocations are for workspace, ~156MB allocations happen somewhere in criterion).\nMy guess is that the bug is at the Python level.", "Either: Something is holding on to a Tensor longer than it should Something is creating a circular reference which prevents quick clean-up\nThe issue is that in creates cyclic references. It's made worse by the fact that it holds on to more memory than it should (like the workspace). You can see this by adding a as the first line in your for loop. With the explicit call to the GC every iteration, the memory usage doesn't explode.\nhow can these cyclic references be fixed?\nNot sure yet, but I will fix it tomorrow with"], "neg": []}
{"query": "getattr(self, self.id().split(\".\")[2])() sys.exit(0) <del> def _join_and_reduce(self): </del> <ins> def _join_and_reduce(self, fn): skip_ok = getattr(fn, \"skip_if_no_cuda_distributed\", False) </ins> for p in self.processes: p.join(self.JOIN_TIMEOUT) <del> self.assertEqual(p.exitcode, 0) </del> <ins> if not skip_ok: self.assertEqual(p.exitcode, 0) if skip_ok: first_process = self.processes[0] # do this first so we don't give an error message about mismatched exit codes if the first isn't valid assert first_process.exitcode == 0 or first_process.exitcode == SKIP_IF_NO_CUDA_EXIT_CODE for p in self.processes: self.assertEqual(p.exitcode, first_process.exitcode) if first_process.exitcode == SKIP_IF_NO_CUDA_EXIT_CODE: raise unittest.SkipTest(\"cuda is not available\") </ins> elif BACKEND == 'mpi': dist.init_process_group(init_method=INIT_METHOD, backend='mpi')", "pos": ["is causing a crash when we try to build pytorch. Crash looks like this: (all times in US/West) Our 7AM, 8AM builds and earlier were all fine but 9AM, 10AM, 11AM builds all crashed with that same traceback. The new import was in which I think went in between 8AM and 9AM today See command below: See here if needed for : <!-- A clear and concise description of what you expected to happen. --Please copy and paste the output from our (or fill out the checklist below manually). You can get the script and run it with: PyTorch Version (e.g., 1.0): building from master OS (e.g., Linux): linux How you installed PyTorch (, , source): source Build command you used (if compiling from source): see above under 'reproduce' Python version: 3.6\nmaybe easier to read for build command used:\nplease try in the CircleCi environment?\nin our build script seems to have fixed this. Thanks !"], "neg": []}
{"query": "<del> d6015d42d9a1834bc7595c4bd6852562fb80b30b </del> <ins> 23512dbebd44a11eb84afbf53c3c071dd105297e </ins>", "pos": ["*Status could be: ongoing Torch Bench testing is broken with errors that look like Inductor CI has been broken since 5:41 PM PT on June 14, 2024. It was detected the next day and root caused the morning on June 17th. The affected inductor ci jobs (those related to torchbench perf) shall be marked as unstable until this issue is mitigated. We suspect it's due to the numpy 2.0 update. () Ongoing Current plan is for to pin numpy for torchbench workflows TBD\nThis can be fixed by the following PRs: PyTorch PR: TorchBench PR:\nI'll close those unstable issue once trunk recovers"], "neg": []}
{"query": "if IS_CONDA: lib_paths.append(os.path.join(CONDA_DIR, 'lib')) include_paths.append(os.path.join(CONDA_DIR, 'include')) <ins> for path in include_paths: if path is None or not os.path.exists(path): continue include_file_path = os.path.join(path, 'cudnn.h') if os.path.exists(include_file_path): CUDNN_INCLUDE_DIR = path CUDNN_INCLUDE_VERSION = -1 with open(include_file_path) as f: for line in f: if \"#define CUDNN_MAJOR\" in line: CUDNN_INCLUDE_VERSION = int(line.split()[-1]) break if CUDNN_INCLUDE_VERSION == -1: raise AssertionError(\"Could not find #define CUDNN_MAJOR in \" + include_file_path) break </ins> for path in lib_paths: if path is None or not os.path.exists(path): continue", "pos": ["Traceback (most recent call last): File \"\", line 3, in <moduleimport model File \"/home/abc/github/chinese-\", line 5, in <modulefrom import crnnOcr File \"/home/abc/github/chinese-\", line 72, in <modulemodel,converter = crnnSource() File \"/home/abc/github/chinese-\", line 64, in crnnSource model = crnn.CRNN(32, 1, len(alphabet)+1, 256, 1).cuda() File \"/usr/local/lib/python2.7/dist-\", line 216, in cuda return (lambda t: t.cuda(device)) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 123, in apply self.flattenparameters() File \"/usr/local/lib/python2.7/dist-\", line 65, in flattenparameters if not or not (anyparam): File \"/usr/local/lib/python2.7/dist-\", line 82, in isacceptable if libcudnn() is None: File \"/usr/local/lib/python2.7/dist-\", line 58, in libcudnn 'but linked against {}'.format(compileversion, _cudnnversion)) RuntimeError: cuDNN version mismatch: PyTorch was compiled against 6021 but linked against 5110 How to solve the problem?\ncould you pull the latest master branch and recompile. I pushed a change yesterday that fixes this:\nI remove the old code and pull new code, then compile, install.It still fails. /usr/local/lib/python2.7/dist- UserWarning: No training configuration found in save file: the model was not compiled. Compile it manually.", "('No training configuration found in save file: ' Traceback (most recent call last): File \"\", line 3, in <moduleimport model File \"/home/qianbb/github/chinese-\", line 5, in <modulefrom import crnnOcr File \"/home/qianbb/github/chinese-\", line 72, in <modulemodel,converter = crnnSource() File \"/home/qianbb/github/chinese-\", line 64, in crnnSource model = crnn.CRNN(32, 1, len(alphabet)+1, 256, 1).cuda() File \"/usr/local/lib/python2.7/dist-\", line 216, in cuda return (lambda t: t.cuda(device)) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 100, in apply self.flattenparameters() File \"/usr/local/lib/python2.7/dist-\", line 66, in flattenparameters if not or not (anyparam): File \"/usr/local/lib/python2.7/dist-\", line 87, in isacceptable if libcudnn() is None: File \"/usr/local/lib/python2.7/dist-\", line 58, in libcudnn 'but linked against {}'.format(compileversion, _cudnnversion)) RuntimeError: cuDNN version mismatch: PyTorch was compiled against 6021 but linked against 5110 The library of cuda is shown as following screenshot. !\nI found the issue, I fixed it in which we will merge shortly.\nThanksDoes it have been merged?\nyes. it is merged.\nIt seems the fix doesn't work.", "The log is shown as following Traceback (most recent call last): File \"\", line 3, in <moduleimport model File \"/home/qianbb/github/chinese-\", line 5, in <modulefrom import crnnOcr File \"/home/qianbb/github/chinese-\", line 72, in <modulemodel,converter = crnnSource() File \"/home/qianbb/github/chinese-\", line 64, in crnnSource model = crnn.CRNN(32, 1, len(alphabet)+1, 256, 1).cuda() File \"/usr/local/lib/python2.7/dist-\", line 222, in cuda return (lambda t: t.cuda(device)) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 100, in apply self.flattenparameters() File \"/usr/local/lib/python2.7/dist-\", line 66, in flattenparameters if not or not (anyparam): File \"/usr/local/lib/python2.7/dist-\", line 87, in isacceptable if libcudnn() is None: File \"/usr/local/lib/python2.7/dist-\", line 58, in libcudnn 'but linked against {}'.format(compileversion, _cudnnversion)) RuntimeError: cuDNN version mismatch: PyTorch was compiled against 6021 but linked against 5110\nthat is weird. Can I see the full log of in pytorch directory. Can you make sure you have latest commits with .\nI am sure the code is up-to-date. First I uninstall the code using the command pip unisntall torch. Then I remove the old code using the command rm -r pytorch and git clone again. The log of install is\nSorry to trouble you. Is there any new information on this issue?\nI spent some time resolving this issue and I don't see too much about it online. Try ensuring that your Pytorch build is correct for your specific versions of Cuda and Cudnn. While I had the correct version of Pytorch, I resolved the issue by installing a different build of the same version.", "The build I installed is . Here are the list of versions and builds: You can download these files locally and install with conda: (assumes command is being run in a directory containing , where this file will be specific to your Cuda and Cudnn version.\nfeedback from this code PackagesNotFoundError\nthank you - I have edited my comment to reflect that I installed from a local zip."], "neg": []}
{"query": "char *bytes = (char *) data; int64_t remaining = sizeof(real) * storage->size; while (remaining > 0) { <del> ssize_t result = read(fd, bytes, remaining); </del> <ins> // we write and read in 1GB blocks to avoid bugs on some OSes ssize_t result = read(fd, bytes, THMin(remaining, 1073741824)); </ins> if (result <= 0) // 0 means EOF, which is also an error throw std::system_error(result, std::system_category()); bytes += result;", "pos": ["Running this on OSX: More details of failure in Ubuntu 14.04.1 in\nMaybe it's caused by a ? I just test this code and it doesn't work (on Mac OS X 10.12.3 with Python 3.6):\nThat thread links to this one , which says it\u2019s an OS bug and proposes a workaround patch that hasn\u2019t been merged. There\u2019s also a numpy issue thread, which links to this Torch7 commit that implements a workaround\nWe're doing a loop in case it writes less, but we always pass the full amount of remaining bytes. Putting a cap (std::min with INT_MAX) on the size of a single write might help.\nI can't reproduce it on CentOS. I think it might have been a report for a PyTorch version prior to fixing ."], "neg": []}
{"query": "input2 = torch.randn(input_size, requires_grad=True) self.assertEqual(F.cosine_similarity(input1, input2, dim=1).size(), expected_size) <ins> def test_grid_sample_unsupported_mode(self): with self.assertRaisesRegex(NotImplementedError, \"nn.functional.grid_sample got unsupported mode: 'garbage'\"): F.grid_sample(torch.tensor([]), torch.tensor([]), mode='garbage') </ins> def test_grid_sample(self): def test_cpu_against_cuda(N, C, H, W, padding_mode): def test_shape(N, C, IH, IW, H, W, padding_mode):", "pos": ["The only valid mode is 'bilinear', but one can pass anything to it and it will not complain:\nApparently, the is not passed to the backend function for . I think it should be removed.\nYes. I think the idea is that we'd like to support other modes at some point."], "neg": []}
{"query": "self.value = value def forward(self, input): <del> return F_ConstantPad2d(pad=self.padding, value=self.value)(input) </del> <ins> return F.pad(input, self.padding, 'constant', self.value) </ins> def __repr__(self): return self.__class__.__name__ + ' ' + str(self.padding)", "pos": ["After merging cuda 12.4 for nightly. I observe following failures: libtorch-cuda124-shared-with-deps-cxx11-abi-build: Build Official Docker Images 124: We need to resolve these before continuing adding CUDA 12.4 CI workflows 0 cc\nattempt 3 worked for libtorch build issue:\nFor Docker image, related discussion:\nI will instance a new bug for the docker job failure.", "Is there a way to use reflection padding instead of zero padding with nn.Conv2d ? I see there is a but the option is not in nn.Conv2d. Thanks,\nare in , but they're not exposed in the docs.\nThanks!"], "neg": []}
{"query": "padding_idx (int, optional): If given, pads the output with zeros whenever it encounters the index. max_norm (float, optional): If given, will renormalize the embeddings to always have a norm lesser than this norm_type (float, optional): The p of the p-norm to compute for the max_norm option <del> scale_grad_by_freq (boolean, optional): if given, this will scale gradients by the frequency of </del> <ins> scale_grad_by_freq (bool, optional): if given, this will scale gradients by the frequency of </ins> the words in the mini-batch. <del> sparse (boolean, optional): if ``True``, gradient w.r.t. weight matrix will be a sparse tensor. See Notes for </del> <ins> sparse (bool, optional): if ``True``, gradient w.r.t. weight matrix will be a sparse tensor. See Notes for </ins> more details regarding sparse gradients. Attributes: weight (Tensor): the learnable weights of the module of shape (num_embeddings, embedding_dim) Shape: <del> - Input: LongTensor `(N, W)`, N = mini-batch, W = number of indices to extract per mini-batch - Output: `(N, W, embedding_dim)` </del> <ins> - Input: LongTensor of arbitrary shape containing the indices to extract - Output: `(*, embedding_dim)`, where `*` is the input shape </ins> Notes: Keep in mind that only a limited number of optimizers support", "pos": ["I have a straightforward use case. Given a token (or word), lookup the embedding of the word, pass that through a linear layer and perform L2 regression. If the token corresponds to the of the embedding layer, throws an error saying OS: macOS High Sierra PyTorch version: 0.3.1 How you installed PyTorch (conda, pip, source): Tried both conda and pip Python version: 3.6.4 CUDA/cuDNN version: CPU GPU models and configuration: NA GCC version (if compiling from source): NA A script to reproduce the bug:\nI got the same trouble, and it throw this error after some epochs of training. Does anyone have some solutions?"], "neg": []}
{"query": "# caffe2::cudart is dealt with separately, due to CUDA_ADD_LIBRARY # design reason (it adds CUDA_LIBRARIES itself). set(Caffe2_PUBLIC_CUDA_DEPENDENCY_LIBS <del> caffe2::cuda caffe2::cufft caffe2::curand caffe2::nvrtc) </del> <ins> caffe2::cufft caffe2::curand) if(BUILD_CAFFE2) # Don't be deceived!  caffe2::cuda is the low-level DRIVER API, # not the actual CUDA library.  Caffe2 depends directly on nvrtc # and needs it, but ATen doesn't use it at all, and so the # dependency is unnecessary. # # BTW, if you change this so that PyTorch has this dependency # again, make sure Mac OS X GPU builds still work; there's # a decent chance the library finding algorithm picked up # on cuda.framework, which is totally not going to work when # linking. list(APPEND Caffe2_PUBLIC_CUDA_DEPENDENCY_LIBS caffe2::cuda caffe2::nvrtc) endif() </ins> if(CAFFE2_FOUND_CUDNN) LIST(APPEND Caffe2_PUBLIC_CUDA_DEPENDENCY_LIBS caffe2::cudnn) else()", "pos": ["error occurs when linking the cuda library ld: can't map file, errno=22 file '/usr/local/cuda/lib/stubs/cuda.framework' for architecture x8664 Provide a short description. I have tracked down in findlibrary(CUDACUDALIB cuda PATHS ${CUDATOOLKITROOTDIR} PATHSUFFIXES lib lib64 lib/stubs lib64/stubs lib/x64) finds the cuda.framework instead of Have confirm it was broken by this commit () Any idea why it is linking the cuda.framework? CUDA used to build PyTorch: 9.2 OS: Mac OSX 10.13.4 GCC version: Could not collect CMake version: version 3.11.1 Python version: 3.6 Is CUDA available: Yes CUDA runtime version: 9.2.64 GPU models and configuration: Could not collect Nvidia driver version: Could not collect cuDNN version: Probably one of the following: /usr/local/cuda/lib/libcudnn_static.a\ncc\nFrom cmake doc , CMake will use a -framework A, and a -F<fullPathto link the framework to the target, however it is not. I have tried separating ::cuda from Caffe2PUBLICCUDADEPENDENCY_LIBS, still not linking properly. Removing ::cuda passes the build though.\ncan you post your cmake output that contains the list of CUDA-related libraries that cmake found? Is this statically linking CUDA? What's the actual linking command that is failing (probably need to pass VERBOSE=1 to the make command to see this)? None of us will be able to repro on our end as we do not have macs with GPUs :( , but we can try to answer questions and give pointers as to where to look.\nthanks for reporting this and helping reproduce. Can you try and add this to the root around line 162? You'll then want to run to clear the directory and try again.", "[ 94%] Linking CXX shared library cd /Users/hongbinliu/pytorch/build/ && /usr/local/miniconda3/bin/cmake -E cmakelinkscript --verbose=1 /Library/Developer/CommandLineTools/usr/bin/clang++ -msse4.1 -msse4.2 --std=c++11 -Wno-deprecated -fvisibility-inlines-hidden -Wno-deprecated-declarations -DONNXNAMESPACE=onnxc2 -O2 -fPIC -Wno-narrowing -Wno-invalid-partial-specialization -Wall -Wextra -Wno-missing-field-initializers -Wno-type-limits -Wno-unused-parameter -Wno-unknown-warning-option -Wno-unknown-pragmas -DHAVEAVXCPUDEFINITION -DHAVEAVX2CPUDEFINITION -O3 -isysroot -mmacosx-version-min=10.10 -dynamiclib -Wl,-headerpadmaxinstallnames -compatibilityversion 1.0.0 -currentversion 1.0.0 -o -installname -Wl,-rpath,/usr/local/miniconda3/lib -Wl,-rpath,/usr/local/cuda/lib -Wl,-rpath,/Users/hongbinliu/pytorch/build/lib /usr/local/cuda/lib/libcudartstatic.a -Wl,-rpath,/usr/local/cuda/lib /usr/local/cuda/lib/libcudartstatic.a -Wl,-rpath,/usr/local/cuda/lib /usr/local/cuda/lib/stubs/cuda.framework /usr/local/cuda/lib/libcublasdevice.a ld: can't map file, errno=22 file '/usr/local/cuda/lib/stubs/cuda.framework' for architecture x8664 clang: error: linker command failed with exit code 1 (use -v to see invocation) make[2]: [] Error 1 make[1]: [] Error 2 make: * [all] Error 2\nThx, adding that did work. However, when I run training with cuda, it throws an RuntimeError: cannot initialize CUDA without ATen_cuda library. I guess this is another issue anyway.\nSame issue.", "<img width=\"545\" alt=\"downloads\" src=\"https://user-\"Also, I found a small problem when copying the dylibs to torch/lib, it is replacing the symbolic link with the actual file (screenshot above). Replacing cp -r to cp -R does the trick, please let me know if we want a pr.\nHave tracked down the problem was in , always return null because in Registry.h line 77 the (key) is always 0; ObjectPtrType Create(const SrcType&amp; key, Args... args) { ` Suspect CUDAHooksRegistry() is creating new instance, because of namespace? Is it working in Linux? Any help will be appreciated.\nI think Mac OS X with CUDA is broken at the moment. We think the problem has something to do with not getting a dependency on (or, I guess, , now)\nHi ezyang, thanks for your hint. I just tried before . It gives me \"Key already registered\". I will leave it to you guys, and I can help with testing, please @ me.\nThanks for giving it a try. Sounds like is being loaded correctly, so my initial guess was wrong. Would it be possible to print the keys of the map to see what is in it when you get the key already registered error message? And maybe also, you noted that there was never anything at the key \"CUDAHooks\" when you didn't have the dlopen line. Is the map always empty in that case, or is there something registered, perhaps at a different name? (I made this bug when I developed the patch and figured it out by seeing that there was another CUDAHooks thing but named slightly differently). I'd like to debug this myself but I can't get PyTorch to build with Mac OS X and CUDA at the moment. (See also the bugs I've recently filed)\nYes, the map was always empty when calling CUDAHooksRegistry(). Loading the dylib works as expected that it is reusing the same static Registry and registers the key. However calling CUDAHooksRegistry() doesn't get the same static Registry instance, instead it creates a new one. Could it be relating to how mac handles dynamic loading? e.g different memory space?\nSuccessfully reproed.", "And yes, it looks like the dynamic dependencies are setup correctly:\nBTW, fix for the cuda.framework error didn't work for me. Where exactly did you end up placing the block in the OTOH, this patch worked for me: (It's totally the wrong way to solve the problem but it might help someone)\nand I diagnosed the problem, fix coming.\nThat is great!\nGo try . You'll need to do a clean build from scratch. Also, you'll still need a cuda.framework fix of some sort or another. Working on a more robust version of this fix.\nThanks ezyang, it is working."], "neg": []}
{"query": "weight_data[:] = 4 self.assertEqual(weight_data, all_vars[4].data) <ins> @unittest.skipIf(not TEST_CUDNN, 'CUDNN not available') def test_cudnn_weight_tying(self): rnns = [ nn.LSTM(10, 20, batch_first=True, bidirectional=True), nn.GRU(10, 20, batch_first=True, bidirectional=True), nn.RNN(10, 20, batch_first=True, bidirectional=True) ] for rnn in rnns: rnn.bias_ih_l0_reverse = rnn.bias_ih_l0 rnn.cuda() input = Variable(torch.randn(5, 4, 10).cuda(), requires_grad=True) hx = Variable(torch.randn(2, 5, 20).cuda(), requires_grad=True) all_vars = [input, hx] + list(rnn.parameters()) opt = torch.optim.SGD(rnn.parameters(), lr=0.1) opt.zero_grad() if isinstance(rnn, nn.LSTM): cx = Variable(torch.randn(2, 5, 20).cuda(), requires_grad=True) all_vars[2:2] = [cx] hx = (hx, cx) with warnings.catch_warnings(record=True) as w: output = rnn(input, hx) self.assertEqual(len(w), 1) self.assertIn('weights are not part of single contiguous chunk of memory', w[0].message.args[0]) output[0].sum().backward() opt.step() with warnings.catch_warnings(record=True) as w: output_cuda = rnn(input, hx) rnn.cpu() hx = (hx[0].cpu(), hx[1].cpu()) if isinstance(rnn, nn.LSTM) else hx.cpu() output_cpu = rnn(input.cpu(), hx) self.assertEqual(output_cuda, output_cpu) </ins> @unittest.skipIf(not TEST_CUDA, 'CUDA not available') def test_cuda_rnn_fused(self): def copy_rnn(rnn1, rnn2):", "pos": ["While the fast path for cuDNN RNNs introduced in includes a check intended to fall back to copying weights if weights are moved or aliased, this only works if the aliasing happens after the weights are first compacted. But the most natural place to tie weights is inside a module's method, which runs before , the typical time when is first invoked. Here's an MWE. The three runs (CPU, weight tying before , and weight tying after ) should all give the same results, and they do before , but they have not since:"], "neg": []}
{"query": "#ifndef TH_GENERAL_INC #define TH_GENERAL_INC <ins> #define __STDC_FORMAT_MACROS </ins> #include <stdint.h> #include <stdlib.h> #include <stdio.h>", "pos": ["I am installing PyTorch from source. Python version: 2.7.14. CUDA: 8.0. GPU: gtx1080. GCC: 5.4. Processor: AMD Opteron(tm) Processor 6128 While executing \"python install\" build completes till 87% but after that I get the following errors And many more messages of the same type regarding errors with 'PRId64' I looked at similar issues online and most of the people facing this issue had an old version of gcc, so I'm a bit confused about what my exact problem is. Also, when I'm installing via conda, the installation completes, but when I run the following bit of code, I get an illegal instruction at the print command That I guess is because of the fact I'm using an AMD processor? Any help will be appreciated. Thanks\nlooks like the same issue as", "I am not able to build the current version. Getting the following error: pytorch/aten/src/TH/generic/THTensorLapack.c:860:7: warning: unused variable \u2018n\u2019 [-Wunused-variable] int n = ra-size[1]; ^ make[2]: [] Error 1 make[1]: [] Error 2 make: * [all] Error 2 I am using Anaconda as directed in the readme. I have created a conda environment and have set: export CMAKEPREFIXPATH=\"/u/anaconda2/envs/pyt4/\" As specified here: Thanks\nWhat you posted here is only a warning, and it won't stop the build. Please paste the full log into a GitHub gist and link it here\nHi, Here are the logs: I see it says Could NOT find CUDNN, I am installing it in an env which submit jobs to a cluster, do I need to install CUDNN locally?\nwhat compiler do you have? and what OS are you using? The issue is this:\nlsbrelease -a LSB Version: :core-4.1-amd64:core-4.1-noarch:cxx-4.1-amd64:cxx-4.1-noarch:desktop-4.1-amd64:desktop-4.1-noarch:languages-4.1-amd64:languages-4.1-noarch:printing-4.1-amd64:printing-4.1-noarch Distributor ID: RedHatEnterpriseServer Description: Red Hat Enterprise Linux Server release 7.2 (Maipo) Release: 7.2 Codename: Maipo uname -a Linux 1 SMP Thu Oct 29 17:29:29 EDT 2015 x8664 x8664 x8664 GNU/Linux\n- We are seeing the same error above with the build scripts that you provided in , on PyTorch master. I can bisect and let you know the commit that is breaking the build.\nThe CPU build script fails with this commit () -\nGot the same error when trying to compile under CentOS 7.2.1511 with GCC 4.8 and GCC 5.4 (no newer gcc available on that server).", "CC\nSearching issues for \"PRId64\" shows that this error popped up and has been fixed in various places in the pytorch codebase, e.g. The problem seems to be that this format specifier is not defined by default in C++ and/or old gcc versions. A hacky way to fix this is replacing with in the two files currently causing this error on master. A better way is to find the proper place for addding as described in the issue linked above.\nCould you try and see if the branch in fixes the issue? I wasn't able to reproduce the compile error locally, but based on the past issues it should be the right fix. Thanks!\nyour branch compiles without error with GCC 5.4"], "neg": []}
{"query": "PyObject *THPFunction_saved_variables(THPFunction *self, void *_unused) { <del> THPUtils_assert(!self->has_freed_buffers, \"Trying to backward through the \" \"graph second time, but the buffers have already been freed. Please \" \"specify retain_variables=True when calling backward for the first time.\"); </del> <ins> THPUtils_assert(!self->has_freed_buffers, ERR_BACKWARD_TWICE); </ins> if (!self->saved_variables) return PyTuple_New(0);", "pos": ["Usually calling backwards twice gives this message: But conv2d just segfaults. $ pt A B Segmentation fault (core dumped)"], "neg": []}
{"query": "tasks or pull requests with your environment details is helpful and appreciated. -  **CI tests failed, what does it mean?** Maybe your PR is based <del> off a broken main bracnh? You can try to rebase your change on top </del> <ins> off a broken main branch? You can try to rebase your change on top </ins> of the latest main branch. You can also see the current status of main branch's CI at https://hud.pytorch.org/. -  **What are the most high risk changes?** Anything that touches build", "pos": ["There is a typo on the , in the Frequently Asked Questions. It says: Is that mean to be a joke (broken branch), or an actual typo? If it is meant to be a joke, please ignore this issue. I will be submitting a PR with the fix for the typo.\nA typo for sure. I've force-merged your PR, thanks!"], "neg": []}
{"query": "l1=l, l2=l ) <ins> self.param = Variable(torch.Tensor(3, 5)) </ins> l = nn.Linear(10, 20) n = Net() <del> s = nn.Sequential(l, l, l, l) </del> <ins> s = nn.Sequential(n, n, n, n) </ins> self.assertEqual(num_params(l), 2) <del> self.assertEqual(num_params(n), 2) self.assertEqual(num_params(s), 2) </del> <ins> self.assertEqual(num_params(n), 3) self.assertEqual(num_params(s), 3) def test_modules(self): class Net(nn.Container): def __init__(self): super(Net, self).__init__() self.l1 = l self.l2 = l self.param = Variable(torch.Tensor(3, 5)) l = nn.Linear(10, 20) n = Net() s = nn.Sequential(n, n, n, n) self.assertEqual(list(s.modules()), [s, n, l]) </ins> def test_Sequential_getitem(self): l1 = nn.Linear(10, 20)", "pos": ["While using torch on a Macbook pro with M1 ship, using python 3.8.10 and torch 1.12.1, I came across a weird bug that happened because I forgot to put a tensor on MPS. The reproducible code is the following: As you can see, if is on MPS and on CPU, the line wrongly replaces the index in place of the index . I would expect to have an error saying both tensors should be on the same device instead. I get an error: cc\nThe error seems to happen somewhere within Also I'm getting a warning \"The operator 'aten::nonzero' is not currently supported on the MPS backend and will fall back to run on the CPU. This may have performance implications\" when printing tensors on MPS.\nHi - thanks for reporting this issue. This is actually an allowed usecase (copying CPU tensors into MPS tensors / MPS -CPU / MPS -MPS): The issue is on MPS side when copying a CPU tensor into a strided MPS tensor. I'll take a look."], "neg": []}
{"query": "main_sources += split_types(\"torch/csrc/cuda/Tensor.cpp\") if WITH_NCCL: <del> if SYSTEM_NCCL: </del> <ins> if WITH_SYSTEM_NCCL: </ins> main_libraries += ['nccl'] <ins> include_dirs.append(NCCL_INCLUDE_DIR) library_dirs.append(NCCL_LIB_DIR) </ins> else: main_link_args += [NCCL_LIB] extra_compile_args += ['-DWITH_NCCL']", "pos": ["Currently, you can have a libnccl on your system, and pytorch setup favors that instead of compiling one itself from source. The code that takes care of this is: However, this breaks gloo build because gloo gets a path to NCCL via:\nI got the error during the installation: Is it related to this problem?\nthat's a dupe of\nno, that's a consequence of , now you have to do before compiling. I think you don't want to do recursive submodule update, because that will pull in gtest and another copy of nccl.\nAlternatively we can pass new flags to gloo to prevent it from using its own NCCL\nThanks works for me.\nYou can prevent it from using its own nccl, but its cmake still has to be able to find some nccl.\nI think should be sufficient. I'm testing.\nfixes it when we dont have an external NCCL and pytorch compiles it's own NCCL?\nI don't see how NCCLEXTERNAL fixes anything, looks like in gloo NCCLEXTERNAL is used just to set dependency if gloo is compiling its own nccl (which we don't want in any case).\nThe only thing I tested when I was that the vendored nccl was not being built (even if its submodule was checked out); I tested this by adding an to one of the headers. I'll test with the other configurations.\nActually, there are a few bugs here. First, even though GLOOFLAGS looks like it's setting NCCLROOTDIR: it actually isn't because the gloo invocation is which expands to but only is passed to cmake. Suppose you fix this. nccl is still not going to be picked up, because INSTALLDIR is PyTorch's local install dir, but if you disabled nccl builds, there will never actually be a nccl here. So cmake is just going to do it's standard search to find a nccl somewhere. Inversely, if you ever built nccl, and then turn it off, we will keep picking up that locally built version until you delete it. So we shouldn't do that. Suppose you fix this.", "Then I think gloo will (continue to) correctly pick up system nccl and be built with it, and the final thing to fix is that build of PyTorch csrc doesn't know how to search for nccl headers somewhere. This is probably related to . I wonder, maybe we shouldn't do any autodetection at all, and force the user to pass us the include and library paths of nccl if they want to build with system.\nFYI, I committed a change to gloo to stop bundling NCCL itself. It will now happily pick up whatever NCCL you give it through NCCLROOTDIR or by trying to find it in the standard systems paths. See facebookincubator/gloo You do have to set to make Gloo try and find it and compile in support.\nSo to fix this, you should be able to not specify anything aside from if you want to use system NCCL, and add if you have some custom location (either a PyTorch bundled version or a custom install altogether)."], "neg": []}
{"query": "r\"\"\" expm1_() -> Tensor <del> In-place version of :meth:`~Tensor.exp` </del> <ins> In-place version of :meth:`~Tensor.expm1` </ins> \"\"\") add_docstr_all('exponential_',", "pos": ["Build error: According to , the should be modified as:\nI the same problem when installing pytorch from source. You mean the file is wrong?\nAlthough the PyYAML documentation site suggests that is possible, I could not perform it. That is a reason for the failing. This has been reported as well: I believe that the should have this instead: In fact when I did this: This indicates that the import didn't happen in PyYAML itself.\nyou guys are right, the case has to be modified to: . The original issue (my guess) is that your pyyaml had some compile error and the CLoader wasn't successfully installed."], "neg": []}
{"query": "if modules is None: raise AttributeError( \"cannot assign module before Module.__init__() call\") <ins> remove_from(self.__dict__, self._parameters, self._buffers) </ins> modules[name] = value elif modules is not None and name in modules: if value is not None:", "pos": ["They do one pair of these calls for each slicing operation.\nDoes this actually have a performance impact? These calls tend to be very quick\nYes, it's a major reason why small unrolled LSTMs (which involve in addition to two GEMMs: a chunk into four gates, for 4 pairs of calls, five unary pointwise ops, for 5 more, and five binary pointwise ops, for another 5) are much slower than they are in Theano and TensorFlow.\nAlso, there are excessive cudaGetDevice calls for each pointwise operation (9 cudaGetDevice) even without chunking, as in the following repro code: As for small unrolled LSTMs it does show up.\nThis should no longer be an issue, now that is implemented as a native ATen function. I think we're still using the Python version for tensors, but this will be resolved soon when we stop using tensors altogether."], "neg": []}
{"query": "fi if [[ \"$BUILD_ENVIRONMENT\" == *pytorch-linux-xenial-cuda9-cudnn7-py3 ]] ||  <del> [[ \"$BUILD_ENVIRONMENT\" == *pytorch-linux-trusty-py3.6-gcc7.2 ]]; then </del> <ins> [[ \"$BUILD_ENVIRONMENT\" == *pytorch-linux-trusty-py3.6-gcc7* ]]; then </ins> BUILD_TEST_LIBTORCH=1 else BUILD_TEST_LIBTORCH=0", "pos": ["We're not testing against CUDA 9.2 and we should. C.f. We should test against GCC 7 (because this is how Anaconda is building the packages), AND we should use test with this in OS X, because CUDA 9.2 now supports the default xcode shipped on OS X.\nBuilding PyTorch with CUDA 9.2 for 7.0 compute compatibility devices produces quite worrisome warnings: warning : Instruction 'vote' without '.sync' may produce unpredictable results on sm70 and later architectures warning : Instruction 'vote' without '.sync' is deprecated since PTX ISA version 6.0 and will be discontinued in a future PTX ISA version warning : Instruction 'shfl' without '.sync' is deprecated since PTX ISA version 6.0 and will be discontinued in a future PTX ISA version warning : Instruction 'shfl' without '.sync' may produce unpredictable results on sm70 and later architectures\nCUDA 9.2 CI on macOS is done in"], "neg": []}
{"query": "This is the reverse operation of the manner described in :meth:`~Tensor.gather`. <del> :attr:`self`, :attr:`index` and :attr:`src` should have same number of dimensions. It is also required that `index.size(d) <= src.size(d)` for all dimensions `d`, and that `index.size(d) <= self.size(d)` for all dimensions `d != dim`. </del> <ins> :attr:`self`, :attr:`index` and :attr:`src` (if it is a Tensor) should have same number of dimensions. It is also required that ``index.size(d) <= src.size(d)`` for all dimensions ``d``, and that ``index.size(d) <= self.size(d)`` for all dimensions ``d != dim``. </ins> Moreover, as for :meth:`~Tensor.gather`, the values of :attr:`index` must be <del> between `0` and `(self.size(dim) -1)` inclusive, and all values in a row along the specified dimension :attr:`dim` must be unique. </del> <ins> between ``0`` and ``self.size(dim) - 1`` inclusive, and all values in a row along the specified dimension :attr:`dim` must be unique. </ins> Args: dim (int): the axis along which to index", "pos": ["is not available in the docs page, but the doc string has been in . This was mentioned earlier in", "Oh, bilinear, I missed that last time I looked at it. I'll add a docstring.", "returns nothing. Might be useful to have some linter in the build process to find Python functions missing", "Pytorch 0.4.0 gives a segmentation fault with dilated 2D convolutions over a specific size. import torch import as nn model = nn.Conv2d(inchannels=64, outchannels=64, kernelsize=5, dilation=2) model(([1, 64, 1166, 1166])) # This runs model(([1, 64, 1167, 1167])) # This crashes PyTorch version: 0.4.0 Is debug build: No CUDA used to build PyTorch: 9.1.85 OS: CentOS Linux 7 (Core) GCC version: (GCC) 4.8.5 (Red Hat 4.8.5-28) CMake version: version 2.8.12.2 Python version: 3.6 Is CUDA available: Yes CUDA runtime version: 8.0.61 GPU models and configuration: GPU 0: GeForce GTX 1080 Ti GPU 1: TITAN X (Pascal) GPU 2: Quadro K2200 Nvidia driver version: 390.59 cuDNN version: Probably one of the following: /usr/local/cuda- /usr/local/cuda- /usr/local/cuda-7.5/lib64/libcudnnstatic.a /usr/local/cuda- /usr/local/cuda-8.0/lib64/libcudnn_static.a\ndo you think this is the same as the segfault you showed me a week ago on convolution?\nah yes. int overflow. I only fixed conv3d because I didn\u2019t want to hurt perf. I\u2019ll fix all of them.", "This should be an easy one.. The function is implemented, just no sign of it in the docs."], "neg": []}
{"query": "*tempValues__data = *t_data; *tempIndices__data = *tempIndices__dimOffset; }); <ins> THTensor_(free)(tempValues_); THLongTensor_free(tempIndices_); </ins> } if (!keepdim) {", "pos": ["There is a possible memory leak in () This causes memory to blow up: When using () memory usage is ok There is no leak for unitary dimensions"], "neg": []}
{"query": "import ctypes import warnings import torch.cuda <ins> import sys </ins> import os.path as path lib = None <del> # TODO: fix libname for OSX / Windows </del> <ins> # TODO: fix libname for Windows </ins> # TODO: dynamic version checks via cudnnGetVersion # TODO: load 5.1.3 if using CUDA 7.5 and 5.1.5 if using CUDA 8.0 thisdir = path.dirname(__file__) libpaths = ['', path.join(thisdir, '../../lib')] <del> libnames = ['libcudnn.so.5.1.5', 'libcudnn.so.5.1.3'] </del> <ins> if sys.platform.startswith('linux'): libnames = ['libcudnn.so.5.1.5', 'libcudnn.so.5.1.3'] elif sys.platform == 'darwin': libnames = ['libcudnn.5.dylib'] else: libnames = [] </ins> def _loadlib(): global lib", "pos": ["The official website does not provide the torchvison version installation method. Here is the error message when torchvison 0.3.0 is used. torchvision/C.cpython-36m-x8664-linux- undefined symbol: ZN2at7getTypeERKNS6TensorE That should be caused by the mismatch version between pytorch-nightly and torchvision 0.3.0.\nIt should be that matches with , but they are not provided currently.\nI tried installing it with pip as per , but now I get when importing . I did a similar thing on June 25th. It worked then. Any known fix for this?\ndo we document what to do here?\nI finally solved this by building from source with Before running that make sure to upgrade and to use . Note: this will not work if you have on your system when building since you'd get version mismatch between and . Tested only on .\nyou build pytorch from source or just torchvision? or both of them.\nI installed from conda with: I then cloned the repo and ran inside the repo. I guess this would also work if you do , but I haven't tried it.\nJust a reminder (what I encountered just now): one should uninstall previous ` from source, or they can mix up and mess up everything.\nThere are nightlies for torchvision under Windows now.\nI have problems using Pytorch with my Geforce RTX 3060, (sm86). I tried to install with help of : Create a new venv, based on python 3.7.10. Activate it. conda install pytorch-nightly cudatoolkit=10.0 -c pytorch (I also did it without the \"=10.0\") After this, Pytorch detects the sm86 with no errors, problem sm86 solved with pytorch-nightly Now, I have a problem with torchvision. tried installing torchvision with File \"C:repovisiontorchvisionmodels\", line 195, in Inception3 AttributeError: module '' has no attribute 'unused' tried installing torchvision version 0.3.0, from , but an error rose: tried installing torchvision at origin/master, from , but another error rose: Here is the list of packages installed:"], "neg": []}
{"query": "def _test_sparse_mask_fixed(self): i = self.IndexTensor([ <del> [1, 3, 3, 0, 4], [2, 1, 1, 2, 3], </del> <ins> [1, 3, 0, 4], [2, 1, 2, 3], </ins> ]) <del> v = self.ValueTensor([1, 2, 3, 4, 5]) x = self.SparseTensor(i, v, torch.Size([5, 4])) </del> <ins> v = self.ValueTensor([1, 2, 3, 4]) x = self.SparseTensor(i, v, torch.Size([5, 4])).coalesce() </ins> dense = self.ValueTensor([ [1, 2, 3, 4], [5, 6, 7, 8],", "pos": ["and I talked about the intended semantics of sparsemask when given a sparse tensor as an argument, and we came to the following conclusions: with an uncoalesced sparse tensor does not make sense, full stop. Unlike values, where your update might be linear and work correctly on an uncoalesced sparse tensor, there isn't any good reason why you want to pluck out values from the dense tensor multiple times, it is basically always a bug. the (eventual) documentation for sparsemask, we will recommend calling coalesce before doing any operations involving sparsemask (or nonlinear updates to indices/values). should add a test to sparsemask which checks if the flag of the input tensor is true, and raise an error if it is not. This is a cheap safety check which will prevent people from forgetting to call coalesce (as in (2)). There is a performance consequence to calling coalesce every iteration of training. If the gradient is producing coalesced vectors, we can alleviate this problem by adding a preprocessing step to coalesce() which tests if the input is already coalesced, and bugs out in that case. However, I think this is out of scope for this particular ticket for now."], "neg": []}
{"query": "from torch.autograd import Variable from torch.nn import Parameter from common_nn import NNTestCase, ModuleTest, CriterionTest, TestBase,  <del> module_tests, criterion_tests, TEST_CUDA, TEST_MULTIGPU, TEST_CUDNN, PRECISION </del> <ins> module_tests, criterion_tests, TEST_CUDA, TEST_MULTIGPU, TEST_CUDNN,  TEST_CUDNN_VERSION, PRECISION </ins> from common import freeze_rng_state, run_tests", "pos": ["cuDNN backend for RNN stabilizes at large mem usage. The following code stabilizes for me at 14+GiB reported by nvidia-smi. The same RNN without cuDNN uses <2.5GiB.\nOops... any idea what's going on with the garbage collector in this case?\nare you running this on a P100 GPU?\nYes, I observe similar behavior on Titan X Pascal\nhmmm, do cudnn RNNs have weird memory-oversubscribing properties? If it wasn't for CuDNN RNN, then it's not unreasonable. we use a custom CUDA allocator that will oversubscribe if needed, but will free the memory blocks it's holding onto if it hits OOM / failed allocations.\nTried the same RNN in torch and am getting ~2.6GiB on nvidia-smi.\nWith cuDNN I'm getting a constant usage of 2241MB on my machine using PyTorch (M40).\nare you using cudnnv5, or v6?\nv5\nno, cudnn does not do memory oversubscribing, and already tried the same network in lua torch with cudnn, where it gets expected memory usage. Also, he is getting OOMs on slightly bigger networks, where apparently caching allocator should have tried to free previous blocks. Interesting that can not repro this.\nJust to make sure did you set when you ran? I left it as False in the repro.\nI can repro in the docker image built from the current version on P100, and I go OOM after 8 batches with the same image on 1080, which generally should not happen - if there is enough memory to go through 1 batch, memory usage should not increase afterwards.\nNo, I just copied your script and ran it. Weird thing is that it goes OOM after 18 batches once I enable . Without it memory remains stable.\nA lot of cudaMalloc allocations go not through THCudaMalloc, and never get freed:\nSo not going through THCudaMalloc is actually expected for storages allocated with THCStorage_(newWithSize), but what's not expected is to every time cudaMalloc bytes of rnn workspace, instead of reusing earlier allocations. (~784 MB allocations are for workspace, ~156MB allocations happen somewhere in criterion).\nMy guess is that the bug is at the Python level.", "Either: Something is holding on to a Tensor longer than it should Something is creating a circular reference which prevents quick clean-up\nThe issue is that in creates cyclic references. It's made worse by the fact that it holds on to more memory than it should (like the workspace). You can see this by adding a as the first line in your for loop. With the explicit call to the GC every iteration, the memory usage doesn't explode.\nhow can these cyclic references be fixed?\nNot sure yet, but I will fix it tomorrow with"], "neg": []}
{"query": "} std::string FunctionSignature::toString() const { <ins> // TODO: consider printing more proper schema strings with defaults, optionals, etc. </ins> std::ostringstream ss; <ins> bool keyword_already = false; </ins> ss << \"(\"; int i = 0; for (auto& param : params) { if (i != 0) { ss << \", \"; } <ins> if (param.keyword_only && !keyword_already) { ss << \"*, \"; keyword_already = true; } </ins> ss << param.type_name() << \" \" << param.name; i++; }", "pos": ["Consider the following: A reasonable reading of this message is that the order of the parameters should be reversed, but this isn't equivalent: The problem is that the second signature has a keyword-only argument but it's printed. It also doesn't include defaults and such, but that's probably secondary. In this case the problem is at least flagged, but it's not clear to me whether there are other deprecated signatures that silently suggest the wrong thing. Please copy and paste the output from our (or fill out the checklist below manually). You can get the script and run it with: Collecting environment information... PyTorch version: 1.5.0a0+ Is debug build: Yes CUDA used to build PyTorch: Could not collect OS: CentOS Linux 7 (Core) GCC version: (GCC) 7.4.0 CMake version: version 3.14.0 Python version: 3.7 Is CUDA available: No CUDA runtime version: 9.2.88 GPU models and configuration: GPU 0: Tesla M40 GPU 1: Tesla M40 Nvidia driver version: 396.69 cuDNN version: /usr/local/cuda-9.2/targets/x8664- Versions of relevant libraries: [pip] numpy==1.17.2 [pip] torch==1.5.0a0+ [conda] blas 1.0 mkl [conda] mkl 2019.4 243 [conda] mkl-include 2019.4 243 [conda] mkl-service 2.3.0 py37he904b0f0 [conda] mklfft 1.0.14 py37ha843d7b0 [conda] mklrandom 1.1.0 py37hd6b4f25_0 [conda] torch 1.5.0a0+ <pip<!-- Add any other context about the problem here. --"], "neg": []}
{"query": "gI = apply_fn<Transpose>(0, 1)(gIt); } } <ins> if (should_compute_output(0) && !ggO.defined()) ggO = at::zeros_like(gO); if (should_compute_output(1) && !gI.defined()) gI = at::zeros_like(input); if (should_compute_output(2) && !gW.defined()) gW = at::zeros_like(weight); </ins> return {ggO, gI, gW}; }", "pos": ["The following code: produces this error Also, removing the ReLU from the network: causes the error to change to: However, when the second convolution is removed as well: then the code runs error free I'm on a ppc64le architecture with a P100 GPU, CUDA 8.0 and PyTorch version '0.4.0a0+'\nI'm also having this problem using branch v0.3.0\ni could reproduce this as well. we'll look into this.\nI'm having the same error :/\nsame error with gradient penalty for WGAN\nWhen you execute , are you cause memory leak\nno there shouldn't be a memory leak.\nI am having an out of memory error as well, only when applying backward to the output of this method. I am using 3 titans GPUs so memory is not a problem I would assume. def computeGP(netD, realdata, realembed, fakedata, LAMBDA): BATCHSIZE = (0) alpha = (BATCHSIZE, 1) alpha = (BATCHSIZE, int(realdata.nelement() / BATCHSIZE)).contiguous().view(BATCHSIZE, 3, 64, 64) alpha = () interpolates = alpha realdata + ((1 - alpha) fakedata) interpolates = () interpolates = autograd.Variable(interpolates, requiresgrad=True) discinterpolates, = netD(interpolates, realembed) gradients = (outputs=discinterpolates, inputs=interpolates, (()).cuda(), creategraph=True, retaingraph=True, onlyinputs=True)[0] gradientpenalty = (((2, dim=1) - 1) * 2).mean() LAMBDA return gradient_penalty `\nI am also using it with gradient penalty for WGAN, But When I use gradient_penalty.backward(), the memory leak is happen, My pytorch is 0.4.0a0+.\nI can reproduce the memory leak with this: However the leak disappears when only one convolution is used: Also, if nn.Linear is used, there is no leak. Torch version: 0.4.0a0+\nI've opened a separate issue for that memory leak:"], "neg": []}
{"query": "results = {} def _worker(i, module, input, kwargs, results, lock): <del> var_input = input while not isinstance(var_input, Variable): var_input = var_input[0] </del> <ins> var_input = get_a_var(input) </ins> try: with torch.cuda.device_of(var_input): output = module(*input, **kwargs)", "pos": ["is the culprit. Assumes that the first input is a variable or a sequence containing a variable in index 0 (recursively indexed). Workaround is to pass a dummy variable as first input, as suggested in\nCould you provide a short reproducer for testing? Think this shouldn't be too hard to fix.\nSure, here's a gist along with commands/outputs:\nfixed in master via"], "neg": []}
{"query": "PyObject *THPFunction_saved_tensors(THPFunction *self, void *_unused) { <del> THPUtils_assert(!self->has_freed_buffers, \"Trying to backward through the \" \"graph second time, but the buffers have already been freed. Please \" \"specify retain_variables=True when calling backward for the first time.\"); </del> <ins> THPUtils_assert(!self->has_freed_buffers, ERR_BACKWARD_TWICE); </ins> if (!self->saved_variables) return PyTuple_New(0);", "pos": ["Usually calling backwards twice gives this message: But conv2d just segfaults. $ pt A B Segmentation fault (core dumped)"], "neg": []}
{"query": "import torch.optim as optim import torch.legacy.optim as old_optim from torch.autograd import Variable <ins> from torch import sparse </ins> from common import TestCase, run_tests", "pos": ["and I talked about the intended semantics of sparsemask when given a sparse tensor as an argument, and we came to the following conclusions: with an uncoalesced sparse tensor does not make sense, full stop. Unlike values, where your update might be linear and work correctly on an uncoalesced sparse tensor, there isn't any good reason why you want to pluck out values from the dense tensor multiple times, it is basically always a bug. the (eventual) documentation for sparsemask, we will recommend calling coalesce before doing any operations involving sparsemask (or nonlinear updates to indices/values). should add a test to sparsemask which checks if the flag of the input tensor is true, and raise an error if it is not. This is a cheap safety check which will prevent people from forgetting to call coalesce (as in (2)). There is a performance consequence to calling coalesce every iteration of training. If the gradient is producing coalesced vectors, we can alleviate this problem by adding a preprocessing step to coalesce() which tests if the input is already coalesced, and bugs out in that case. However, I think this is out of scope for this particular ticket for now."], "neg": []}
{"query": "PyArrayObject *array_ptr = (PyArrayObject*)object.get(); if (!array_ptr || ptr != PyArray_DATA(array_ptr)) throw std::logic_error(\"invalid call to NumpyArrayAllocator::free()\"); <del> object = nullptr; </del> <ins> { AutoGIL gil; object = nullptr; } </ins> delete this; } #endif", "pos": ["When both PyTorch and Pandas are imported, calling results in a segfault. The issue was originally reported by It looks like calling returns a null pointer here: Examining in GDB: I suspect it's both PyTorch and Pandas doing threading magic and somehow getting in each other's way. I haven't yet been able to provide a minimal example that triggers this behaviour and does not use my package; it's likely that I'm doing something odd that causes this. More documentation here: Possibly related issues: and Python version: 3.6 PyTorch version: v0.2.0 Pandas version: 0.20.1"], "neg": []}
{"query": "using TensorRef = std::reference_wrapper<const Tensor>; using TensorRefList = std::initializer_list<TensorRef>; <ins> // ArrayRef is not covariant, which means there is no // implicit conversion between TensorList (aka ArrayRef<Tensor>) // and ArrayRef<Variable>.  What we do instead is manually // construct a variable_list, which itself is implicitly convertible // into an ArrayRef<Variable> (but don't return an ArrayRef<Variable>; // ArrayRef is non-owning!) static variable_list cast_tensor_list(const TensorList& tensors) { // TODO: Eliminate the intermediate vector allocation return variable_list(tensors.begin(), tensors.end()); } </ins> static VariableFlags compute_flags(const TensorRefList& tensors) { return compute_flags_tmpl(tensors); }", "pos": ["when executing:\nI'm seeing this as well as of last night. Its preceded by many warnings about : But the first error is the one about . I went back, and I get a good install as late as commit"], "neg": []}
{"query": "The lower triangular part of the matrix is defined as the elements on and below the diagonal. <del> The argument :attr:`k` controls which diagonal to consider. </del> <ins> The argument :attr:`diagonal` controls which diagonal to consider. </ins> <del> - :attr:`k` = 0, is the main diagonal. - :attr:`k` > 0, is above the main diagonal. - :attr:`k` < 0, is below the main diagonal. </del> <ins> - :attr:`diagonal` = 0, is the main diagonal. - :attr:`diagonal` > 0, is above the main diagonal. - :attr:`diagonal` < 0, is below the main diagonal. </ins> Args: input (Tensor): the input `Tensor` <del> k (int, optional): the diagonal to consider </del> <ins> diagonal (int, optional): the diagonal to consider </ins> out (Tensor, optional): The result `Tensor` Example::", "pos": ["The states that the argument controls which diagonal to consider. However, this argument is actually named in pytorch 0.2.0."], "neg": []}
{"query": "#ifdef USE_MAGMA THArgCheck(a_->nDimension == 2, 1, \"A should be 2 dimensional\"); THArgCheck(b_->nDimension == 2, 1, \"b should be 2 dimensional\"); <del> THArgCheck(a_->size[0] == b_->size[0], 2, \"size incompatible A,b\"); THArgCheck(a_->size[0] >= a_->size[1], 2, \"A should have m >= n\"); </del> <ins> THArgCheck(a_->size[0] == b_->size[0], 2, \"Expected A and b to have same size \" \"at dim 0, but they have incompatible sizes\"); THArgCheck(a_->size[0] >= a_->size[1], 2, \"Expected A with shape (m x n) to have \" \"m >= n. The case for m < n is not implemented yet.\"); </ins> THCTensor *a = THCTensor_(newColumnMajor)(state, ra_, a_); THCTensor *b = THCTensor_(newColumnMajor)(state, rb_, b_);", "pos": ["I am currently using PyTorch 0.3.0 and I am noticing some (possibly strange) differences between the behavior of when operating on CPU tensors versus GPU ones, in particular relating to whether the number of rows of A is allowed to be larger than the number of columns or not. Here is some minimal code that reproduces my problem: The first solve on the CPU works fine; however, the second solve gives me the error: Is this a limitation of the CUDA implementation of that PyTorch uses, or otherwise intended? If so, any guidance about a possible work-around would be greatly appreciated. Our only thought on a work-around was to manually QR decompose A with and then use , but doesn't support CUDA tensors.\nMAGMA gels doesn't implement the m < n case."], "neg": []}
{"query": "(Cross, (), ((S, 3), (S, 3))), (Cross, (), ((S, 3, S), (S, 3, S), 1), 'dim'), (Inverse, (), ((S, S),), '', (), [skipIfNoLapack]), <ins> (Gesv, (), ((S, S), (S, S)), '', (), [skipIfNoLapack]), </ins> (Clone, (), ((S, M, S),)), (Squeeze, (), ((S, 1, M, 1),)), # TODO: enable neg dim checks", "pos": ["There seems to be regression in the dynamo export of the MONAI VNet model as regards . Export script: The versions I'm using is torch 2.5.0.dev20240625+cu121 Previous torch versions such as were able to export this model successfully. Perhaps dynamo was generating FX graph without the expm1 node previously? cc: PyTorch version: 2.5.0.dev20240625+cu121 Is debug build: False CUDA used to build PyTorch: 12.1 ROCM used to build PyTorch: N/A OS: Ubuntu 22.04.4 LTS (x8664) GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 Clang version: Could not collect CMake version: version 3.29.0 Libc version: glibc-2.35 Python version: 3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0] (64-bit runtime) Python platform: Linux-6.5.0-35-generic-x8664-with-glibc2.35 Is CUDA available: True CUDA runtime version: 12.4.131 CUDAMODULELOADING set to: LAZY GPU models and configuration: GPU 0: NVIDIA RTX A5000 Nvidia driver version: 535.171.04 cuDNN version: Probably one of the following: /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- HIP runtime version: N/A MIOpen runtime version: N/A Is XNNPACK available: True CPU: Architecture: x8664 CPU op-mode(s): 32-bit, 64-bit Address sizes: 46 bits physical, 48 bits virtual Byte Order: Little Endian CPU(s): 20 On-line CPU(s) list: 0-19 Vendor ID: GenuineIntel Model name: Intel(R) Core(TM) i9-9820X CPU @ 3.", "30GHz CPU family: 6 Model: 85 Thread(s) per core: 2 Core(s) per socket: 10 Socket(s): 1 Stepping: 4 CPU max MHz: 4200.0000 CPU min MHz: 1200.0000 BogoMIPS: 6599.98 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constanttsc art archperfmon pebs bts repgood nopl xtopology nonstoptsc cpuid aperfmperf pni pclmulqdq dtes64 monitor dscpl vmx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse41 sse42 x2apic movbe popcnt tscdeadlinetimer aes xsave avx f16c rdrand lahflm abm 3dnowprefetch cpuidfault epb catl3 cdpl3 invpcidsingle pti ssbd mba ibrs ibpb stibp tprshadow flexpriority ept vpid eptad fsgsbase tscadjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm mpx rdta avx512f avx512dq rdseed adx smap clflushopt clwb intelpt avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqmllc cqmoccupllc cqmmbmtotal cqmmbmlocal dtherm ida arat pln pts hwp hwpactwindow hwpepp hwppkgreq vnmi mdclear flushl1d archcapabilities Virtualization: VT-x L1d cache: 320 KiB (10 instances) L1i cache: 320 KiB (10 instances) L2 cache: 10 MiB (10 instances) L3 cache: 16.5 MiB (1 instance) NUMA node(s): 1 NUMA node0 CPU(s): 0-19 Vulnerability Gather data sampling: Mitigation;", "Microcode Vulnerability Itlb multihit: KVM: Mitigation: VMX disabled Vulnerability L1tf: Mitigation; PTE Inversion; VMX conditional cache flushes, SMT vulnerable Vulnerability Mds: Mitigation; Clear CPU buffers; SMT vulnerable Vulnerability Meltdown: Mitigation; PTI Vulnerability Mmio stale data: Mitigation; Clear CPU buffers; SMT vulnerable Vulnerability Retbleed: Mitigation; IBRS Vulnerability Spec rstack overflow: Not affected Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and _user pointer sanitization Vulnerability Spectre v2: Mitigation; IBRS; IBPB conditional; STIBP conditional; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Mitigation; Clear CPU buffers; SMT vulnerable Versions of relevant libraries: [pip3] numpy==1.24.4 [pip3] onnx==1.16.0 [pip3] onnxscript==0.1.0.dev20240625 [pip3] optree==0.11.0 [pip3] pytorch-quantization==2.1.2 [pip3] pytorch-triton==3.0.0+ [pip3] torch==2.5.0.dev20240625+cu121 [pip3] torch-tensorrt==2.3.0a0 [pip3] torchaudio==2.4.0.dev20240625+cu121 [pip3] torchdata==0.7.1a0 [pip3] torchtext==0.17.0a0 [pip3] torchvision==0.20.0.dev20240625+cu121 [conda] Could not collect\nI am not very sure, but this looks like the kind of ops that should get decomp'ed. do you know? could you help take a look as part of the op support work you are doing? Thanks!\nYes, it looks like there is a graph-changed in dynamo, which dynamo now decomposes elu for some reasons. Using can preserve elu and then successfully export though.", "Yes can run but we would need to be resolved first since the weights are still not exported properly.\nI can take a look at this.\nshould fix this.", "Here, I can't export trivial net with sort() using dynamoexport(). Legacy () works fine, dynamoexport errors out : Unsupported FX nodes: {'callfunction': ['']}. Pytorch nightly 06/03 cc\na missing op, could you take this?\nNot sure if related:\nHmmm. Ideally, ONNX should add a sort op. It doesn't have one. I wonder what the legacy exporter converts it to.\nshould fix this issue.", "Are there any plans to implement something like tensorflow's and ? exists but it doesn't have the backward step implemented.\nOne workaround for the moment would be to write as , which should support backprop since . Adding support for should be trivial, the gradient for it can be found in\nI'll add backward support for , which should be equivalent to . I'll try to do it later today.\nHow difficult would it be to add in batched versions (where only the final two dimensions are square) of GESV and inverse? Are there any plans for this?\nIf we have batched version of , then adapting the autograd part is very simple. There is an open issue already that discuss implementing batched gesv\nActually, there is an open PR for that, so it might be coming soon\nI'll rebase the batched version of gesv. It'll come in soon\nMay I ask why gesv only support . As I saw from , it support that A is a MxN matrix, and B is a NxK matrix, then solve a matrix X (NxK). If I want to use a non-square matrix A in gesv, would you mind to let me know how to implement in PyTorch?\nthis doesn't exactly answer your question but you can add zero-rows to the A matrix until it becomes square.\nthanks a lot, I find the solution, can solve the problem.\nAny plan to add gradient calculation for gels?\nI am also wondering if will have differentiation support. My current workaround is to use instead\nI actually implemented overdetermined case following tensorflow implementation. For backward pass I only did basic grad test. On Monday, March 18, 2019, Johnson Zhong <:\nThanks, seems to work fine!"], "neg": []}
{"query": "return None return match.group(1) <ins> def run_and_return_first_line(run_lambda, command): \"\"\"Runs command using run_lambda and returns first line if output is not empty\"\"\" rc, out, _ = run_lambda(command) if rc != 0: return None return out.split('n')[0] </ins> def get_conda_packages(run_lambda): if get_platform() == 'win32':", "pos": ["TensorPipe uses some advanced features of the Linux kernel. It does so on a \"best effort\" basis (i.e., use them if available, fall back on something else if not). This logic is unfortunately hard to test because it would require specific kernel versions on our CI machines. So it occasionally happens that we find out about issues when users report them. To help investigate these reports we've found it useful to know the kernel version and the glibc version used by the user (both when building and when running, if they're different). It would be great if these could be to the report produced by collect_env.\nCC"], "neg": []}
{"query": "from tools.setup_helpers.env import check_env_flag from tools.setup_helpers.cuda import WITH_CUDA, CUDA_HOME from tools.setup_helpers.cudnn import WITH_CUDNN, CUDNN_LIB_DIR, CUDNN_INCLUDE_DIR <ins> from tools.setup_helpers.nccl import WITH_NCCL, WITH_SYSTEM_NCCL, NCCL_LIB_DIR, NCCL_INCLUDE_DIR, NCCL_ROOT_DIR </ins> from tools.setup_helpers.split_types import split_types DEBUG = check_env_flag('DEBUG') WITH_DISTRIBUTED = not check_env_flag('NO_DISTRIBUTED') WITH_DISTRIBUTED_MW = WITH_DISTRIBUTED and check_env_flag('WITH_DISTRIBUTED_MW') <del> WITH_NCCL = WITH_CUDA and platform.system() != 'Darwin' SYSTEM_NCCL = False </del> ################################################################################", "pos": ["Currently, you can have a libnccl on your system, and pytorch setup favors that instead of compiling one itself from source. The code that takes care of this is: However, this breaks gloo build because gloo gets a path to NCCL via:\nI got the error during the installation: Is it related to this problem?\nthat's a dupe of\nno, that's a consequence of , now you have to do before compiling. I think you don't want to do recursive submodule update, because that will pull in gtest and another copy of nccl.\nAlternatively we can pass new flags to gloo to prevent it from using its own NCCL\nThanks works for me.\nYou can prevent it from using its own nccl, but its cmake still has to be able to find some nccl.\nI think should be sufficient. I'm testing.\nfixes it when we dont have an external NCCL and pytorch compiles it's own NCCL?\nI don't see how NCCLEXTERNAL fixes anything, looks like in gloo NCCLEXTERNAL is used just to set dependency if gloo is compiling its own nccl (which we don't want in any case).\nThe only thing I tested when I was that the vendored nccl was not being built (even if its submodule was checked out); I tested this by adding an to one of the headers. I'll test with the other configurations.\nActually, there are a few bugs here. First, even though GLOOFLAGS looks like it's setting NCCLROOTDIR: it actually isn't because the gloo invocation is which expands to but only is passed to cmake. Suppose you fix this. nccl is still not going to be picked up, because INSTALLDIR is PyTorch's local install dir, but if you disabled nccl builds, there will never actually be a nccl here. So cmake is just going to do it's standard search to find a nccl somewhere. Inversely, if you ever built nccl, and then turn it off, we will keep picking up that locally built version until you delete it. So we shouldn't do that. Suppose you fix this.", "Then I think gloo will (continue to) correctly pick up system nccl and be built with it, and the final thing to fix is that build of PyTorch csrc doesn't know how to search for nccl headers somewhere. This is probably related to . I wonder, maybe we shouldn't do any autodetection at all, and force the user to pass us the include and library paths of nccl if they want to build with system.\nFYI, I committed a change to gloo to stop bundling NCCL itself. It will now happily pick up whatever NCCL you give it through NCCLROOTDIR or by trying to find it in the standard systems paths. See facebookincubator/gloo You do have to set to make Gloo try and find it and compile in support.\nSo to fix this, you should be able to not specify anything aside from if you want to use system NCCL, and add if you have some custom location (either a PyTorch bundled version or a custom install altogether)."], "neg": []}
{"query": "export PATH=\"$PWD:$PATH\" popd <del> time test/run_test.sh </del> <ins> time test/run_test.sh -- -v </ins> rm -rf ninja", "pos": ["If a test segfaults, it's hard to determine which test it is because the output looks like: ! I can work on this if we think this is a good idea.\nLast time I asked about this, he didn't want to because it adds a lot of scrolling. I also did some work to move the CI tests to so that we can output XML status files, but (1) this is unlikely to help in case of segfault, and (2) you can't do this straightforwardly, because we run some test files multiple times with different environment variables, and the XML exporter isn't really set up to handle this case. For the record, I am personally OK with just slapping on the tests and calling it a day.\nin the conbuild sounds good to me too\nclosed via"], "neg": []}
{"query": "std::unique_ptr<npy_intp[]> sizes; std::unique_ptr<npy_intp[]> strides; <ins> THPUtils_assert(PyTuple_Size(args) == 0, \"numpy() takes no positional arguments\"); </ins> // Numpy and Torch disagree on empty tensors. In Torch, an empty tensor // is a tensor with zero dimensions. In Numpy, a tensor with zero dimensions // is a scalar (with one element). So we'll convert an empty Torch tensor", "pos": ["Currently due to the way cwrap generates wrapping code for the c source, passing an argument to .numpy() would not result in any error. I think probably this should not be the desired behaviour for the API?\nfixed via"], "neg": []}
{"query": "// See Note [Numpy memory management] void* NumpyArrayAllocator::realloc(void* ptr, ptrdiff_t size) { <del> PyArrayObject *array_ptr = (PyArrayObject*)object.get(); if (array_ptr && ptr == PyArray_DATA(array_ptr)) { void* newPtr = this->malloc(size); memcpy(newPtr, ptr, std::min((size_t) size, (size_t) PyArray_NBYTES(array_ptr))); // Whee! We're done! object = nullptr; return newPtr; } return allocator->realloc(allocatorContext, ptr, size); </del> <ins> throw std::logic_error(\"NumpyArrayAllocator::realloc() not supported\"); </ins> } // See Note [Numpy memory management]", "pos": ["When both PyTorch and Pandas are imported, calling results in a segfault. The issue was originally reported by It looks like calling returns a null pointer here: Examining in GDB: I suspect it's both PyTorch and Pandas doing threading magic and somehow getting in each other's way. I haven't yet been able to provide a minimal example that triggers this behaviour and does not use my package; it's likely that I'm doing something odd that causes this. More documentation here: Possibly related issues: and Python version: 3.6 PyTorch version: v0.2.0 Pandas version: 0.20.1"], "neg": []}
{"query": "\"\"\"Fills the input Tensor or Variable with values according to the method described in \"Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification\" - He, K. et al. (2015), using a uniform distribution. The resulting tensor will have values sampled from :math:`U(-bound, bound)` where <del> :math:`bound = sqrt{2 / ((1 + a^2) times fan_in)} times sqrt{3}`. Also known as He initialisation. </del> <ins> :math:`bound = sqrt{2 / ((1 + a^2) times fan_in)} times sqrt{3}`. Also known as He initialisation. </ins> Args: tensor: an n-dimensional torch.Tensor or autograd.Variable", "pos": ["why 1.21.6? because that's the oldest supported numpy on windows/python3.10: all cc\nHere is the test for pinning numpy to 1.21.6 The reason its was not pinned to this version is that conda can't properly resolve all the dependencies at this step here:\nIt looks like you are correct, I will put up a PR to pin lower then already pinned:\nHere is the PR:"], "neg": []}
{"query": "sequence.add_(start - 1) return sequence.resize_(*size) <del> def test_index(self): reference = self._consecutive((3, 3, 3)) </del> <ins> @staticmethod def _test_index(self, conv_fn): def consec(size, start=1): sequence = torch.ones(int(torch.Tensor(size).prod(0)[0])).cumsum(0) sequence.add_(start - 1) return sequence.view(*size) reference = conv_fn(consec((3, 3, 3))) </ins> # empty tensor indexing <del> self.assertEqual(reference[torch.LongTensor()], reference.new()) </del> <ins> self.assertEqual(reference[conv_fn(torch.LongTensor())], reference.new()) </ins> <del> self.assertEqual(reference[0], self._consecutive((3, 3)), 0) self.assertEqual(reference[1], self._consecutive((3, 3), 10), 0) self.assertEqual(reference[2], self._consecutive((3, 3), 19), 0) self.assertEqual(reference[0, 1], self._consecutive((3,), 4), 0) self.assertEqual(reference[0:2], self._consecutive((2, 3, 3)), 0) </del> <ins> self.assertEqual(reference[0], consec((3, 3)), 0) self.assertEqual(reference[1], consec((3, 3), 10), 0) self.assertEqual(reference[2], consec((3, 3), 19), 0) self.assertEqual(reference[0, 1], consec((3,), 4), 0) self.assertEqual(reference[0:2], consec((2, 3, 3)), 0) </ins> self.assertEqual(reference[2, 2, 2], 27, 0) <del> self.assertEqual(reference[:], self._consecutive((3, 3, 3)), 0) </del> <ins> self.assertEqual(reference[:], consec((3, 3, 3)), 0) </ins> # indexing with Ellipsis self.assertEqual(reference[..., 2], torch.Tensor([[3, 6, 9],", "pos": ["Arguably, indexing a Tensor by a LongTensor with no dimension should return a LongTensor with no dimension. Consider the following case: This breaks indexing by indices when the list of indices is empty, which can be hard to detect. In my use-case, I use such a masking to select rows of a tensor corresponding to \"valid\" entries in a vector; I had not considered that it would not work when there are no valid entries.\nThanks! I'll take a look."], "neg": []}
{"query": "<del> import torch from torch.autograd.function import Function from torch._thnn import type2backend import torch.backends.cudnn as cudnn class BatchNorm(Function): def __init__(self, running_mean, running_var, training, momentum, eps): super(BatchNorm, self).__init__() self.running_mean = running_mean self.running_var = running_var self.training = training self.momentum = momentum self.eps = eps def forward(self, input, weight=None, bias=None): self.save_for_backward(input, weight, bias) # don't use cuDNN for half inputs because cuDNN requires the weight and # bias tensors to be floats, unlike THCUNN which requires half tensors. self.use_cudnn = (cudnn.is_acceptable(input) and cudnn.version() > 5110 and weight is not None and bias is not None and not isinstance(input, torch.cuda.HalfTensor)) # temporary buffers used in forward and backward num_features = input.size(1) _save_mean = input.new(num_features) _save_std = input.new(num_features) output = input.new(input.size()) if self.use_cudnn: torch._C._cudnn_batch_norm_forward( input, output, weight, bias, self.running_mean, self.running_var, _save_mean, _save_std, self.training, self.momentum, self.eps) else: backend = type2backend[type(input)] backend.BatchNormalization_updateOutput( backend.library_state, input, output, weight, bias, self.running_mean, self.running_var, _save_mean, _save_std, self.training, self.momentum, self.eps) if self.requires_grad: self._save_mean = _save_mean self._save_std = _save_std return output def backward(self, grad_output): input, weight, bias = self.saved_tensors grad_input, grad_weight, grad_bias = None, None, None if self.needs_input_grad[0] or self.use_cudnn: grad_input = input.new(input.size()) if (len(self.needs_input_grad) > 1 and self.needs_input_grad[1]) or self.use_cudnn: grad_weight = weight.new(weight.size()).zero_() if (len(self.needs_input_grad) > 1 and self.needs_input_grad[2]) or self.use_cudnn: grad_bias = bias.new(bias.size()).zero_() if self.use_cudnn and self.training: # cudnn does not support backward in evaluate mode torch._C._cudnn_batch_norm_backward( input, grad_output, grad_input, grad_weight, grad_bias, weight, self.running_mean, self.running_var, self._save_mean, self._save_std, self.training, self.eps) else: grad_output = grad_output.contiguous() backend = type2backend[type(input)] backend.BatchNormalization_backward( backend.library_state, input, grad_output, grad_input, grad_weight, grad_bias, weight, self.running_mean, self.running_var, self._save_mean, self._save_std, self.training, 1.0, self.eps) return grad_input, grad_weight, grad_bias </del>", "pos": ["As Arthur Szlam reports, fb-internal cudnn is still lagging behind, and giving batch-size 1024 with batchnorm is raising an error from cudnn. Need to check for compile-time version and disable this codepath\nThe 1024 limitation was removed in 5.1.10"], "neg": []}
{"query": "for i in range(10): Variable(torch.randn(10, 10), _grad_fn=CollectOnDelete()) <del> @unittest.skipIf(not torch.cuda.is_available() or torch.cuda.device_count() < 2, \"CUDA not available or <2 GPUs detected\") </del> <ins> @unittest.skipIf(torch.cuda.device_count() < 2, \"no multi-GPU\") </ins> def test_unused_output_gpu(self): from torch.nn.parallel._functions import Broadcast x = Variable(torch.randn(5, 5).float().cuda(), requires_grad=True)", "pos": ["Consider following code, and run it with multi gpus (e.g. 4): It will output: That is, the default device would always be zero even in DataParallel. My pytorch version is 0.1.122. I think this is not the desired behavior. It would cause some troubles. I tried to insert my own cuda kernel into backward to calculate the gradients, it become very slow, and I fixed it by (grad.getdevice()). Anyway, I think currentdevice in forward and backward should be the same, or could anyone explain to me why they are different?\nthis seems like something we should fix: ?\n<!-- drci-comment-start --:pagefacingup: Preview :pagefacingup: Preview :question: Need help or want to give feedback on the CI? Visit our Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. <!-- drci-comment-end --"], "neg": []}
{"query": "res = [] for g in range(self.groups): def group(tensor, dim=None): <ins> if tensor is None: return None </ins> if dim is None: dim = 0 if tensor.dim() == 1 else 1 n = tensor.size(dim) // self.groups", "pos": ["x = Variable(((16, 64, 8, 8))) res = nn.Conv2d(16, 16, kernelsize=3, groups=2, bias=False)(x) results in: return ('updateoutput', input, weight, bias) File \"\", line 158, in thnn groupedargs += [group(t) for t in args] File \"\", line 153, in group dim = 0 if () == 1 else 1 AttributeError: 'NoneType' object has no attribute 'dim' The in line 158 is because there is no bias. I am running on CPU so this is in the code."], "neg": []}
{"query": "self.assertEqual(n[1], l2) self.assertEqual(n[2], l3) self.assertEqual(n[3], l4) <ins> self.assertEqual(n[1:], nn.Sequential(l2, l3, l4)) self.assertEqual(n[3:], nn.Sequential(l4)) self.assertEqual(n[:-1], nn.Sequential(l1, l2, l3)) self.assertEqual(n[:-3], nn.Sequential(l1)) self.assertEqual(n[::-1], nn.Sequential(l4, l3, l2, l1)) </ins> def test_ModuleList(self): modules = [nn.ReLU(), nn.Linear(5, 5)]", "pos": ["it'd be great to pick out elements of of a sequential with array slicing:\nI agree this idea!\nHowever,\nthis has come up again (last in ). Shall we reconsider?\nI attempted getting a to work. Is this fine? If this is fine, I can issue a PR too. Let me know what you think. I believe this modification needs to be done in the class.\ncc: This issue has been open for quite sometime."], "neg": []}
{"query": "id_map = {old_id: p for old_id, p in zip(chain(*(g['params'] for g in saved_groups)), chain(*(g['params'] for g in groups)))} <del> state = {id_map.get(k, k): v for k, v in state_dict['state'].items()} </del> <ins> state = defaultdict( dict, {id_map.get(k, k): v for k, v in state_dict['state'].items()}) </ins> # Update parameter groups, setting their 'params' value def update_group(group, new_group):", "pos": ["The state attribute in the opt. class is a defaultdict which helps to add new parameter states in the state dict as they are being used while training. Once the optimizer's statedict is saved and loaded, the state attribute is a dict and not a defaultdict which assumes that before saving, all the parameters in the network are present in the state dict. When loading the opt. using a saved statedict the state attribute should be a defaultdict. I currently do the following to have this behavior Corresponding\nHi all, it is my first time trying to contribute to an open-source project. I took a look on the bug. It seems that it is related to the function \"statedict()\" of class \"\". When the state is transformed to packedstate, it uses dict instead of defaultdict, which is inconsistent with the type of state. The solution I am proposing is to let the packedstate be defaultdict to maintain consistency with Here is my proposed change: packedstate = defaultdict(dict, {(id(k) if isinstance(k, Variable) else k): v for k, v in ()}) Is there anything I am missing? Is \"\" in the pytorch/test enough to validate my change? Where could I add a test point to ensure the type of state remain the same after saving and loading? Thanks a lot for your patience.\nclosed via thanks to !! Thanks a lot !"], "neg": []}
{"query": "std::tie(self, indices) = transposeToFront(self, indices); } auto linearIndex = computeLinearIndex(self, indices); <del> return std::make_tuple<>(self, linearIndex); </del> <ins> return std::make_tuple(self, linearIndex); </ins> } Tensor index(const Tensor & self, TensorList indices) {", "pos": ["In PyTorch master 0.4 , variables cannot be indexed. This used to work in PyTorch 0.2 :\nThis is breaking some of Pyro's inference algorithms on PyTorch master:"], "neg": []}
{"query": "if not self.shutdown: self.shutdown = True self.done_event.set() <del> # if worker_manager_thread is waiting to put while not self.data_queue.empty(): self.data_queue.get() </del> <ins> # if worker_manager_thread is waiting to put, make place for it try: while not self.data_queue.empty(): self.data_queue.get() except FileNotFoundError: # FileNotFoundError can happen when we rebuild the fd # fetched from the queue but the socket is already closed # from the worker side (e.g. due to Python shutting down). pass </ins> for _ in self.workers: self.index_queue.put(None) # done_event should be sufficient to exit worker_manager_thread,", "pos": ["I'm creating a in my script ([]), and I get an error when this script is close to ending. I am using MPI through mpi4py, not the MPI PyTorch distributed backend. This error only happens when under , not under (which completes successfully). Nothing in the traceback reveals this is an issue with my code, and I haven't touched the DataLoader at all. I've been running this code for a couple months in the same environment, and I haven't seen this bug pop before. That is, this command does not raise an error: However, this command does raise an error: Traceback: PyTorch version: 0.3.1.post2 How you installed PyTorch (conda, pip, source): conda Python version: 3.6.1 OS/CUDA/cuDNN version: A modification of the Deep Learning AMI GPU models and configuration: Amazon instance []\nI do not see this bug with Pytorch 0.3.0.post4`"], "neg": []}
{"query": "@property def saved_tensors(self): flat_tensors = super(NestedIOFunction, self).saved_tensors <del> return _map_tensor_fromiter(iter(flat_tensors))(self._to_save_nested) </del> <ins> return _unflatten(flat_tensors, self._to_save_nested) </ins> def mark_dirty(self, *args, **kwargs): self.dirty_tensors = tuple(_iter_tensors((args, kwargs)))", "pos": ["cuDNN backend for RNN stabilizes at large mem usage. The following code stabilizes for me at 14+GiB reported by nvidia-smi. The same RNN without cuDNN uses <2.5GiB.\nOops... any idea what's going on with the garbage collector in this case?\nare you running this on a P100 GPU?\nYes, I observe similar behavior on Titan X Pascal\nhmmm, do cudnn RNNs have weird memory-oversubscribing properties? If it wasn't for CuDNN RNN, then it's not unreasonable. we use a custom CUDA allocator that will oversubscribe if needed, but will free the memory blocks it's holding onto if it hits OOM / failed allocations.\nTried the same RNN in torch and am getting ~2.6GiB on nvidia-smi.\nWith cuDNN I'm getting a constant usage of 2241MB on my machine using PyTorch (M40).\nare you using cudnnv5, or v6?\nv5\nno, cudnn does not do memory oversubscribing, and already tried the same network in lua torch with cudnn, where it gets expected memory usage. Also, he is getting OOMs on slightly bigger networks, where apparently caching allocator should have tried to free previous blocks. Interesting that can not repro this.\nJust to make sure did you set when you ran? I left it as False in the repro.\nI can repro in the docker image built from the current version on P100, and I go OOM after 8 batches with the same image on 1080, which generally should not happen - if there is enough memory to go through 1 batch, memory usage should not increase afterwards.\nNo, I just copied your script and ran it. Weird thing is that it goes OOM after 18 batches once I enable . Without it memory remains stable.\nA lot of cudaMalloc allocations go not through THCudaMalloc, and never get freed:\nSo not going through THCudaMalloc is actually expected for storages allocated with THCStorage_(newWithSize), but what's not expected is to every time cudaMalloc bytes of rnn workspace, instead of reusing earlier allocations. (~784 MB allocations are for workspace, ~156MB allocations happen somewhere in criterion).\nMy guess is that the bug is at the Python level.", "Either: Something is holding on to a Tensor longer than it should Something is creating a circular reference which prevents quick clean-up\nThe issue is that in creates cyclic references. It's made worse by the fact that it holds on to more memory than it should (like the workspace). You can see this by adding a as the first line in your for loop. With the explicit call to the GC every iteration, the memory usage doesn't explode.\nhow can these cyclic references be fixed?\nNot sure yet, but I will fix it tomorrow with"], "neg": []}
{"query": "using thpp::Tensor; <ins> #ifndef CUDNN_BN_MIN_EPSILON #define CUDNN_BN_MIN_EPSILON 0 #endif </ins> auto BatchNormForward::apply(const variable_list& inputs) -> variable_list { check_input_variables(\"BatchNorm\", inputs, 3, 1);", "pos": ["Hi, I have two linux servers. My code works in one machine but not the other. In the one that it doesn't work, the error message is: occured during batch normalization( convolution is fine). Also, the same code works in CPU but not in GPU. I guess this problem is mainly related to GPU libraries like cuda or cudnn. The one that doesn't work has the following configurations: Nvidia driver 375.26 cuda 8.0.61 cudnn 5.1.10 While in the one it does work, I have the following configurations: Nvidia driver 367.48 cuda 8.0.44 cudnn 5.1.5 TensorFlow can run in the above environment flawless though. So I am wondering if this is caused by the incompatibility between PyTorch 0.1.12 and cuda 8.0.61/cudnn 5.1.10 ?\nyou use the pytorch binaries, we ship cudnn6 in them, use it it's great. 5.1.10 maybe has a bug. cc: . do you have a small test case to reproduce this? does not use CuDNN batchnorm by default, so maybe that's why?\nSmall test case would help, thanks.\nAfter some experimentation, I have pin pointed where exactly the error comes from. It seems batch norm in cudnn 5.1.10 doesn't support eps as small as 1e-6. A small test case would be:\nYes, cudnn.h for 5.1.10 defines CUDNNBNEPSILON to be 1e-5 (minimum allowable value). I can't check header for 5.1.5. Even with 1e-5 epsilon for many input data distributions the results become pretty unstable. Is there a case for this restriction to be lifted, does using very small epsilon provide benefits?\nI see. Not sure if small epsilon is necessary. I am using 1e-6 just to replicate others' configurations. Should PyTorch provide more informative error message about this restriction?\nIdeally, CUDNN should provide more informative error messages (pytorch can not know why cudnn is returning an error code in each particular case) , but that's a long standing discussion.\nIt just hits on me.", "As mentioned by , TensorFlow does not use CuDNN batchnorm by default. So is this possible in Pytorch, to not use default CuDNN batchnorm?\nNever mind. I found the switch in ."], "neg": []}
{"query": "self.assertEqual(list(tensor), [4, 4, 4, 4]) p.join() <ins> def _test_empty_tensor_sharing(self, dtype): q = mp.Queue() empty = torch.tensor([], dtype=dtype) q.put(empty) out = q.get(timeout=1) self.assertEqual(out, empty) def test_empty_tensor_sharing(self): self._test_empty_tensor_sharing(torch.float32) self._test_empty_tensor_sharing(torch.int64) @unittest.skipIf(not torch.cuda.is_available(), 'CUDA not available') def test_empty_tensor_sharing_cuda(self): self._test_empty_tensor_sharing(torch.cuda.float32) self._test_empty_tensor_sharing(torch.cuda.int64) </ins> def _test_autograd_sharing(self, var): ready = mp.Event() master_modified = mp.Event()", "pos": ["PyTorch version: 0.4.0a0+ Repro: Result:\nEmm. I fail to reproduce this bug in Python 3.5.0 and Mac OS 10.13.3. What's your system environment?\nThe bug is probably specific to the sharing mode. Mac OS uses by default."], "neg": []}
{"query": ".. automethod:: repeat .. automethod:: resize_ .. automethod:: resize_as_ <del> .. automethod:: retain </del> .. automethod:: round .. automethod:: round_ .. automethod:: rsqrt", "pos": ["Dataloader call getitem as many times as indices in the current batch. In case datasets support a list of indices in one call, or a native python slice object, add a getbatch (optional) to pytorch dataset class. If dataloader sees the dataset has such a method implemented, make it fetch the batch corresping to the list of indices with one getbatch call rather than many getitem calls. This saves the trouble of calling getitem many times in cases where for example a disk file has to be opened then closed for every call and the underlying data structure supports a slicing or lists of indices for example hdf5 files.\nwill support something similar\nis the latest version of this."], "neg": []}
{"query": "// This Engine's ReadyQueues and their corresponding threads are leaked here Engine::~Engine() = default; <del> auto Engine::thread_main(std::shared_ptr<ReadyQueue> queue) -> void { </del> <ins> auto Engine::thread_main(std::shared_ptr<ReadyQueue> queue, int device) -> void { </ins> THInferNumThreads(); <ins> AutoGPU guard(device); </ins> while (1) { FunctionTask task = queue->pop_back(); if (!task.base->has_error.load()) {", "pos": ["Consider following code, and run it with multi gpus (e.g. 4): It will output: That is, the default device would always be zero even in DataParallel. My pytorch version is 0.1.122. I think this is not the desired behavior. It would cause some troubles. I tried to insert my own cuda kernel into backward to calculate the gradients, it become very slow, and I fixed it by (grad.getdevice()). Anyway, I think currentdevice in forward and backward should be the same, or could anyone explain to me why they are different?\nthis seems like something we should fix: ?\n<!-- drci-comment-start --:pagefacingup: Preview :pagefacingup: Preview :question: Need help or want to give feedback on the CI? Visit our Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. <!-- drci-comment-end --"], "neg": []}
{"query": "return _iter <ins> def _unflatten(input, proto): # unflatten a list or tuple input into a nested list/tuple structure # specified by proto def unflatten_helper(input, proto): res = [] if not isinstance(proto, (list, tuple)): return input[0], input[1:] for e in proto: res_e, input = unflatten_helper(input, e) res.append(res_e) return type(proto)(res), input return unflatten_helper(input, proto)[0] </ins> _iter_variables = _iter_filter(lambda o: isinstance(o, torch.autograd.Variable)) _iter_tensors = _iter_filter(torch.is_tensor) _iter_None_tensors = _iter_filter(lambda o: o is None or torch.is_tensor(o)) _map_variable_tensor = _nested_map(lambda o: isinstance(o, torch.autograd.Variable), lambda o: o.data) <del> def _map_tensor_fromiter(itr): return _nested_map(lambda o: torch.is_tensor(o), lambda o: next(itr)) </del> class NestedIOFunction(Function): def _do_forward(self, *input):", "pos": ["cuDNN backend for RNN stabilizes at large mem usage. The following code stabilizes for me at 14+GiB reported by nvidia-smi. The same RNN without cuDNN uses <2.5GiB.\nOops... any idea what's going on with the garbage collector in this case?\nare you running this on a P100 GPU?\nYes, I observe similar behavior on Titan X Pascal\nhmmm, do cudnn RNNs have weird memory-oversubscribing properties? If it wasn't for CuDNN RNN, then it's not unreasonable. we use a custom CUDA allocator that will oversubscribe if needed, but will free the memory blocks it's holding onto if it hits OOM / failed allocations.\nTried the same RNN in torch and am getting ~2.6GiB on nvidia-smi.\nWith cuDNN I'm getting a constant usage of 2241MB on my machine using PyTorch (M40).\nare you using cudnnv5, or v6?\nv5\nno, cudnn does not do memory oversubscribing, and already tried the same network in lua torch with cudnn, where it gets expected memory usage. Also, he is getting OOMs on slightly bigger networks, where apparently caching allocator should have tried to free previous blocks. Interesting that can not repro this.\nJust to make sure did you set when you ran? I left it as False in the repro.\nI can repro in the docker image built from the current version on P100, and I go OOM after 8 batches with the same image on 1080, which generally should not happen - if there is enough memory to go through 1 batch, memory usage should not increase afterwards.\nNo, I just copied your script and ran it. Weird thing is that it goes OOM after 18 batches once I enable . Without it memory remains stable.\nA lot of cudaMalloc allocations go not through THCudaMalloc, and never get freed:\nSo not going through THCudaMalloc is actually expected for storages allocated with THCStorage_(newWithSize), but what's not expected is to every time cudaMalloc bytes of rnn workspace, instead of reusing earlier allocations. (~784 MB allocations are for workspace, ~156MB allocations happen somewhere in criterion).\nMy guess is that the bug is at the Python level.", "Either: Something is holding on to a Tensor longer than it should Something is creating a circular reference which prevents quick clean-up\nThe issue is that in creates cyclic references. It's made worse by the fact that it holds on to more memory than it should (like the workspace). You can see this by adding a as the first line in your for loop. With the explicit call to the GC every iteration, the memory usage doesn't explode.\nhow can these cyclic references be fixed?\nNot sure yet, but I will fix it tomorrow with"], "neg": []}
{"query": "- THTensor* tensor2 ]] <del> #ifndef THP_LAPACK_CONSTANTS #define THP_LAPACK_CONSTANTS </del> <ins> #if defined(TH_REAL_IS_FLOAT) || defined(TH_REAL_IS_DOUBLE) || CUDA_FLOAT || CUDA_DOUBLE </ins> // We need to pass pointers to chars to tensor lapack functions... static const char __U = 'U'; static const char __L = 'L';", "pos": ["Trying to run on github actions via AWS instance gives me: I will of course set up now, but it shouldn't crash like that. May be it should use as a fallback? pt-2.0 cc\nIs this a multitenant setup? Also curious which distro are you running, I was under the impression that most linux distros set home by default\nthis is AWS image. one default user. I think it's user, but it wants you to run as root, which is weird. I have manually now set to overcome this.\nWe should update triton to use pathlib or something"], "neg": []}
{"query": ".. autofunction:: einsum .. autofunction:: flip .. autofunction:: histc <ins> .. autofunction:: meshgrid </ins> .. autofunction:: renorm .. autofunction:: trace .. autofunction:: tril", "pos": ["is not available in the docs page, but the doc string has been in . This was mentioned earlier in", "Oh, bilinear, I missed that last time I looked at it. I'll add a docstring.", "returns nothing. Might be useful to have some linter in the build process to find Python functions missing", "Pytorch 0.4.0 gives a segmentation fault with dilated 2D convolutions over a specific size. import torch import as nn model = nn.Conv2d(inchannels=64, outchannels=64, kernelsize=5, dilation=2) model(([1, 64, 1166, 1166])) # This runs model(([1, 64, 1167, 1167])) # This crashes PyTorch version: 0.4.0 Is debug build: No CUDA used to build PyTorch: 9.1.85 OS: CentOS Linux 7 (Core) GCC version: (GCC) 4.8.5 (Red Hat 4.8.5-28) CMake version: version 2.8.12.2 Python version: 3.6 Is CUDA available: Yes CUDA runtime version: 8.0.61 GPU models and configuration: GPU 0: GeForce GTX 1080 Ti GPU 1: TITAN X (Pascal) GPU 2: Quadro K2200 Nvidia driver version: 390.59 cuDNN version: Probably one of the following: /usr/local/cuda- /usr/local/cuda- /usr/local/cuda-7.5/lib64/libcudnnstatic.a /usr/local/cuda- /usr/local/cuda-8.0/lib64/libcudnn_static.a\ndo you think this is the same as the segfault you showed me a week ago on convolution?\nah yes. int overflow. I only fixed conv3d because I didn\u2019t want to hurt perf. I\u2019ll fix all of them.", "This should be an easy one.. The function is implemented, just no sign of it in the docs."], "neg": []}
{"query": "[ 0.0000,  0.0000,  0.0000,  1.2300]]) \"\"\") <ins> add_docstr_all('scatter_add_', r\"\"\" scatter_add_(dim, index, other) -> Tensor Adds all values from the tensor :attr:`other` into :attr:`self` at the indices specified in the :attr:`index` tensor in a similar fashion as :meth:`~torch.Tensor.scatter_`. For each value in :attr:`other`, it is added to an index in :attr:`self` which is specified by its index in :attr:`other` for ``dimension != dim`` and by the corresponding value in :attr:`index` for ``dimension = dim``. For a 3-D tensor, :attr:`self` is updated as:: self[index[i][j][k]][j][k] += other[i][j][k]  # if dim == 0 self[i][index[i][j][k]][k] += other[i][j][k]  # if dim == 1 self[i][j][index[i][j][k]] += other[i][j][k]  # if dim == 2 :attr:`self`, :attr:`index` and :attr:`other` should have same number of dimensions. It is also required that ``index.size(d) <= other.size(d)`` for all dimensions ``d``, and that ``index.size(d) <= self.size(d)`` for all dimensions ``d != dim``. Moreover, as for :meth:`~Tensor.gather`, the values of :attr:`index` must be between ``0`` and ``self.size(dim) - 1`` inclusive, and all values in a row along the specified dimension :attr:`dim` must be unique. Args: dim (int): the axis along which to index index (LongTensor): the indices of elements to scatter and add other (Tensor): the source elements to scatter and add Example:: >>> x = torch.rand(2, 5) >>> x tensor([[0.7404, 0.0427, 0.6480, 0.3806, 0.8328], [0.7953, 0.2009, 0.9154, 0.6782, 0.9620]]) >>> torch.ones(3, 5).scatter_add_(0, torch.tensor([[0, 1, 2, 0, 0], [2, 0, 0, 1, 2]]), x) tensor([[1.7404, 1.2009, 1.9154, 1.3806, 1.8328], [1.0000, 1.0427, 1.0000, 1.6782, 1.0000], [1.7953, 1.0000, 1.6480, 1.0000, 1.9620]]) \"\"\") </ins> add_docstr_all('select', r\"\"\" select(dim, index) -> Tensor", "pos": ["is not available in the docs page, but the doc string has been in . This was mentioned earlier in", "Oh, bilinear, I missed that last time I looked at it. I'll add a docstring.", "returns nothing. Might be useful to have some linter in the build process to find Python functions missing", "Pytorch 0.4.0 gives a segmentation fault with dilated 2D convolutions over a specific size. import torch import as nn model = nn.Conv2d(inchannels=64, outchannels=64, kernelsize=5, dilation=2) model(([1, 64, 1166, 1166])) # This runs model(([1, 64, 1167, 1167])) # This crashes PyTorch version: 0.4.0 Is debug build: No CUDA used to build PyTorch: 9.1.85 OS: CentOS Linux 7 (Core) GCC version: (GCC) 4.8.5 (Red Hat 4.8.5-28) CMake version: version 2.8.12.2 Python version: 3.6 Is CUDA available: Yes CUDA runtime version: 8.0.61 GPU models and configuration: GPU 0: GeForce GTX 1080 Ti GPU 1: TITAN X (Pascal) GPU 2: Quadro K2200 Nvidia driver version: 390.59 cuDNN version: Probably one of the following: /usr/local/cuda- /usr/local/cuda- /usr/local/cuda-7.5/lib64/libcudnnstatic.a /usr/local/cuda- /usr/local/cuda-8.0/lib64/libcudnn_static.a\ndo you think this is the same as the segfault you showed me a week ago on convolution?\nah yes. int overflow. I only fixed conv3d because I didn\u2019t want to hurt perf. I\u2019ll fix all of them.", "This should be an easy one.. The function is implemented, just no sign of it in the docs."], "neg": []}
{"query": "} } <del> if (use_cudnn) { </del> <ins> if (use_cudnn && eps >= CUDNN_BN_MIN_EPSILON) { </ins> #ifdef WITH_CUDNN torch::cudnn::cudnn_batch_norm_backward( state,", "pos": ["Hi, I have two linux servers. My code works in one machine but not the other. In the one that it doesn't work, the error message is: occured during batch normalization( convolution is fine). Also, the same code works in CPU but not in GPU. I guess this problem is mainly related to GPU libraries like cuda or cudnn. The one that doesn't work has the following configurations: Nvidia driver 375.26 cuda 8.0.61 cudnn 5.1.10 While in the one it does work, I have the following configurations: Nvidia driver 367.48 cuda 8.0.44 cudnn 5.1.5 TensorFlow can run in the above environment flawless though. So I am wondering if this is caused by the incompatibility between PyTorch 0.1.12 and cuda 8.0.61/cudnn 5.1.10 ?\nyou use the pytorch binaries, we ship cudnn6 in them, use it it's great. 5.1.10 maybe has a bug. cc: . do you have a small test case to reproduce this? does not use CuDNN batchnorm by default, so maybe that's why?\nSmall test case would help, thanks.\nAfter some experimentation, I have pin pointed where exactly the error comes from. It seems batch norm in cudnn 5.1.10 doesn't support eps as small as 1e-6. A small test case would be:\nYes, cudnn.h for 5.1.10 defines CUDNNBNEPSILON to be 1e-5 (minimum allowable value). I can't check header for 5.1.5. Even with 1e-5 epsilon for many input data distributions the results become pretty unstable. Is there a case for this restriction to be lifted, does using very small epsilon provide benefits?\nI see. Not sure if small epsilon is necessary. I am using 1e-6 just to replicate others' configurations. Should PyTorch provide more informative error message about this restriction?\nIdeally, CUDNN should provide more informative error messages (pytorch can not know why cudnn is returning an error code in each particular case) , but that's a long standing discussion.\nIt just hits on me.", "As mentioned by , TensorFlow does not use CuDNN batchnorm by default. So is this possible in Pytorch, to not use default CuDNN batchnorm?\nNever mind. I found the switch in ."], "neg": []}
{"query": "return ret; } <del> Variable VariableType::as_variable(Tensor tensor) const { </del> <ins> static Variable as_variable(Tensor tensor) { </ins> return make_variable(std::move(tensor)); } <del> std::tuple<Variable, Variable> VariableType::as_variable(std::tuple<Tensor, Tensor> tensors) const { </del> <ins> static std::tuple<Variable, Variable> as_variable(std::tuple<Tensor, Tensor> tensors) { </ins> return std::make_tuple<>( make_variable(std::move(std::get<0>(tensors))), make_variable(std::move(std::get<1>(tensors)))); } <del> std::tuple<Variable, Variable, Variable> VariableType::as_variable(std::tuple<Tensor, Tensor, Tensor> tensors) const { </del> <ins> static std::tuple<Variable, Variable, Variable> as_variable(std::tuple<Tensor, Tensor, Tensor> tensors) { </ins> return std::make_tuple<>( make_variable(std::move(std::get<0>(tensors))), make_variable(std::move(std::get<1>(tensors))), make_variable(std::move(std::get<2>(tensors)))); } <del> std::tuple<Variable, Variable, Variable, Variable> VariableType::as_variable(std::tuple<Tensor, Tensor, Tensor, Tensor> tensors) const { </del> <ins> static std::tuple<Variable, Variable, Variable, Variable> as_variable(std::tuple<Tensor, Tensor, Tensor, Tensor> tensors) { </ins> return std::make_tuple<>( make_variable(std::move(std::get<0>(tensors))), make_variable(std::move(std::get<1>(tensors))),", "pos": ["See Docker run fails with , which is due to the fact that is specified in docker container: Which results in: CI cc\nCheck that adding to manywheels image results in a failure:\nTentatively marking hi-pri since nightly testing are failing\nThis is resolved, here is the successful run:"], "neg": []}
{"query": "self.assertRaises(TypeError, lambda: reference[0.0, ..., 0.0:2.0]) self.assertRaises(TypeError, lambda: reference[0.0, :, 0.0]) <ins> def test_index(self): self._test_index(self, lambda x: x) </ins> @staticmethod def _test_advancedindex(self, conv_fn): # Tests for Integer Array Indexing, Part I - Purely integer array", "pos": ["Arguably, indexing a Tensor by a LongTensor with no dimension should return a LongTensor with no dimension. Consider the following case: This breaks indexing by indices when the list of indices is empty, which can be hard to detect. In my use-case, I use such a masking to select rows of a tensor corresponding to \"valid\" entries in a vector; I had not considered that it would not work when there are no valid entries.\nThanks! I'll take a look."], "neg": []}
{"query": "\"\"\" # TODO: try to find an input on another gpu, copy it, # and accumulate into the copy <ins> if destination is None: destination = torch.cuda.current_device() </ins> input_size = inputs[0].size() <ins> nccl_root = None </ins> for i, inp in enumerate(inputs): assert inp.is_cuda, \"reduce_add expects all inputs to be on GPUs\" <ins> if inp.get_device() == destination: nccl_root = i </ins> if inp.size() != input_size: got = 'x'.join(str(x) for x in inp.size()) expected = 'x'.join(str(x) for x in input_size) raise ValueError(\"input {} has invalid size: got {}, but expected \" \"{}\".format(i, got, expected)) <del> if destination is None: destination = torch.cuda.current_device() </del> <ins> assert nccl_root is not None, \"reduce_add expects destination to be on the same GPU with one of the tensors\" </ins> with torch.cuda.device(destination): result = type(inp)(input_size).zero_() if nccl.is_available(inputs) and inputs[0].get_device() == destination: outputs = [result] + [t.new(t.size()) for t in inputs[1:]] <del> nccl.reduce(inputs, outputs, root=destination) </del> <ins> nccl.reduce(inputs, outputs, root=nccl_root) </ins> return result for inp in inputs:", "pos": ["The loss will never decrease when I train a model using multi-gpus but without gpu0. This is an example I modified from the : If the result after one epoch is If the result after one epoch is\nshould always contain . If you want to specify which GPUs to use, you can set the environment variable , either by running your script like this: or set it programmatically in your script like this: So that deviceids 0, 1 will map to GPUs 1, 2, respectively. (Note that for the second method, you should put that line before any cuda relevant function call.)\nThank you.\n<!-- drci-comment-start --:pagefacingup: Preview :pagefacingup: Preview :question: Need help or want to give feedback on the CI? Visit our Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. <!-- drci-comment-end --"], "neg": []}
{"query": ":members: <ins> Padding Layers -------------- :hidden:`ReflectionPad2d` ~~~~~~~~~~~~~~~~~~~~~~~~~ .. autoclass:: ReflectionPad2d :members: :hidden:`ReplicationPad2d` ~~~~~~~~~~~~~~~~~~~~~~~~~~ .. autoclass:: ReplicationPad2d :members: :hidden:`ReplicationPad3d` ~~~~~~~~~~~~~~~~~~~~~~~~~~ .. autoclass:: ReplicationPad3d :members: :hidden:`ZeroPad2d` ~~~~~~~~~~~~~~~~~~~ .. autoclass:: ZeroPad2d :members: :hidden:`ConstantPad2d` ~~~~~~~~~~~~~~~~~~~~~~~ .. autoclass:: ConstantPad2d :members: </ins> Non-linear Activations ----------------------------------", "pos": ["After merging cuda 12.4 for nightly. I observe following failures: libtorch-cuda124-shared-with-deps-cxx11-abi-build: Build Official Docker Images 124: We need to resolve these before continuing adding CUDA 12.4 CI workflows 0 cc\nattempt 3 worked for libtorch build issue:\nFor Docker image, related discussion:\nI will instance a new bug for the docker job failure.", "Is there a way to use reflection padding instead of zero padding with nn.Conv2d ? I see there is a but the option is not in nn.Conv2d. Thanks,\nare in , but they're not exposed in the docs.\nThanks!"], "neg": []}
{"query": "self.assertEqual(inp_i.grad, i.grad) assert not hasattr(TestAutograd, test_name), 'Two tests have the same name: ' + test_name <ins> for skip in skipTestIf: do_test = skip(do_test) </ins> setattr(TestAutograd, test_name, do_test)", "pos": ["Command to install the latest nightly binary: installs : However, PyTorch itself reports : I guess this version tag might be defined during the build process from e.g. and then forwarded from , but also source builds should use based on the tag. Unsure, if this is just a reporting issue or if we are building the nightly binaries with and executing them with , so creating this issue to dig a bit further into it. CC cc\nThis is really weird, I've put up measure in place to validate against it... Will investigate today (and moving to pytorch repo)\ncan you please check that 2.2.0 reports (and is build) against the correct version of nccl library?\nConfirmed for nightlies : Same issue for RC:\nLooking at the build log Looks like we are compiling with: USESYSTEMNCCL\nOn Windows I see following exception: We compiling with should we return None rather then exception ?\nIn this case we are indeed using 2.18.1: Do you know why we are using the system-wide NCCL installation instead of the one? Should we delete the system NCCL and use by default?\nLooks like its set here: Looks like it was by: Let me propose the PR to unset this flag : USESYSTEM_NCCL\nI guess we are using the system NCCL installation to be able to during the runtime. If we would use the module I assume we would link NCCL into the PyTorch binaries. If so, we should make sure to update the NCCL version whenever is updated. This could break our dependency on the NCCL pip package.\nYes. Also when we update this nccl here:\nWe used following path  What would be good path for nccl 2.19.3 and cuda 12.1 ? As per this issue: Will be using 12.0 ncck 2.19.3\nInstalling system NCCL into the nightly build docker container feels great, but we need to figure out, how to keep it in sync with PyTorch. I wonder if one can, for example, just check nccl version from this requirements file and build from source?\nyes probably we should build and upload nccl for our supported cuda versions, same way as we build docker containers, by following details here:\nVerified with final rc"], "neg": []}
{"query": "return r; } <del> variable_list flatten(const Tensor& x, const TensorList& y, const Tensor& z) { </del> <ins> static variable_list flatten(const Tensor& x, const TensorList& y, const Tensor& z) { </ins> std::vector<Variable> r; r.reserve(2 + y.size()); r.emplace_back(x);", "pos": ["See Docker run fails with , which is due to the fact that is specified in docker container: Which results in: CI cc\nCheck that adding to manywheels image results in a failure:\nTentatively marking hi-pri since nightly testing are failing\nThis is resolved, here is the successful run:"], "neg": []}
{"query": "arguments(), is_variadic(is_variadic), has_out(has_out) {}; Option(const Option&) = delete; Option(Option&& other): <del> arguments(std::move(other.arguments)), is_variadic(other.is_variadic) {}; </del> <ins> arguments(std::move(other.arguments)), is_variadic(other.is_variadic), has_out(other.has_out) {}; </ins> std::vector<Argument> arguments; bool is_variadic;", "pos": ["! Maybe we should change it to `? No response cc\nare these old installation commands also auto-tested with CI?\nLooks no longer relevant."], "neg": []}
{"query": "auto m = self.size(0); auto n = self.size(1); auto k = sigma.size(0); <ins> auto gsigma = grads[1]; auto u = raw_u; auto v = raw_v; auto gu = grads[0]; auto gv = grads[2]; </ins> <del> Tensor u, v; </del> if (!some) { <del> // ignore the free subspace </del> <ins> // We ignore the free subspace here because possible base vectors cancel // each other, e.g., both -v and +v are valid base for a dimension. // Don't assume behavior of any particular implementation of svd. </ins> u = raw_u.narrow(1, 0, k); v = raw_v.narrow(1, 0, k); <del> } else { u = raw_u; v = raw_v; </del> <ins> if (gu.defined()) { gu = gu.narrow(1, 0, k); } if (gv.defined()) { gv = gv.narrow(1, 0, k); } </ins> } <del> auto gu = grads[0]; auto gsigma = grads[1]; auto gv = grads[2]; </del> auto vt = v.t(); Tensor sigma_term;", "pos": ["I am having problem in running the backward propagation with svd. In particular, it seems working when the first dimension of the matrix x (below) is greater than the second. But I get the following error if the first is smaller than the second. PyTorch version: 2018.04.20 Is debug build: No CUDA used to build PyTorch: 8.0.61 OS: Ubuntu 16.04.4 LTS GCC version: (Ubuntu 5.4.0-6ubuntu1~16.04.9) 5.4.0 CMake version: Could not collect Python version: 3.5 Is CUDA available: No CUDA runtime version: No CUDA GPU models and configuration: No CUDA Nvidia driver version: No CUDA cuDNN version: No CUDA Versions of relevant libraries: [pip3] numpy (1.13.3) [pip3] torch (2018.4.20) [pip3] torchvision (0.2.0) [conda] pytorch 0.3.1 py35cuda8.0.61cudnn7.0.52 pytorch [conda] pytorch-nightly 2018.04.20 py35cuda8.0.61cudnn7.1.21 pytorch [conda] torchvision 0.2.0 py35heaa392f_1 pytorch\nRepro'ed on 0.4 branch. works btw.\nWhat does \"Repro'ed\" means? Should I install from that branch? I need to compute the kernel of a matrix, therefore I need the last columns of the matrix V (which is why I have to pass ).\nIt's a shorthand for \"reproduced\", which means that the bug still hasn't been fixed. Sorry, we'll try to get to it soon."], "neg": []}
{"query": "raise ValueError(\"output_size should be a sequence containing \" \"2 or 4 elements, but it has a length of {}\".format( len(output_size))) <del> out_sizew, out_sizeh = output_size sizew = ((input.size(3) - 1) * self.dw - 2 * self.padw + self.kw) </del> <ins> out_sizeh, out_sizew = output_size </ins> sizeh = ((input.size(2) - 1) * self.dh - 2 * self.padh + self.kh) <del> out_padw = out_sizew - sizew </del> <ins> sizew = ((input.size(3) - 1) * self.dw - 2 * self.padw + self.kw) </ins> out_padh = out_sizeh - sizeh <del> out_padw_ok = 0 <= out_padw < self.dw </del> <ins> out_padw = out_sizew - sizew </ins> out_padh_ok = 0 <= out_padh < self.dh <del> if not out_padw_ok or not out_padh_ok: </del> <ins> out_padw_ok = 0 <= out_padw < self.dw if not out_padh_ok or not out_padw_ok: </ins> raise ValueError((\"requested an output size of {}x{}, but \" \"valid sizes range from {}x{} to {}x{} (for an input of \" \"{}x{})\").format(out_sizeh, out_sizew, sizeh, sizew,", "pos": ["Repro is torchtrain w/ it happens more often if in my experience but it still happens either way. Logically its clear we have a race: rank0 is free to exit and shut down TCPStore before or after other ranks shut themselves down. we have to consider either a) rank 0 waits for all other ranks to exit before it exists b) we make other ranks tolerant of tcpstore going away (A) has potential downsides in efficiency even in the best case. But also practically, we could use tcpstore barrier, but it wont work reliably at large scales. we'd have to consider something like 'scalable coordinator' to go this route (B) we could probably implement w/ a try/catch in the heartbeat monitor for the connection error, similar to how we catch cuda driver exit. This feels a little hacky. Its important we somehow distinguish between the cases when we are actually shutting down and want to swallow the error, from cases where the is a tcpstore/network error during runtime and we actually want to raise the error. This points to (a) as a more correct direction. cc\ndo you have any update on \"make torchrun share TCPStore instance with pytorch\"? I think that is our \"proper\" fix for this problem, and in the meantime people can work around it by 1) sleeping on rank0 for a sufficiently long time (between 5-10s seemed plenty for 128rank jobs) simplest possible to not sleep long enough in edge cases 2) using a barrier at job shutdown such that rank0 (and tcsptore) are gauranteed alive.\nwas landed as a fix"], "neg": []}
{"query": "self.assertIn('buf', l.state_dict()) self.assertIs(l.state_dict()['buf'], buf) <ins> def test_Conv2d_inconsistent_types(self): inputs = Variable(torch.randn(4, 1, 7, 7).float()) weights = Variable(torch.randn(1, 1, 3, 3).double()) # inconsistent types should raise an exception self.assertRaises(RuntimeError, lambda: nn.functional.conv2d(inputs, weights)) # but it should work with the same type nn.functional.conv2d(inputs.float(), weights.float()) </ins> @unittest.skipIf(not TEST_CUDA, 'CUDA not available') def test_Conv2d_large_workspace(self): # These sizes require huge cuDNN workspaces. Make sure we choose a", "pos": ["the error only happens in very specific conditions. (change the view of input tensor and use weight converted from numpy array.) The output of functional conv2d is far off in scale (e12 or so.)\nthis happens because generates double precision tensor by default, whereas torch generates single precision. We need to add guards for tensor types in THNN and THCUNN.\nIt's a problem with function, that only . Should be a simple fix.\nWhoops! Yeah it just ignores the \"thpp::Type type\" parameter\nwhat type is expected here?\na is a FloatTensor but weights is a DoubleTensor. Change: to: (or change a to (...).double())\ncool thanks!"], "neg": []}
{"query": "THArgCheckWithCleanup(n_sample > 0, THCleanup(if (start_dim == 1) THTensor_(resize1d)(prob_dist, n_categories);), 2, <del> \"cannot sample n_sample < 0 samples\"); </del> <ins> \"cannot sample n_sample <= 0 samples\"); </ins> if (!with_replacement) {", "pos": ["Looking at the values returned from tn.multinomial(1) and counting them, it seems like it returns the values of the multinomial distribution taken from t, i.e. that it has ignored the signs of the values. However, it seem like always returns a tensor filled with zeros.\nlooking into this"], "neg": []}
{"query": "def in_cooldown(self): return self.cooldown_counter > 0 <del> def _init_is_better(self, mode, threshold, threshold_mode): if mode not in {'min', 'max'}: raise ValueError('mode ' + mode + ' is unknown!') if threshold_mode not in {'rel', 'abs'}: raise ValueError('threshold mode ' + threshold_mode + ' is unknown!') </del> <ins> def _cmp(self, mode, threshold_mode, threshold, a, best): </ins> if mode == 'min' and threshold_mode == 'rel': rel_epsilon = 1. - threshold <del> self.is_better = lambda a, best: a < best * rel_epsilon self.mode_worse = float('Inf') </del> <ins> return a < best * rel_epsilon </ins> elif mode == 'min' and threshold_mode == 'abs': <del> self.is_better = lambda a, best: a < best - threshold self.mode_worse = float('Inf') </del> <ins> return a < best - threshold </ins> elif mode == 'max' and threshold_mode == 'rel': rel_epsilon = threshold + 1. <del> self.is_better = lambda a, best: a > best * rel_epsilon self.mode_worse = -float('Inf') </del> <ins> return a > best * rel_epsilon </ins> else:  # mode == 'max' and epsilon_mode == 'abs': <del> self.is_better = lambda a, best: a > best + threshold self.mode_worse = -float('Inf') </del> <ins> return a > best + threshold def _init_is_better(self, mode, threshold, threshold_mode): if mode not in {'min', 'max'}: raise ValueError('mode ' + mode + ' is unknown!') if threshold_mode not in {'rel', 'abs'}: raise ValueError('threshold mode ' + threshold_mode + ' is unknown!') if mode == 'min': self.mode_worse = float('inf') else:  # mode == 'max': self.mode_worse = (-float('inf')) self.is_better = partial(self._cmp, mode, threshold_mode, threshold) </ins>", "pos": ["I'm working on checkpointing in our nn trainer. When saving the state objects, I get an error: PicklingError: Can't pickle <type 'function': attribute lookup builtin.function failed I tracked it down to trying to save the scheduler ReduceLROnPlateau. I think the issue is due to the scheduler referencing the optimizer, which has to be saved with statedict(). I presume the same error will occur with other schedulers too. It seems like the problem could be solved by providing a statedict() and loadstatedict() in the scheduler classes. pytorch master rev (Jan 22) installed from source x86_64 on centos 7.4 cuda 9.0 cudnn 7 k80 gcc 5.4 python 2.7.14\nThe following script reproduces the crash: import torch import from import ReduceLROnPlateau if name == 'main': model = (10,10) optim = (model.parameters()) sched = ReduceLROnPlateau(optim, mode='max', factor=0.5) with open('','wb') as f: # Works # ((model.statedict(), optim.statedict()), f) # Crashes ((model.statedict(), optim.state_dict(), sched), f)"], "neg": []}
{"query": "check_file(os.path.join(lib_path, \"pybind11\", \"CMakeLists.txt\")) check_file(os.path.join('aten', 'src', 'ATen', 'cpu', 'cpuinfo', 'CMakeLists.txt')) check_file(os.path.join('aten', 'src', 'ATen', 'cpu', 'tbb', 'tbb_remote', 'Makefile')) <ins> check_file(os.path.join('aten', 'src', 'ATen', 'utils', 'catch', 'CMakeLists.txt')) </ins> check_pydep('yaml', 'pyyaml') check_pydep('typing', 'typing')", "pos": ["Trying to build master (), I'm seeing I tried in to no avail.\nNM, also needed a .\nNeed a check to see if this is folder is checked out in , like the other submodules. Should be a simple patch."], "neg": []}
{"query": "r__->stride[(transpose_r == 'n' ? 1 : 0)]); /* free intermediate variables */ <del> if(m1_ != m1) </del> <ins> if(free_m1) </ins> THTensor_(free)(m1_); <del> if(m2_ != m2) </del> <ins> if(free_m2) </ins> THTensor_(free)(m2_); if(r__ != r_)", "pos": ["Here is an example of what seems to be a memory leak in pytorch, although it could be numpy? The leak was only occurring when using the or array unsqueezing methods in numpy. Also worth noting that the issue is resolved by removing the from and casting the torch tensor to type. An easy workaround is to reshape the array after turning it into a torch Tensor."], "neg": []}
{"query": "#   method name, #   input size/constructing fn, #   args (tuple represents shape of a tensor arg), <del> #   test variant name (will be used at test name suffix),  // optional #   indices for possible dim arg,                          // optional #   output indices that should be gradcheck'ed,            // optional </del> <ins> #   test variant name (will be used at test name suffix),    // optional #   indices for possible dim arg,                            // optional #   fn mapping output to part that should be gradcheck'ed,   // optional </ins> # ) method_tests = [ ('add', (S, S, S), ((S, S, S),)),", "pos": ["I am having problem in running the backward propagation with svd. In particular, it seems working when the first dimension of the matrix x (below) is greater than the second. But I get the following error if the first is smaller than the second. PyTorch version: 2018.04.20 Is debug build: No CUDA used to build PyTorch: 8.0.61 OS: Ubuntu 16.04.4 LTS GCC version: (Ubuntu 5.4.0-6ubuntu1~16.04.9) 5.4.0 CMake version: Could not collect Python version: 3.5 Is CUDA available: No CUDA runtime version: No CUDA GPU models and configuration: No CUDA Nvidia driver version: No CUDA cuDNN version: No CUDA Versions of relevant libraries: [pip3] numpy (1.13.3) [pip3] torch (2018.4.20) [pip3] torchvision (0.2.0) [conda] pytorch 0.3.1 py35cuda8.0.61cudnn7.0.52 pytorch [conda] pytorch-nightly 2018.04.20 py35cuda8.0.61cudnn7.1.21 pytorch [conda] torchvision 0.2.0 py35heaa392f_1 pytorch\nRepro'ed on 0.4 branch. works btw.\nWhat does \"Repro'ed\" means? Should I install from that branch? I need to compute the kernel of a matrix, therefore I need the last columns of the matrix V (which is why I have to pass ).\nIt's a shorthand for \"reproduced\", which means that the bug still hasn't been fixed. Sorry, we'll try to get to it soon."], "neg": []}
{"query": "TensorTriOp(T *start_, int64_t stride0_, int64_t stride1_, int64_t k_) : start(start_), stride0(stride0_), stride1(stride1_), k(k_) {} <del> __device__ __forceinline__ int mask(T *in) { ptrdiff_t n = in - start; </del> <ins> __device__ __forceinline__ int mask(T *out) { ptrdiff_t n = out - start; </ins> int64_t row, col; if (stride0 > stride1) {", "pos": ["On recent PyTorch master (built from source; 0.4.0a0+) on a P100, CUDA 9:\nI think the results on CPU meets the defination of triu (upper triangular part of the matrix). It is really hard to handle with a tensor from expand operation with multi-thread not only on GPU but also on CPU. The unexpected result will also happen on CPU if we apply openmp to it. I think it is necessary to eleminate the difference between the implementation and defination.\nIt doesn\u2019t matter if it\u2019s cpu or not. triu doesn\u2019t have an underscore which means it doesn\u2019t operate in-place and should always give the same result (in new memory)"], "neg": []}
{"query": "LDFLAGS=\"-L$INSTALL_DIR/lib \" LD_POSTFIX=\".so.1\" LD_POSTFIX_UNVERSIONED=\".so\" <del> if [[ $(uname) == 'Darwin' ]]; then </del> <ins> if [[ $(uname) == 'Darwin' ]]; then </ins> LDFLAGS=\"$LDFLAGS -Qunused-arguments -Wl,-rpath,@loader_path\" LD_POSTFIX=\".1.dylib\" LD_POSTFIX_UNVERSIONED=\".dylib\"", "pos": ["there is a bug. the code is from The code for installing cuda10.2 should be corrected to The code for installing cuda10.2 should be corrected to cc\nThank you for reporting the typo", "This is how I implement the decoder of a sequence to sequence model I'm not sure about the new autograd mechanics but this worked in the previous version. If I didn't make the unrolling codes a function it will work. It will also work with CPU. I compiled from source with Python 2.7, CUDA 8.0 and Cudnn 6\nThanks for the minimal reproduction! It broke OpenNMT and a big internal model but we weren't looking forward to isolating what parts of those had the issue.\nThat's great, I'll take a look today. Thanks!\nI've pushed a fix to .\nHow does the inference work in this case? I mean, once you train it how do you test it in the test phase?"], "neg": []}
{"query": "return t(*sizes).copy_(torch.randn(*sizes)) return tmp <ins> # Content of each tuple: # - function name # - constructor for the tensor,    signature: fn(tensor_type) -> tensor # - constructor for the arguments, signature: fn(tensor_type) -> list # - postfix name for the test (must be unique for a given function) (default='') # - tensor types to use (default=types) # - disable inplace test, if set to True, no inplace test will be done (default=False) </ins> tests = [ ('add', small_3d, lambda t: [number(3.14, 3, t)]), ('add', small_3d, lambda t: [small_3d_positive(t)], 'tensor'),", "pos": ["On recent PyTorch master (built from source; 0.4.0a0+) on a P100, CUDA 9:\nI think the results on CPU meets the defination of triu (upper triangular part of the matrix). It is really hard to handle with a tensor from expand operation with multi-thread not only on GPU but also on CPU. The unexpected result will also happen on CPU if we apply openmp to it. I think it is necessary to eleminate the difference between the implementation and defination.\nIt doesn\u2019t matter if it\u2019s cpu or not. triu doesn\u2019t have an underscore which means it doesn\u2019t operate in-place and should always give the same result (in new memory)"], "neg": []}
{"query": "} void THCTensor_(sparseMask)(THCState *state, THCSTensor *r_, THCTensor *t, THCSTensor *mask) { <ins> THArgCheck(mask->coalesced, 2, \"mask is uncoalesced\"); </ins> THCAssertSameGPU(THCSTensor_(checkGPU)(state, 2, 3, r_, mask, t)); if(!THCSTensor_(isSameSizeAsDense)(state, mask, t)) { THError(\"sparseMask operands have incompatible sizes\");", "pos": ["and I talked about the intended semantics of sparsemask when given a sparse tensor as an argument, and we came to the following conclusions: with an uncoalesced sparse tensor does not make sense, full stop. Unlike values, where your update might be linear and work correctly on an uncoalesced sparse tensor, there isn't any good reason why you want to pluck out values from the dense tensor multiple times, it is basically always a bug. the (eventual) documentation for sparsemask, we will recommend calling coalesce before doing any operations involving sparsemask (or nonlinear updates to indices/values). should add a test to sparsemask which checks if the flag of the input tensor is true, and raise an error if it is not. This is a cheap safety check which will prevent people from forgetting to call coalesce (as in (2)). There is a performance consequence to calling coalesce every iteration of training. If the gradient is producing coalesced vectors, we can alleviate this problem by adding a preprocessing step to coalesce() which tests if the input is already coalesced, and bugs out in that case. However, I think this is out of scope for this particular ticket for now."], "neg": []}
{"query": "(Addcdiv,       (),                 ((S, S), (S, S), torch.rand(S, S) + 1e-2)   ), (Addcdiv,       (0.6,),             ((S, S), (S, S), torch.rand(S, S) + 1e-2), 'scale'), (IndexAdd,      (0,),               ((S, S), index_variable(2, S), (2, S))      ), <del> (IndexCopy,     (0,),               ((S, S), index_variable(2, S), (2, S))      ), </del> <ins> # (IndexCopy,     (0,),               ((S, S), index_variable(2, S), (2, S))      ), </ins> (IndexFill,     (0, 2),             ((S, S), index_variable(2, S))              ), (IndexSelect,   (0,),               ((S, S), index_variable(2, S))              ), (Gather,        (0,),               ((M, S), gather_variable((S, S), 1, M))     ),", "pos": ["This happens when the index tensor contains duplicate elements. If this is not allowed indexcopy should raise an exception when this happens and we should fix the test. Otherwise we need to fix the autograd op. Here's an example: cc\nWhy would you with repeated indices? Sounds for me like UB is a reasonable. We should check that in C\nI ran into this in and fixed the test for now. I'm not sure how to interpret last comment: did you mean UB behavior is \"unreasonable\" rather than \"a reasonable\"? The last sentence reads that way.\nI'd be ok with saying that it's UB in such case. There's no natural choice, and it can decrease performance.\nIs this issue still worth working on? If adding the additional check can decrease performance, can we add another argument specifying that indexes contain duplicate elements?"], "neg": []}
{"query": "#include <Python.h> #include \"THP.h\" <ins> #include \"torch/csrc/utils/auto_gil.h\" </ins> // Adapted from fblualib void* ObjectPtrAllocator::malloc(ptrdiff_t size) {", "pos": ["When both PyTorch and Pandas are imported, calling results in a segfault. The issue was originally reported by It looks like calling returns a null pointer here: Examining in GDB: I suspect it's both PyTorch and Pandas doing threading magic and somehow getting in each other's way. I haven't yet been able to provide a minimal example that triggers this behaviour and does not use my package; it's likely that I'm doing something odd that causes this. More documentation here: Possibly related issues: and Python version: 3.6 PyTorch version: v0.2.0 Pandas version: 0.20.1"], "neg": []}
{"query": "#include <vector> #include <string> <ins> #include \"torch/csrc/utils/object_ptr.h\" </ins> #define THPUtils_(NAME) TH_CONCAT_4(THP,Real,Utils_,NAME) #define THPUtils_typename(obj) (Py_TYPE(obj)->tp_name)", "pos": ["Go to Look at the html. It contains the following: ! This problem manifests in google search. Searching for 1.8.1 docs brings up the \"1.8.0 docs\": ! We can probably fix this by doing a quick sed on the website but we should develop a solution for future releases. I think this has happened before in the past and am not sure what causes it. cc"], "neg": []}
{"query": "name: binary_linux_manywheel_3_7m_cu102_devtoolset7_test requires: - binary_linux_manywheel_3_7m_cu102_devtoolset7_build <del> resource_class: gpu.medium </del> <ins> resource_class: gpu.nvidia.small </ins> use_cuda_docker_runtime: \"1\" - binary_linux_test: build_environment: libtorch 3.7m cpu devtoolset7", "pos": ["As it still using deprecated resource class, see See example of the failure in\nFollowup: we need to plan better --we knew it was gonna happen and none of us fixed it early enough. Action item: go through tasks about deprecation in sprint planning"], "neg": []}
{"query": "optimizer (Optimizer): Wrapped optimizer. step_size (int): Period of learning rate decay. gamma (float): Multiplicative factor of learning rate decay. <del> Default: -0.1. </del> <ins> Default: 0.1. </ins> last_epoch (int): The index of last epoch. Default: -1. Example:", "pos": ["as shown in source code, this parameters is not used. Any reason that we keep it?\nthis can be cleaned up."], "neg": []}
{"query": "from .serialization import save, load <ins> from ._tensor_str import set_printoptions </ins> ################################################################################ # Define Storage and Tensor classes", "pos": ["It seems that the following link that can be reached from here (check the screenshot) is broken: Page: Broken link: Screenshot: ! cc\nthis should be transferred to , but I don't have permissions to do that. Fix is available at pytorch/hub."], "neg": []}
{"query": "\"torch/csrc/Generator.cpp\", \"torch/csrc/Tensor.cpp\", \"torch/csrc/Storage.cpp\", <ins> \"torch/csrc/byte_order.cpp\", </ins> \"torch/csrc/utils.cpp\", \"torch/csrc/allocators.cpp\", \"torch/csrc/serialization.cpp\",", "pos": ["<!-- A clear and concise description of what the bug is. --The target name is used by a test executable in mkldnn. Steps to reproduce the behavior: BUILDTEST=1 build <!-- If you have a code sample, error messages, stack traces, please provide it here as well --<!-- A clear and concise description of what you expected to happen. --CMake configure passes. Please copy and paste the output from our (or fill out the checklist below manually). You can get the script and run it with: PyTorch Version (e.g., 1.0): v1.5.0-rc2 OS (e.g., Linux): Windows How you installed PyTorch (, , source): source Build command you used (if compiling from source): python build Python version: 3.x CUDA/cuDNN version: 9.2 / 10.1 / 10.2 / NA GPU models and configuration: Tesla M60 Any other relevant information: <!-- Add any other context about the problem here. --We should not build tests for mkldnn.\nI have encountered the same error, how do I resolve it?\nIt is actually mkldnn that is out of sync. So running the following script should help.\nThanks a lot for this I posted this on the PyTorch Discussions but to no avail, and now I've finally understood the reason of this error. Thanks a lot. Surprisingly, for me the build works out if I build from the master repo, but changing to a different version always fails."], "neg": []}
{"query": "grouped_args += [group(t) for t in args] res.append(impl[fn_name](self, self._bufs[g], *grouped_args)) if fn_name == 'grad_params': <del> return [torch.cat(t, 0) for t in zip(*res)] </del> <ins> return [torch.cat(t, 0) if t[0] is not None else None for t in zip(*res)] </ins> else: return torch.cat(res, 1)", "pos": ["x = Variable(((16, 64, 8, 8))) res = nn.Conv2d(16, 16, kernelsize=3, groups=2, bias=False)(x) results in: return ('updateoutput', input, weight, bias) File \"\", line 158, in thnn groupedargs += [group(t) for t in args] File \"\", line 153, in group dim = 0 if () == 1 else 1 AttributeError: 'NoneType' object has no attribute 'dim' The in line 158 is because there is no bias. I am running on CPU so this is in the code."], "neg": []}
{"query": "self.assertEqual(counter[0], 1, 'bw_hook not called') self.assertEqual(x.grad.data, torch.ones(5, 5) * 2) <del> @unittest.skipIf(sys.version_info[0] == 2, \"Python 2 doesn't collect cycles involving __del__\") def test_hooks_cycle(self): import gc counter = [0] class GradHook(object): def __init__(self, var): self.var = var def __del__(self): counter[0] += 1 def __call__(self, *args): pass def run_test(): x = Variable(torch.ones(5, 5), requires_grad=True) y = x * 2 x.register_hook(GradHook(x)) y.register_hook(GradHook(y)) y._backward_hooks[1] = GradHook(y) run_test() gc.collect() self.assertEqual(counter[0], 3) </del> def test_hook_none(self): # WARNING: this is a test for autograd internals. # You should never have to use such things in your code.", "pos": ["My code runs well for a while , it occurs randomly..... Traceback (most recent call last): File \"\", line 236, in &lt;moduletrainvgg(\"\",dset='luna16') File \"\", line 54, in trainvgg lossvalue.backward() File \"/home/uih/anaconda2/lib/python2.7/site-\", line 169, in backward torch.autograd.backward(self, gradient, retaingraph, creategraph, retainvariables) File \"/home/uih/anaconda2/lib/python2.7/site-\", line 99, in backward variables, gradvariables, retaingraph) RuntimeError: add: Assertion \"pos = 0 &amp;&amp; pos &lt; ()\" failed. Any idea what's happend? Thanks in advance~\nYes, it looks like we have a bug somewhere in the code. Can you please try to isolate a small snippet that still triggers the same error so we can look into it? Thank you!\nIt's hard for me to isolate, the error occurs when I do the train and validation in turn, if I just do either the train or val, everthing runs well.\nThe problem is that we won\u2019t be able to fix it, until we actually see under what conditions does the error happen\nI got the same error in v0.3.0 today. when I trained my model with single GPU, it worked perfectly;with multi-gpu the error occured. RuntimeError: add: Assertion failed.\nWe've also been struggling with this issues for a long time. Unfortunately, finding which part of our code specifically is triggering the assert (which used to be a segfault before ) has proven extremely hard and, as of now, unsuccessful. In fact, we never submitted a bug report as we couldn't create a self-contained piece of code to trigger it. I'll try to summarize all the information we've gathered: We know for a fact that the issue appeared sometime after this commit: . All our code works perfectly fine on or v0.2. The assert is always triggered during the backward pass, but the exact moment when it happens seems to be completely random: when training over Imagenet we've seen this after 50 iterations as well as after 5000.", "This makes it super-hard to understand whether a particular piece of code is triggering the assert or not, as you can never be sure how many iterations are enough to exclude that the assert will ever trigger! The issue seems to be somehow related to the use of multiple GPUs (via ): we've never observed it while training with a single GPU. The issue seems also to be related to the use of user-defined classes: we've never seen it when training networks composed of \"standard\" Pytorch functions / modules only... ... however, simply running forward / backward passes with our custom functions doesn't seem to be enough to trigger the assertion. In our experience, working with \"complex\" computational graphs increases the likelihood of encountering this issue. Recently we've been able to re-create the issue using and a modified version of , where we replace some of the layers with our custom-made ones. Unfortunately I can't share the code for now (our paper describing the new layers is still not out), but I'll come back to you as soon as I can! Some additional information: all our experiments are on Ubuntu 16.04 using CUDA 8, CUDNN 7 and TITAN X and Xp GPUs.\nalmost the same condition as us\nEncounter the same condition after serveral imagenet epochs on a NasNet after update to 3.0. Pytorch 2.0 is working well.\nI am seeing this too -- I have two runs of the same model running with different hyperparameters, one is running smoothly, this one quit with the exact same error avgloss = self.trainepoch(trainingdata,validationdata,batchsize,saveprefix,epoch) File \"\", line 759, in trainepoch gradients = loss.backward() File \"/u/dcontrac/anaconda2/lib/python2.7/site-\", line 167, in backward torch.autograd.backward(self, gradient, retaingraph, creategraph, retainvariables) File \"/u/dcontrac/anaconda2/lib/python2.7/site-\", line 99, in backward variables, gradvariables, retaingraph) RuntimeError: add: Assertion failed. Running PyTorch v0.3 and training on 4 K-80 GPUs. My other run with different parameters is also running on the exact same setup but its still running.", "Both runs differ only in shuffle of data and a different batch size (the one that failed had a slightly larger batch size)\n: I'm not using any user define autograd.Function but I see the error too, so maybe thats not the primary place with a bug. Using nn.DataParallel though.\nI'll try again to train a network using standard layers only, maybe I'll also encounter the error in the no-custom-functions case if I let it run a little bit longer. If this is the case, then I guess we have some strong indication that (or multi-GPU computation in general) could indeed be the source of the error!\n: I dont think the duration has got anything to do with it --- One model run has been running without a problem for the last 18 hours, the other one had quit last night pretty early on (Crashed after the backward pass of the 8th batch). I have just restarted it and watching it.\nI agree that the bug is probably not related to training length. My guess is that each backward pass has a certain probability of randomly failing with this assert, so running for longer just increases the overall chances of encountering the bug.\nThanks for the information. We'd really like to fix this bug, and so there are a few things that would help us out a lot (in order of difficulty): you run your training process with gdb attached with a breakpoint, so that we can get a backtrace when the assertion fails? Also, if you can go one frame up and print the name of the function (run in gdb) that would also be helpful. We basically want to know if it's either a C++ function or a Python function. soon as you can publish your latest layers, uploading a script that reproduces (even if it takes a while to reproduce); we would be happy to run the script to try to reproduce, which would be very helpful for us. We'd love a bisection, but given that it can sometimes take 18hrs+ to repro, it doesn't sound like that would be easy to do.\nThis is the backtrace: For the function name I'll probably have to recompile pytorch in debug mode, since in gdb returns . Edit - I managed to get the name of the function by catching the exception raised by the assert: . Edit 2 - I did a second run and the name of the function is still the same.", "Actually I just realized that what we really need is Can you please try printing that (might require a debug build as well)?\nPrinting instead of I get , which is the name of our custom function's backward pass. More importantly, our code and paper are now public so I can share this with you: As I mentioned in a previous comment, this code is just the Pytorch ImageNet + ResNet examples, except that standard BN is replaced with our custom layer. When running this on our machines with 4 GPUs we always encounter the assert within the first epoch. Please feel free to ask if you need help running the code.\nI also encountered this assert. I have two machines, one with 2, and other one with 4 GPUs. Both machines are Ubuntu 16.04, python 2.7, cuda 9.0, cudnn 7, and pytorch v0.3.0. I never encountered this assert on my 2 GPUs machine. This assert always appeared on my 4 GPUs machine.\nsame as here. I have python 2.7, cuda 8.0, cudnn 7 and pytorch v0.3.0 installed from binaries. The assertion was triggered when I was training networks on 4 GPUs. It happened at a random time during training.\nSame here, with multiple pixel-wise prediction models with just CNNs. Seems to occur only with multiple GPUs, has not happened so far with single GPU. python 2.7, CUDA 8.0, CUDNN 7, PyTorch v0.3.0.\ncan you take a look at this asap\nedit: resolved, I ended up adding a compilation flag for compute capacity 3.5 to pytorchbugreport's building scripts (to work with my gpus) I'm getting the following after trying to run pytorchbugreport. Have you seen this before? !\nyou have to change to cover your GPU, which is compute35,code=sm_35 i think (if it is K40)\nI ran your code for six epochs on a machine with 4 gpus on pytorch 0.3.0 but haven't been able to trigger the assert. I'll try letting it run for longer and see if anything pops up\nFor us the assert almost always triggers during the first epoch.", "There might be some important difference in our setups. This is the exact sequence of steps we have used to reproduce the error across several machines: from a clean Ubuntu 16.04 installation with CUDA 8.0, CUDNN 7.0, pip, virtualenv and the compilers. and activate a new python2 virtualenv. necessary python packages . the pytorch repository and checkout and install with our repository and compile our native module with the training as described before. Our machines all have either Titan X or Titan Xp GPUs, so a different architecture from the one you are using, I don't know if this could have an influence on the error.\nMy machine has 4 GTX 1080 Ti which triggers the error.\nwe (finally) found a different repro that triggers this issue internally on one of our machines, a fix will be issued soon.\nQuick update: we're still looking into this, but so far, I've been able to get this bug to trigger with two machines using python 2.7 but never with python 3. If you're in need of a quick fix using python 3 might help.\nI also encountered this propblem, with pytorch 0.3.0, python2.7, CUDA8.0, Titan X GPUs. My session always crashed around 8 epochs. And I degrade pytorch to 0.2.0. now the training seems work well. Someone in emergency can try this.\nThank you , I'll try to switch to python 3 for now!\nlet me know if the bug comes back in python 3. As far as I can tell the python version shouldn't matter in the triggering of this bug... but it's never happened to me with python 3 yet.\nCrashed in Python2.7 RuntimeError: add: Assertion failed Absolutely random crash using multi gpu.\nThis has been fixed in HEAD but there has not been a release of 0.3 with the fix yet.\nOh, Thanks for your reply. I will try the HEAD version. Thank you .\ni have encountered this error , but change the pytorch to the latest version(0.4) will solve it.\nEncountered same error when I add dropout layer to a Resnet which I modified. Trained on multi-GPU. Does latest conda release fix the problem?\nit is not yet on latest conda release.", "to get the fix at the moment you will have to install pytorch from source:\nThe same bug ! !\ndid you install pytorch from source?\nIs there a simple way to block this error without changingthe version of python or pyotrch?\nYou could disable the python reference cycle . This would probably lead to memory leaks though.\nIt should be safe if you do this once all the graphs are deleted, and there will be no leaks unless you create cycles in your data structures\nAfter 2 days of struggle, I finally changed back to 0.2.0 and expects the release of binary 0.4.0. Anyway, thank you for your advices. :)\nI encountered the same error with version 0.3.0.post4. What was the proper fix here?\nupgrade to version 0.3.1 or newer."], "neg": []}
{"query": "if mode not in {'min', 'max'}: raise ValueError('mode ' + mode + ' is unknown!') if threshold_mode not in {'rel', 'abs'}: <del> raise ValueError('threshold mode ' + mode + ' is unknown!') </del> <ins> raise ValueError('threshold mode ' + threshold_mode + ' is unknown!') </ins> if mode == 'min' and threshold_mode == 'rel': rel_epsilon = 1. - threshold self.is_better = lambda a, best: a < best * rel_epsilon", "pos": ["Trivial report, unfortunately I can't make a PR at the moment, but I can submit a report. line 329 in should replace with"], "neg": []}
{"query": ".. automethod:: log2 .. automethod:: log2_ .. automethod:: log_normal_ <ins> .. automethod:: logsumexp </ins> .. automethod:: long .. automethod:: lt .. automethod:: lt_", "pos": ["Leaving this issue here for higher visibility. Originally reported by here: Also mentioned by Piggybacking off of this issue, has strange behavior (reported here: )\nI have fixed this locally. Could I send a PR?\nYes, go ahead!"], "neg": []}
{"query": "] extra_compile_args += ['-DWITH_DISTRIBUTED_MW'] include_dirs += [tmp_install_path + \"/include/THD\"] <del> main_link_args += [THD_LIB] </del> <ins> main_link_args += [THD_LIB, THPP_LIB] </ins> if platform.system() == 'Linux': main_link_args += [GLOO_LIB]", "pos": ["It seems that the shared object file is not being built correctly and this is leading to build errors on Travis.\nthis is weird, i saw this intermittently on travis as well. But couldn't reproduce on subsequent builds on Travis or locally.\nI ran a clean build on the latest commit locally just now and I can't reproduce it either. Is it that ccache is preserving some build files and those are interfering with subsequent builds?\nActually, I apologize, I am able to reproduce. Importing torch in an ipython shell gives the same issue.\nI meet the same problem. I am using the newest version \"commit \". And I checked all dependent libraries of , all of them are available. Any advices are welcome. Much thanks.\nI have reproduced something similar. It seems to only happen when you do . EDIT: No, my problem was actually that I was running Python inside my PyTorch checkout.\nIt also happens when you do\nVarun, i cant reproduce this. What's your OS version and python version?\nI have met the same problem as is mentioned above. I'm using Ubuntu 16.04.3 and Anaconda python 2.7.13.\nUbuntu 16.04.2 Python 3.6.2 Cuda 8.0\none last thing, what's your gcc version, and did you do python clean, or clone a fresh new folder\nI did a , but am running it again to make sure.\nOn definitively running , I am not getting the error.\nthanks everyone, should be fixed on master now."], "neg": []}
{"query": "num_devices = 0; } #endif <del> ready_queues = std::vector<std::shared_ptr<ReadyQueue>>(num_devices + 1); for (auto& queue : ready_queues) { </del> <ins> int num_threads = num_devices + 1; ready_queues = std::vector<std::shared_ptr<ReadyQueue>>(num_threads); for (int i = 0; i < num_threads; ++i) { auto& queue = ready_queues[i]; </ins> queue.reset(new ReadyQueue()); <del> std::thread t(&Engine::thread_main, this, queue); </del> <ins> std::thread t(&Engine::thread_main, this, queue, i - 1); </ins> t.detach(); } }", "pos": ["Consider following code, and run it with multi gpus (e.g. 4): It will output: That is, the default device would always be zero even in DataParallel. My pytorch version is 0.1.122. I think this is not the desired behavior. It would cause some troubles. I tried to insert my own cuda kernel into backward to calculate the gradients, it become very slow, and I fixed it by (grad.getdevice()). Anyway, I think currentdevice in forward and backward should be the same, or could anyone explain to me why they are different?\nthis seems like something we should fix: ?\n<!-- drci-comment-start --:pagefacingup: Preview :pagefacingup: Preview :question: Need help or want to give feedback on the CI? Visit our Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. <!-- drci-comment-end --"], "neg": []}
{"query": "return make_tensor(t, M, M) <ins> def medium_2d_expanded(t): return t(1).expand(M, M) </ins> def medium_2d_scaled(t, scale=10): return make_tensor(t, M, M).mul(scale)", "pos": ["On recent PyTorch master (built from source; 0.4.0a0+) on a P100, CUDA 9:\nI think the results on CPU meets the defination of triu (upper triangular part of the matrix). It is really hard to handle with a tensor from expand operation with multi-thread not only on GPU but also on CPU. The unexpected result will also happen on CPU if we apply openmp to it. I think it is necessary to eleminate the difference between the implementation and defination.\nIt doesn\u2019t matter if it\u2019s cpu or not. triu doesn\u2019t have an underscore which means it doesn\u2019t operate in-place and should always give the same result (in new memory)"], "neg": []}
{"query": "ctx.keepdim = False if keepdim is None else keepdim if dim is None: <del> ctx.norm = input.norm(p) ctx.save_for_backward(input) return input.new((ctx.norm,)) </del> <ins> norm = input.norm(p) output = input.new((norm,)) </ins> else: if keepdim is not None: output = input.norm(p, dim, keepdim=keepdim) else: output = input.norm(p, dim) <del> ctx.save_for_backward(input, output) return output </del> <ins> ctx.save_for_backward(input, output) return output </ins> @staticmethod def backward(ctx, grad_output): <del> if ctx.dim is None: input, = ctx.saved_variables if ctx.p == 2: scale_v = (grad_output / ctx.norm).expand_as(input) return input.mul(scale_v), None, None, None else: pow = input.abs().pow(ctx.p - 2) scale_v = (grad_output / ctx.norm ** (ctx.p - 1)).expand_as(input) return input.mul(pow).mul(scale_v), None, None, None </del> <ins> input, output = ctx.saved_variables if ctx.dim is not None and ctx.keepdim is False and input.dim() != 1: grad_output = grad_output.unsqueeze(ctx.dim) output = output.unsqueeze(ctx.dim) if ctx.p == 2: grad_input = input.mul(grad_output).div(output) </ins> else: <del> input, output = ctx.saved_variables </del> <ins> input_pow = input.abs().pow(ctx.p - 2) output_pow = output.pow(ctx.p - 1) grad_input = input.mul(input_pow).mul(grad_output).div(output_pow) </ins> <del> if ctx.keepdim is False and input.dim() != 1: grad_output = grad_output.unsqueeze(ctx.dim) output = output.unsqueeze(ctx.dim) </del> <ins> # Special case at 0 where we return a subgradient containing 0 grad_input.masked_fill_(output == 0, 0) </ins> <del> big_grad_output = grad_output.expand_as(input) if ctx.p == 2: big_output = output.expand_as(input) return input.mul(big_grad_output).div(big_output), None, None, None else: pow = input.abs().pow(ctx.p - 2) big_output = output.pow(ctx.p - 1).expand_as(input) return input.mul(pow).mul(big_grad_output).div(big_output), None, None, None </del> <ins> return grad_input, None, None, None </ins> # TODO: renorm", "pos": ["If a norm is zero, its gradient returns nan: Obviously just happening because the gradient divides by the norm, but the (sub)gradient here should probably be zero, or at least not nan, since that will propagate to make all updates nan. Probably low priority, as it's not going to be an issue in 99% of cases, but we're doing a few things with (exact) line searches where this caused a nan to appear, breaking everything downstream.\nI'm encountering exactly the same issue! Spent hours on debugging, just to find PyTorch has a bug on such basic thing.\n+1 just found this bug too\n+1 for this bug. Temporarily changing my code to something like the following for the sake of debugging. x = Variable((1), requires_grad=True) y = x + 1e-16 y.norm().backward() print x.grad\nThe thing is that in the 2 norm, there is a square root, which has a gradient of at 0. The gradient gives you because you then multiply 0 and an infinity during the backward pass.\nFor a scalar, norm 2 is basically abs. But x.abs().backward() gives you 0 gradient. In this sense, it's not coherent.\nI found this error, too\nAlban fixed this behavior in\nHi, the norm function can give use the 0 gradients now. However, the following code still has the nan gradient problem\nHo The square root has no gradient at 0. This is expected behavior.\nHi, but the sub-gradient of the square root should be zero? Also, y = ( x * x ) should equal to x.norm(), why they have different gradient ( 0 and nan )?\nI think was right. The left-side derivative of sqrt(x) at x=0 is undefined, so it doesn't even have a subgradient at x=0.\nsquare root has no subgradient at 0. You could define a gradient by continuity but then it would be ... Given that pytorch is using autograd, and (equivalent to your ) are completely different: The first one is a single function that is convex and defined on R, it has a subgradient of 0 at 0. The second one is composed of two function, the first function is the square function which is differentiable and outputs values in .", "The second function is the square root that is not convex and even though it is defined on , it is only differentiable on and it's gradient in 0 in undefined. Given that, even though and will return the same value, their gradients may differ at points where it is not differentiable, this is because automatic differentiation looks at each step of the computation one by one and even though in some cases a subgradient exist (because we look at multiple operations as a single function), it is not always the case and the gradient remains undefined.\nI think math is math. Any root's gradient at zero is either inf or undefined. This issue shall be handled by the users' themselves by adding a small value(as did), but an error(warning) message may be helpful since it is pretty hard to debug. Say: Infinite/Undefined gradient is detected at XfunctionX at line Y. Exit.\nAgree, norm is not differentiable at 0 , the bandaid that Alban put there in is wrong (even in the limit sense the gradient at 0 should be 1 not 0), but it should not have been there at all. Norm is norm, if someone want to add epsilons to their norms (like batchrnorm, e.g.) they are welcome to do so in the user code. What would numpy do?\nI've also run into a number of problems related to the change introduced in . Is there a reason the subgradient is set to 0, rather than the 1? (The limit as norm-0?) As a minimal example: Produces\nWell any value between [-1, 1] is a valid subgradient for the 2-norm. More genereally, any vector in the 1 ball for the dual norm is a valid subgradient. This means that 0 is always going to be a subgradient, while 1 will not be for all p. Anyway, the theory says that any of them could be taken and subgradient descent will work. I'm sure that depending on the application, one will be better than the other. For example, the relu function will also give a 0 subgradient at 0, you could have given 1. The main point here was to remove nans that make your network give nan for everything which is not convenient."], "neg": []}
{"query": "else: sizes = torch.LongStorage(args) sizes = _infer_sizes(sizes, self.nelement()) <del> if reduce(lambda a,b: a * b, sizes) != self.nelement(): raise RuntimeError('Invalid size for view. Input size: ' + 'x'.join(map(lambda v: str(v), self.size())) + ', output size: ' + 'x'.join(map(lambda v: str(v), sizes)) + '.') assert self.is_contiguous(), \"expecting a contiguous tensor\" dst.set_(self.storage(), self.storage_offset(), sizes) </del> <ins> numel = reduce(lambda a,b: a * b, sizes) if len(sizes) > 0 else 0 if numel != self.nelement(): def format_size(size): return 'x'.join(str(v) for v in size) if len(size) > 0 else '0' raise ValueError( \"view of size '{0}' is invalid for input of size '{1}'\" .format(format_size(sizes), format_size(self.size()))) if not self.is_contiguous(): raise ValueError(\"input should be contiguous\") if self.storage() is not None: dst.set_(self.storage(), self.storage_offset(), sizes) </ins> return dst def view_as(self, tensor):", "pos": ["Please check and fix all the PyTorch web links format from to , like in the official site.\nclosed via"], "neg": []}
{"query": "-DNO_NNPACK=%NO_NNPACK% ^ -DCUDNN_INCLUDE_DIR=\"%CUDNN_INCLUDE_DIR%\" ^ -DCUDNN_LIB_DIR=\"%CUDNN_LIB_DIR%\" ^ <del> -DCMAKE_BUILD_TYPE=Release </del> <ins> -DCMAKE_BUILD_TYPE=%BUILD_TYPE% </ins> %MAKE_COMMAND% cd ../..", "pos": ["The current master triggers the following error messages from Visual Studio 2015 (Debug build): 2------ Build started: Project: ATen, Configuration: Debug x64 ------ 2cl : Command line error D8016: '/Ox' and '/RTC1' command-line options are incompatible ... 14LINK : fatal error LNK1104: cannot open file '....srcATenDebug' It appears that they are triggered by *.c files in the ATen project, like \"srcTHvectorAVX.c\"\nWhy do you need Debug build?\nIsn't ATen meant to support both Debug and Release builds? FYI I'm involved with the project at which does support both build configurations, and we would like to be able to link our project with ATen.\nWell, you can remove the flags in as a workground.\nWe should patch this file to ignore this flag when building the library in debug mode. It shouldn't be too hard\nI catched similar errors , when I Tryed to build Debug Version I try to build Release and catched next errors e:pytorchatensrcatencpuvec256vec256int.h(184): note: see reference to function template instantiation 'at::vec256::anonymous-namespace'::Vec256<int16t::blend<1(at::vec256::anonymous-namespace'::Vec256<int16t)' being compiled 28e:pytorchatensrcatencpuvec256vec256int.h(147): error C3861: 'mm256extract_epi16': identifier not found What I should do?\nThere're two issues in your description. The first question is easy. You'll also have to specify the flag to 1. The second question is about your compiler. I guess you are using VS2015 or earlier versions of VS2017. Make sure the version in the output is 19.11 after your enter in CMD.\nYes I am using VS2015. My Compiler is 19.0.2. Must I use VS 2017 in any case?\nSorry, but the answer is yes. The AVX2 functions was in VS2017 and they are used in PyTorch. So VS2017 19.11 is a must.\nNo problem:) Thank you."], "neg": []}
{"query": "for i, (sample, target) in enumerate(loader): idx = i * batch_size self.assertEqual(sample, self.data[idx:idx + batch_size]) <del> self.assertEqual(target, self.labels[idx:idx + batch_size].view(-1, 1)) </del> <ins> self.assertEqual(target, self.labels[idx:idx + batch_size]) </ins> self.assertEqual(i, math.floor((len(self.dataset) - 1) / batch_size)) def _test_shuffle(self, loader):", "pos": ["TensorDataset adds casts 1D target tensor to 2D which is not acceptable by NLL loss layer.\nFixed in ."], "neg": []}
{"query": "CUDNN_LIB_DIR = path break else: <del> libraries = sorted(glob.glob(os.path.join(path, 'libcudnn*'))) </del> <ins> libraries = sorted(glob.glob(os.path.join(path, 'libcudnn*' + str(CUDNN_INCLUDE_VERSION) + \"*\"))) </ins> if libraries: CUDNN_LIBRARY = libraries[0] CUDNN_LIB_DIR = path break <del> for path in include_paths: if path is None or not os.path.exists(path): continue if os.path.exists((os.path.join(path, 'cudnn.h'))): CUDNN_INCLUDE_DIR = path break </del> # Specifying the library directly will overwrite the lib directory library = os.getenv('CUDNN_LIBRARY') if library is not None and os.path.exists(library):", "pos": ["Traceback (most recent call last): File \"\", line 3, in <moduleimport model File \"/home/abc/github/chinese-\", line 5, in <modulefrom import crnnOcr File \"/home/abc/github/chinese-\", line 72, in <modulemodel,converter = crnnSource() File \"/home/abc/github/chinese-\", line 64, in crnnSource model = crnn.CRNN(32, 1, len(alphabet)+1, 256, 1).cuda() File \"/usr/local/lib/python2.7/dist-\", line 216, in cuda return (lambda t: t.cuda(device)) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 123, in apply self.flattenparameters() File \"/usr/local/lib/python2.7/dist-\", line 65, in flattenparameters if not or not (anyparam): File \"/usr/local/lib/python2.7/dist-\", line 82, in isacceptable if libcudnn() is None: File \"/usr/local/lib/python2.7/dist-\", line 58, in libcudnn 'but linked against {}'.format(compileversion, _cudnnversion)) RuntimeError: cuDNN version mismatch: PyTorch was compiled against 6021 but linked against 5110 How to solve the problem?\ncould you pull the latest master branch and recompile. I pushed a change yesterday that fixes this:\nI remove the old code and pull new code, then compile, install.It still fails. /usr/local/lib/python2.7/dist- UserWarning: No training configuration found in save file: the model was not compiled. Compile it manually.", "('No training configuration found in save file: ' Traceback (most recent call last): File \"\", line 3, in <moduleimport model File \"/home/qianbb/github/chinese-\", line 5, in <modulefrom import crnnOcr File \"/home/qianbb/github/chinese-\", line 72, in <modulemodel,converter = crnnSource() File \"/home/qianbb/github/chinese-\", line 64, in crnnSource model = crnn.CRNN(32, 1, len(alphabet)+1, 256, 1).cuda() File \"/usr/local/lib/python2.7/dist-\", line 216, in cuda return (lambda t: t.cuda(device)) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 100, in apply self.flattenparameters() File \"/usr/local/lib/python2.7/dist-\", line 66, in flattenparameters if not or not (anyparam): File \"/usr/local/lib/python2.7/dist-\", line 87, in isacceptable if libcudnn() is None: File \"/usr/local/lib/python2.7/dist-\", line 58, in libcudnn 'but linked against {}'.format(compileversion, _cudnnversion)) RuntimeError: cuDNN version mismatch: PyTorch was compiled against 6021 but linked against 5110 The library of cuda is shown as following screenshot. !\nI found the issue, I fixed it in which we will merge shortly.\nThanksDoes it have been merged?\nyes. it is merged.\nIt seems the fix doesn't work.", "The log is shown as following Traceback (most recent call last): File \"\", line 3, in <moduleimport model File \"/home/qianbb/github/chinese-\", line 5, in <modulefrom import crnnOcr File \"/home/qianbb/github/chinese-\", line 72, in <modulemodel,converter = crnnSource() File \"/home/qianbb/github/chinese-\", line 64, in crnnSource model = crnn.CRNN(32, 1, len(alphabet)+1, 256, 1).cuda() File \"/usr/local/lib/python2.7/dist-\", line 222, in cuda return (lambda t: t.cuda(device)) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 100, in apply self.flattenparameters() File \"/usr/local/lib/python2.7/dist-\", line 66, in flattenparameters if not or not (anyparam): File \"/usr/local/lib/python2.7/dist-\", line 87, in isacceptable if libcudnn() is None: File \"/usr/local/lib/python2.7/dist-\", line 58, in libcudnn 'but linked against {}'.format(compileversion, _cudnnversion)) RuntimeError: cuDNN version mismatch: PyTorch was compiled against 6021 but linked against 5110\nthat is weird. Can I see the full log of in pytorch directory. Can you make sure you have latest commits with .\nI am sure the code is up-to-date. First I uninstall the code using the command pip unisntall torch. Then I remove the old code using the command rm -r pytorch and git clone again. The log of install is\nSorry to trouble you. Is there any new information on this issue?\nI spent some time resolving this issue and I don't see too much about it online. Try ensuring that your Pytorch build is correct for your specific versions of Cuda and Cudnn. While I had the correct version of Pytorch, I resolved the issue by installing a different build of the same version.", "The build I installed is . Here are the list of versions and builds: You can download these files locally and install with conda: (assumes command is being run in a directory containing , where this file will be specific to your Cuda and Cudnn version.\nfeedback from this code PackagesNotFoundError\nthank you - I have edited my comment to reflect that I installed from a local zip."], "neg": []}
{"query": "(hx + cx).sum().backward() <del> @unittest.skipIf(not TEST_CUDNN, \"needs cudnn\") @default_tensor_type(torch.FloatTensor)  # FIXME: just until torch.cuda.DoubleTensor.sum() implemented def test_RNN_cpu_vs_cudnn(self): </del> <ins> def _test_RNN_cpu_vs_cudnn(self, dropout): </ins> def forward_backward(cuda, rnn, input_val, hx_val, weights_val): is_lstm = type(rnn) == nn.LSTM", "pos": ["cuDNN backend for RNN stabilizes at large mem usage. The following code stabilizes for me at 14+GiB reported by nvidia-smi. The same RNN without cuDNN uses <2.5GiB.\nOops... any idea what's going on with the garbage collector in this case?\nare you running this on a P100 GPU?\nYes, I observe similar behavior on Titan X Pascal\nhmmm, do cudnn RNNs have weird memory-oversubscribing properties? If it wasn't for CuDNN RNN, then it's not unreasonable. we use a custom CUDA allocator that will oversubscribe if needed, but will free the memory blocks it's holding onto if it hits OOM / failed allocations.\nTried the same RNN in torch and am getting ~2.6GiB on nvidia-smi.\nWith cuDNN I'm getting a constant usage of 2241MB on my machine using PyTorch (M40).\nare you using cudnnv5, or v6?\nv5\nno, cudnn does not do memory oversubscribing, and already tried the same network in lua torch with cudnn, where it gets expected memory usage. Also, he is getting OOMs on slightly bigger networks, where apparently caching allocator should have tried to free previous blocks. Interesting that can not repro this.\nJust to make sure did you set when you ran? I left it as False in the repro.\nI can repro in the docker image built from the current version on P100, and I go OOM after 8 batches with the same image on 1080, which generally should not happen - if there is enough memory to go through 1 batch, memory usage should not increase afterwards.\nNo, I just copied your script and ran it. Weird thing is that it goes OOM after 18 batches once I enable . Without it memory remains stable.\nA lot of cudaMalloc allocations go not through THCudaMalloc, and never get freed:\nSo not going through THCudaMalloc is actually expected for storages allocated with THCStorage_(newWithSize), but what's not expected is to every time cudaMalloc bytes of rnn workspace, instead of reusing earlier allocations. (~784 MB allocations are for workspace, ~156MB allocations happen somewhere in criterion).\nMy guess is that the bug is at the Python level.", "Either: Something is holding on to a Tensor longer than it should Something is creating a circular reference which prevents quick clean-up\nThe issue is that in creates cyclic references. It's made worse by the fact that it holds on to more memory than it should (like the workspace). You can see this by adding a as the first line in your for loop. With the explicit call to the GC every iteration, the memory usage doesn't explode.\nhow can these cyclic references be fixed?\nNot sure yet, but I will fix it tomorrow with"], "neg": []}
{"query": "# Make sure state dict wasn't modified self.assertEqual(state_dict, state_dict_c) <ins> # Check that state dict can be loaded even when we cast parameters # to a different type and move to a different device. if not torch.cuda.is_available(): return input_cuda = Variable(input.data.float().cuda()) weight_cuda = Variable(weight.data.float().cuda(), requires_grad=True) bias_cuda = Variable(bias.data.float().cuda(), requires_grad=True) optimizer_cuda = constructor(weight_cuda, bias_cuda) fn_cuda = functools.partial(fn_base, optimizer_cuda, weight_cuda, bias_cuda) state_dict = deepcopy(optimizer.state_dict()) state_dict_c = deepcopy(optimizer.state_dict()) optimizer_cuda.load_state_dict(state_dict_c) # Make sure state dict wasn't modified self.assertEqual(state_dict, state_dict_c) for i in range(20): optimizer.step(fn) optimizer_cuda.step(fn_cuda) self.assertEqual(weight, weight_cuda) self.assertEqual(bias, bias_cuda) </ins> def _test_basic_cases(self, constructor, ignore_multidevice=False): self._test_state_dict( torch.randn(10, 5),", "pos": ["Hi, I encountered this bug: The code skeleton is like: It seems the loaded are , and I've tried a workaround to move to , but it still has the same bug.\nCould you provide a full script to reproduce the problem?\nmaybe you can try like this, () (beta1).add_(1 - beta1, ())\nSorry, I missed the reply email. I am afraid that I am unable to provide a reproducer now. It is a work I am doing for the OpenNMT-py , trying to use for doing lr update. And I encoutered this problem when testing the case. So I factored out the code skeleton about this problem above. I've tried several methods, including tricks like what suggests, but it still happens. Per my analysis, it is because the previous training was done on gpu, so when saving the , the stored states(tensors) are of version. During resuming, when we load the saved optimizer, loads this version to cpu(the model() can be moved to gpu easily, but seems lacking this ability?) , so this problem emerges.\nTry moving optimizer state to the GPU memory manually after loading it from the checkpoint. I agree that having an method for this operation would be nice.\nthanks. My work was suspended due to other problems, when resumed, I will try your method.\nI'm afraid solution won't work. It will make the error go away, but your optimizer will no longer be training the model. You should recreate optimizers after casting modules to a different type or device, and you can use to restore the state from a previous copy. This currently doesn't work, but we should fix it (by copying from the data from the state dict, instead of using the tensors directly - this allows for cross-device or cross-type updates).\n, yep, your method is what I currently use, it works. But I will wait for upstream to fix this problem though. Thanks for your great works!\nAh, my bad. I forgot to update the line where the optimizer is recreated. But otherwise, the following should do the job, right?\nah, right. That should work\nExcept that you should use instead of\nI had a similar problem. When I save the optimizer state from a GPU other than GPU 0 and then load the state it still loads everything to GPU 0.", "Specifying in didn't work either. 's solution solves this though.\nHi guys, I have a very similar problem as the one in this thread, here's my code: And then once I resume, I got KeyErrors on my optimizer: Do you guys know how to fix this issue? BTW, I have 8 GPUs used, I'm guessing if this issue was because of that?\nwere you able to solve your issue with multiple-GPUs?\nThanks for asking. Yeah, I'm able to, turns out the issue is because I used an early version of PyTorch, after I updated the version, it works like a charm!\nJust a comment, this problem is caused by makes all state tensor are moved into GPU0 so by moving the state of an optimizer to specific GPU will fix this problem.\nHi how do you know the specific GPU location of different tensors in advance?\nWould a feature where all () calls always makes use of an automatically generated CPU version be feasible ? And then at resume the () would make use of the \"current\" device being used (or any better strategy). At the moment it seems we need a lot of boilerplater code to ensure saving and loading are consistents across devices for models/optimizer/scheduler/etc.\nI have met similar problem, I recreated Adam optimizer without () after reloading model, () and DataParallel(model) according to solution.\nthanks, it work!\nHi, as you say that: every time moving model to other device, we should build optimizer, but, if we move the model to other device and move back, should we build the optimizer again? here is an example code: does optimizer run as expected? also, if doing , should we rebuild optimizer ?\nIs there a problem if you switch the order to something like this? Meaning moving the model to 'cuda' but only loading it's state dict from checkpoint after loading the optimizer's state dict first?\nThe problem can be concluded that the optimizer's state will be loaded to the device as same as the model. You must load the model to GPU at first, and then load the optimizer's state. So that both the model and the optimizer's state are loaded in GPU.", "Instead of moving optimizer to cuda after loading it in cpu, you could load the checkpoint directly in cuda:\nI've independently rediscovered that this works :) Should read until the end of the thread next time\nI find my codes still have the problem. I tried my best to range the modules as the examples shown in the above. can anyone give me some hints? I also tried this\nFor me, doesn't solve the issue, but the one by works. Any idea why? I create the optimizer, load the state (with map location to CUDA), pass to the train loop, where the mode is pushed to device (though I assume it's not needed if the loaded model is pushed to cuda by maplocation). Specifically, I would get stuck at call, throwing the device mismatch error.\nStill it doesn't work for me. Could you please help me?"], "neg": []}
{"query": "std::vector<at::Tensor> unpack(at::TensorList tl, const char *name, int pos) const; std::vector<at::Tensor> unpack_idxs(at::TensorList tl, const char *name, int pos) const; <del> Variable as_variable(Tensor tensor) const; std::tuple<Variable, Variable> as_variable(std::tuple<Tensor, Tensor> tensor) const; std::tuple<Variable, Variable, Variable> as_variable(std::tuple<Tensor, Tensor, Tensor> tensor) const; std::tuple<Variable, Variable, Variable, Variable> as_variable(std::tuple<Tensor, Tensor, Tensor, Tensor> tensor) const; std::vector<Variable> as_variable(TensorList tensor) const; Variable maybe_wrap(Tensor data, const Variable & self, bool inplace) const; </del> private: at::Type* baseType; std::string str;", "pos": ["See Docker run fails with , which is due to the fact that is specified in docker container: Which results in: CI cc\nCheck that adding to manywheels image results in a failure:\nTentatively marking hi-pri since nightly testing are failing\nThis is resolved, here is the successful run:"], "neg": []}
{"query": "python ..ci_scriptsdelete_image.py 7z x %IMAGE_COMMIT_TAG%.7z <del> sh run_test.sh </del> <ins> sh run_test.sh -- -v </ins> EOL", "pos": ["If a test segfaults, it's hard to determine which test it is because the output looks like: ! I can work on this if we think this is a good idea.\nLast time I asked about this, he didn't want to because it adds a lot of scrolling. I also did some work to move the CI tests to so that we can output XML status files, but (1) this is unlikely to help in case of segfault, and (2) you can't do this straightforwardly, because we run some test files multiple times with different environment variables, and the XML exporter isn't really set up to handle this case. For the record, I am personally OK with just slapping on the tests and calling it a day.\nin the conbuild sounds good to me too\nclosed via"], "neg": []}
{"query": "int collapseIndicesDim = indicesInfo.collapseDims(dim);                 int64_t inputSlices = 1;                                               <del> int64_t topKSlices = 1;                                                for (int i = 0; i < numDims; ++i) {                                    </del> <ins> for (int i = 0; i < inputInfo.dims; ++i) {                             </ins> inputSlices *= inputInfo.sizes[i];                                   <ins> }                                                                      int64_t topKSlices = 1;                                                for (int i = 0; i < topKInfo.dims; ++i) {                              </ins> topKSlices *= topKInfo.sizes[i];                                     }                                                                      ", "pos": ["Repro: Expect two small tensors of small integers to be printed. Instead, only the first case works. In the second case, you get a series of failed assertions starting with this: I wonder if it is related to\nis this still an outstanding issue? I cannot reproduce this with an install from the master branch.\nThe size which causes failure doesn't seem deterministic. Using a pretty recent build (a few days old), I now have to increase the \"9\" to a larger number to make it fail. Sometimes it will fail at 10, sometimes at 11. Here's an updated case that seems to fail reliably at some size, for me:"], "neg": []}
{"query": ">>> hx = Variable(torch.randn(3, 20)) >>> output = [] >>> for i in range(6): <del> ...     hx = rnn(input, hx) ...     output[i] = hx </del> <ins> ...     hx = rnn(input[i], hx) ...     output.append(hx) </ins> \"\"\" def __init__(self, input_size, hidden_size, bias=True, nonlinearity=\"tanh\"):", "pos": ["Previously, worked to install CPU version explicitly. However, does not work. Are the whl files for 2.4 missing? N/A cc\nHi, this link is deprecated since 2.4. Please use the official installation guide instead.\nStill it is not available for macosx x86 platform\nHi and Please use official install commands from For linux its: MacOS x86 binaries where deprecated since release 2.3.0.\nClosing as resolved. We previously used the command in original post for building dockerfiles for specific versions. I have now replaced it with:\nThis still installs torch 2.2.2 but not 2.4.0 On Monday, August 19, 2024 at 10:41:25 PM GMT+5:30, Mihir Patel Closing as resolved. We previously used the command in original post for building dockerfiles for specific versions. I have now replaced it with: pip${PYTHONVERSION} install --no-cache-dir --find-links  torch=={CUDAVERSIONTAG} \u2014 Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID:\nThis still installs torch 2.2.2 but not 2.4.0\npython --version Python 3.12.5 pip3 install torch torchvision torchaudio --index-url Collecting torch Downloading (151.0 MB) 151.0/151.0 MB 11.3 MB/s eta 0:00:00 Successfully installed torch-2.2.2 torchaudio-2.2.2 torchvision-0.17.2 but not 2.4.0"], "neg": []}
{"query": "cache_key = metadata[1] rebuild = rebuild_storage_filename storage._shared_incref() <ins> elif storage.size() == 0: # This is special cased because Empty tensors # (with size 0) cannot be mmapped. return (rebuild_storage_empty, (type(storage),)) </ins> else: fd, size = storage._share_fd_() if sys.version_info[0] == 2:", "pos": ["PyTorch version: 0.4.0a0+ Repro: Result:\nEmm. I fail to reproduce this bug in Python 3.5.0 and Mac OS 10.13.3. What's your system environment?\nThe bug is probably specific to the sharing mode. Mac OS uses by default."], "neg": []}
{"query": "from tools.setup_helpers.env import check_env_flag from tools.setup_helpers.cuda import WITH_CUDA, CUDA_HOME from tools.setup_helpers.cudnn import WITH_CUDNN, CUDNN_LIB_DIR, CUDNN_INCLUDE_DIR <ins> from tools.setup_helpers.split_types import split_types </ins> DEBUG = check_env_flag('DEBUG') WITH_DISTRIBUTED = check_env_flag('WITH_DISTRIBUTED') WITH_DISTRIBUTED_MW = WITH_DISTRIBUTED and check_env_flag('WITH_DISTRIBUTED_MW')", "pos": ["Trying to run on github actions via AWS instance gives me: I will of course set up now, but it shouldn't crash like that. May be it should use as a fallback? pt-2.0 cc\nIs this a multitenant setup? Also curious which distro are you running, I was under the impression that most linux distros set home by default\nthis is AWS image. one default user. I think it's user, but it wants you to run as root, which is weird. I have manually now set to overcome this.\nWe should update triton to use pathlib or something"], "neg": []}
{"query": "\"\"\"Fills the input Tensor or Variable with values according to the method described in \"Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification\" - He, K. et al. (2015), using a normal distribution. The resulting tensor will have values sampled from :math:`N(0, std)` where <del> :math:`std = sqrt{2 / ((1 + a^2) times fan_in)}`. Also known as He initialisation. </del> <ins> :math:`std = sqrt{2 / ((1 + a^2) times fan_in)}`. Also known as He initialisation. </ins> Args: tensor: an n-dimensional torch.Tensor or autograd.Variable", "pos": ["why 1.21.6? because that's the oldest supported numpy on windows/python3.10: all cc\nHere is the test for pinning numpy to 1.21.6 The reason its was not pinned to this version is that conda can't properly resolve all the dependencies at this step here:\nIt looks like you are correct, I will put up a PR to pin lower then already pinned:\nHere is the PR:"], "neg": []}
{"query": "PyObject *THSPTensorClass = NULL; <del> static void THSPTensor_(initStorage)(THSTensor* tensor) </del> <ins> static void THSTensor_(initStorage)(THSTensor* tensor) </ins> { // Ensure that PyTorch's \"storage is not NULL\" invariant is upheld. // See Note [Storage is not NULL]", "pos": ["This causes to not work properly on newly created empty sparse tensors. In the example below, both should print 1 instead."], "neg": []}
{"query": "def test_view(self): tensor = torch.rand(15) template = torch.rand(3, 5) <ins> empty = torch.Tensor() </ins> target = template.size().tolist() self.assertEqual(tensor.view_as(template).size().tolist(), target) self.assertEqual(tensor.view(3, 5).size().tolist(), target)", "pos": ["Please check and fix all the PyTorch web links format from to , like in the official site.\nclosed via"], "neg": []}
{"query": "std::unique_ptr<Tensor> save_std(output->newTensor()); save_std->resizeAs(*running_var); <del> if (use_cudnn) { </del> <ins> if (use_cudnn && eps >= CUDNN_BN_MIN_EPSILON) { </ins> #ifdef WITH_CUDNN torch::cudnn::cudnn_batch_norm_forward( state,", "pos": ["Hi, I have two linux servers. My code works in one machine but not the other. In the one that it doesn't work, the error message is: occured during batch normalization( convolution is fine). Also, the same code works in CPU but not in GPU. I guess this problem is mainly related to GPU libraries like cuda or cudnn. The one that doesn't work has the following configurations: Nvidia driver 375.26 cuda 8.0.61 cudnn 5.1.10 While in the one it does work, I have the following configurations: Nvidia driver 367.48 cuda 8.0.44 cudnn 5.1.5 TensorFlow can run in the above environment flawless though. So I am wondering if this is caused by the incompatibility between PyTorch 0.1.12 and cuda 8.0.61/cudnn 5.1.10 ?\nyou use the pytorch binaries, we ship cudnn6 in them, use it it's great. 5.1.10 maybe has a bug. cc: . do you have a small test case to reproduce this? does not use CuDNN batchnorm by default, so maybe that's why?\nSmall test case would help, thanks.\nAfter some experimentation, I have pin pointed where exactly the error comes from. It seems batch norm in cudnn 5.1.10 doesn't support eps as small as 1e-6. A small test case would be:\nYes, cudnn.h for 5.1.10 defines CUDNNBNEPSILON to be 1e-5 (minimum allowable value). I can't check header for 5.1.5. Even with 1e-5 epsilon for many input data distributions the results become pretty unstable. Is there a case for this restriction to be lifted, does using very small epsilon provide benefits?\nI see. Not sure if small epsilon is necessary. I am using 1e-6 just to replicate others' configurations. Should PyTorch provide more informative error message about this restriction?\nIdeally, CUDNN should provide more informative error messages (pytorch can not know why cudnn is returning an error code in each particular case) , but that's a long standing discussion.\nIt just hits on me.", "As mentioned by , TensorFlow does not use CuDNN batchnorm by default. So is this possible in Pytorch, to not use default CuDNN batchnorm?\nNever mind. I found the switch in ."], "neg": []}
{"query": "#define IS_CUDA false #define CUDA_FLOAT false #else #define IS_CUDA true <del> #define CUDA_BYTE defined(THC_REAL_IS_BYTE) #define CUDA_CHAR defined(THC_REAL_IS_CHAR) #define CUDA_SHORT defined(THC_REAL_IS_SHORT) #define CUDA_INT defined(THC_REAL_IS_INT) #define CUDA_LONG defined(THC_REAL_IS_LONG) #define CUDA_FLOAT defined(THC_REAL_IS_FLOAT) #define CUDA_DOUBLE defined(THC_REAL_IS_DOUBLE) #define CUDA_HALF defined(THC_REAL_IS_HALF) </del> <ins> #if defined(THC_REAL_IS_BYTE) #define CUDA_BYTE 1 #else #define CUDA_BYTE 0 </ins> #endif <ins> #if defined(THC_REAL_IS_CHAR) #define CUDA_CHAR 1 #else #define CUDA_CHAR 0 #endif #if defined(THC_REAL_IS_SHORT) #define CUDA_SHORT 1 #else #define CUDA_SHORT 0 #endif #if defined(THC_REAL_IS_INT) #define CUDA_INT 1 #else #define CUDA_INT 0 #endif #if defined(THC_REAL_IS_LONG) #define CUDA_LONG 1 #else #define CUDA_LONG 0 #endif #if defined(THC_REAL_IS_FLOAT) #define CUDA_FLOAT 1 #else #define CUDA_FLOAT 0 #endif #if defined(THC_REAL_IS_DOUBLE) #define CUDA_DOUBLE 1 #else #define CUDA_DOUBLE 0 #endif #if defined(THC_REAL_IS_HALF) #define CUDA_HALF 1 #else #define CUDA_HALF 0 #endif #endif // ifndef THC_GENERIC_FILE </ins> #if IS_CUDA #define THIndexTensor THCudaLongTensor #define THIndexTensor_(NAME) TH_CONCAT_2(THCudaLongTensor_,NAME)", "pos": ["In we define a number of macros which are subsequently used in if statements: This is undefined. On clang 3.4.2, these warnings don't seem to actually get printed out () but on clang 4.0.0 they seem to always get emitted."], "neg": []}
{"query": "END_HANDLE_TH_ERRORS } <ins> // The type for PyInt_FromLong varies between Windows (int64_t) and // Linux (long).  Unfortunately, on Clang, if the template type argument // does not *exactly* line up, Clang discards the template // specialization.  So we write a little stub function which does have // the right type. static PyObject* PortablePyInt_FromLong(int64_t ival) { return PyInt_FromLong(ival); } </ins> static struct PyGetSetDef batch_norm_forward_properties[] = { THP_FUNCTION_DEFAULT_PROPERTIES, {(char*)\"running_mean\", (getter)getTensorAttr<BatchNormForward, BatchNormParams,", "pos": ["One representative error: The cause of the problem is : It seems that on clang, changing the type parameter here is sufficient to cause template instantiation to fail. Maybe the easiest way to fix this is to write a more portable version of PyIntFromLong (and friends) which always returns ."], "neg": []}
{"query": "('ones', small_3d, lambda t: [1, 2, 3, 4, 5],), ('permute', new_t(1, 2, 3, 4), lambda t: [2, 1, 3, 0],), ('put_', new_t(2, 5, 3), lambda t: [long_type(t)([[0], [-2]]), t([[3], [4]])],), <ins> ('put_', new_t(2, 3), lambda t: [long_type(t)([]), t([])], 'empty'), </ins> ('put_', new_t(2, 2), lambda t: [long_type(t)([[1], [-3]]), t([[1], [2]]), True], 'accumulate'), ('prod', small_2d_oneish, lambda t: [],), ('prod', small_3d, lambda t: [1], 'dim'),", "pos": ["odd difference between cuda Tensor and Variable: encountered in torch.distributions.Categorical, built yesterday from master\nrelated to hmm, here's the issue is for variables, so probably not (immediately) related\ndifferent code path but similar issue\nThis is a regression from the latest 0.3 binaries, which don't have this error."], "neg": []}
{"query": "endif() # ---[ Exract versions <del> message(STATUS \"Caffe2: CUDA detected: \" ${CUDA_VERSION}) </del> if (CAFFE2_FOUND_CUDNN) # Get cuDNN version file(READ ${CUDNN_INCLUDE_DIR}/cudnn.h CUDNN_HEADER_CONTENTS)", "pos": ["Here's the situation: Ideally, you'd say something like to make a build of PyTorch with a non-standard version of CUDA. When this happens, needs to find the correct copy of nvcc to use. If you are using ccache as per the instructions in you'll have set to your ccache binary, so that ccache happens. cmake will correctly respect this variable. However, here's the problem: ccache's nvcc will use your path to determine what the real nvcc to use is, this means that you'll pick up whatever nvcc is in your PATH, which is probably not the one you want, because you're trying to use a different nvcc. So in sum, you have to do the non-intuitive to get the whole shebang to work. The worst part is that it won't tell you if you've got it wrong, it will happily use the wrong nvcc. I haven't gone through user submitted issues to see if anyone has succumbed to this particular bug.\nA long-term solution might be getting to work\nworks now."], "neg": []}
{"query": "from functools import reduce from ._utils import _range <ins> class __PrinterOptions(object): precision = 4 threshold = 1000 edgeitems = 3 linewidth = 80 PRINT_OPTS = __PrinterOptions() </ins> SCALE_FORMAT = '{:.5e} *n' <del> def _number_format(tensor): min_sz = 2 </del> <ins> # We could use **kwargs, but this will give better docs def set_printoptions( precision=None, threshold=None, edgeitems=None, linewidth=None, profile=None, ): \"\"\"Set options for printing. Items shamelessly taken from Numpy Args: precision: Number of digits of precision for floating point output (default 8). threshold: Total number of array elements which trigger summarization rather than full repr (default 1000). edgeitems: Number of array items in summary at beginning and end of each dimension (default 3). linewidth: The number of characters per line for the purpose of inserting line breaks (default 80). Thresholded matricies will ignore this parameter. profile: Sane defaults for pretty printing. Can override with any of the above options. (default, short, full) \"\"\" if profile is not None: if profile == \"default\": PRINT_OPTS.precision = 4 PRINT_OPTS.threshold = 1000 PRINT_OPTS.edgeitems = 3 PRINT_OPTS.linewidth = 80 elif profile == \"short\": PRINT_OPTS.precision = 2 PRINT_OPTS.threshold = 1000 PRINT_OPTS.edgeitems = 2 PRINT_OPTS.linewidth = 80 elif profile == \"full\": PRINT_OPTS.precision = 4 PRINT_OPTS.threshold = float('inf') PRINT_OPTS.edgeitems = 3 PRINT_OPTS.linewidth = 80 if precision is not None: PRINT_OPTS.precision = precision if threshold is not None: PRINT_OPTS.threshold = threshold if edgeitems is not None: PRINT_OPTS.edgeitems = edgeitems if linewidth is not None: PRINT_OPTS.linewidth = linewidth def _number_format(tensor, min_sz=-1): min_sz = max(min_sz, 2) </ins> tensor = torch.DoubleTensor(tensor.nelement()).copy_(tensor).abs_() pos_inf_mask = tensor.eq(float('inf'))", "pos": ["It seems that the following link that can be reached from here (check the screenshot) is broken: Page: Broken link: Screenshot: ! cc\nthis should be transferred to , but I don't have permissions to do that. Fix is available at pytorch/hub."], "neg": []}
{"query": "msg='LR is wrong in epoch {}: expected {}, got {}'.format( epoch, target[epoch], param_group['lr']), delta=1e-5) if __name__ == '__main__': run_tests()", "pos": ["I'm working on checkpointing in our nn trainer. When saving the state objects, I get an error: PicklingError: Can't pickle <type 'function': attribute lookup builtin.function failed I tracked it down to trying to save the scheduler ReduceLROnPlateau. I think the issue is due to the scheduler referencing the optimizer, which has to be saved with statedict(). I presume the same error will occur with other schedulers too. It seems like the problem could be solved by providing a statedict() and loadstatedict() in the scheduler classes. pytorch master rev (Jan 22) installed from source x86_64 on centos 7.4 cuda 9.0 cudnn 7 k80 gcc 5.4 python 2.7.14\nThe following script reproduces the crash: import torch import from import ReduceLROnPlateau if name == 'main': model = (10,10) optim = (model.parameters()) sched = ReduceLROnPlateau(optim, mode='max', factor=0.5) with open('','wb') as f: # Works # ((model.statedict(), optim.statedict()), f) # Crashes ((model.statedict(), optim.state_dict(), sched), f)"], "neg": []}
{"query": "python_name: apply_ defined_if: \"!IS_CUDA\" only_register: True <ins> override_method_flags: METH_O </ins> ]] static PyObject * THPTensor_(apply)(THPTensor *self, PyObject *arg) {", "pos": ["On Python 3.5.2 errors out with . map and map2 seem to be work just fine."], "neg": []}
{"query": "import os import torch import traceback <ins> import warnings </ins> from torch._six import raise_from from subprocess import Popen, PIPE from multiprocessing.util import register_after_fork as _register_after_fork", "pos": ["There's quite a few user-reports of installing the CUDA 7.5 binary of pytorch and then running on GTX 1080 / 1070, and there will be reports of installing 7.5 / 8.0 binaries and running on V100. Raising a user-warning in such situations and letting them know will fix the problem.\nThe version of CUDA used to compile the binary should be available thanks to this PR:\nI think is a more robust way of getting the CUDA used to compile PyTorch, as currently if there are multiple CUDA installed in the machine, the heuristics used in might not match those used by CMake FindCUDA. I think that we should modify the to return the cuda version used during compilation. I have the impression that it will not require changing much because doesn't require cuda to execute. What do you think?"], "neg": []}
{"query": "composed of several input planes. This module can be seen as the gradient of Conv1d with respect to its input. <del> It is sometimes (but incorrectly) refered to as a deconvolutional operation. </del> <ins> It is also known as a fractionally-strided convolution or a deconvolution (although it is not an actual deconvolution operation). | :attr:`stride` controls the stride for the cross-correlation. | If :attr:`padding` is non-zero, then the input is implicitly zero-padded on both sides for :attr:`padding` number of points. | If :attr:`output_padding` is non-zero, then the output is implicitly zero-padded on one side for :attr:`output_padding` number of points. | :attr:`dilation` controls the spacing between the kernel points; also known as the \u00e0 trous algorithm. It is harder to describe, but this `link`_ has a nice visualization of what :attr:`dilation` does. | :attr:`groups` controls the connections between inputs and outputs. `in_channels` and `out_channels` must both be divisible by `groups`. |       At groups=1, all inputs are convolved to all outputs. |       At groups=2, the operation becomes equivalent to having two conv layers side by side, each seeing half the input channels, and producing half the output channels, and both subsequently concatenated. At groups=`in_channels`, each input channel is convolved with its own set of filters (of size `out_channels // in_channels`). </ins> .. note::", "pos": ["Hi, I have installed two versions of CUDA. But I want to use one of it, can I specify it during setup process?\nTry running\nI have tried it, () still returns False. Thx for your reply!\nDo you have the cuda driver installed?\nOf course I do. It raise assertion error which shows \"The NVIDIA driver on your system is too old(found version 8000)\".\nthat's the answer to your problem. upgrade your driver.\nThx!", "When building from source, I got this error. My system is OSX 10.12.4, conda, CUDA 8.0. Any thoughts? Thanks!\ncan you try adding to the file and see if that fixes things? if so i'll commit it.\nactually, that file already has memory included. Are you following instructions from Especially: This part is important:\nworks nicely! Thanks a lot!", "The suggest that takes a as a parameter, but it actually must be wrapped in a Variable first.\nyou are looking at the Tensor docs. When operating on Variable, it is implied from the tutorials that all functions operating on Variables have to use Variables. We didn't write separate docs for Variable itself.", "something wrong with import torchvision Traceback (most recent call last): File \"\", line 3, in <moduleimport torchvision File \"\", line 4, in <moduleimport torchvision.transforms as transforms ImportError: No module named transforms\nyou have a file called for some reason. delete it or change your pythonpath to ignore it.\nHey, I have the same problem persisting when I import torchvision. I have no other file by the same name. ImportError Traceback (most recent call last) <ipython-input-6-3baf6837d8f1in <module() 8 from import Dataset 9 from import DataLoader ---10 from torchvision import transforms ImportError: No module named torchvision Any idea why this is happening?\nDo you have torchvision installed (it is a separate package)?\npip Install torchvision try this\nfrom data import BaseTransform,VOC_CLASSES as labelmap this is not working\npip install torchvision", "Hi, I need to pass input through one nn.Module, then argmax the output, and then pass that output to a second nn.Module. I want to backprop through the argmax back to the weights of the first module. That's impossible right now, though, it seems. When I loop through the gradient of the weights of my first module (with the following code): firstmodel.zerograd() secondmodel.zerograd() loss.backward() # retainvariables=True for param in secondmodel.parameters(): print( 'data: ', ) print( 'grad data: ', ) for param in firstmodel.parameters(): print( 'data: ', ) print( 'grad data: ', ) I get the following error, stating that those gradients can no longer be viewed/exist: -main()\nargmax is not differentiable (and even if you try to think about it as a function differentiable almost everywhere, its derivative is always 0, so you won't get anything meaningful).", "def inithidden(self): h0 = Variable(FloatTensor(self.numlayers, self.batchlen, self.hiddendim).zero(),  requiresgrad=True) c0 = Variable(FloatTensor(self.numlayers, self.batchlen, self.hiddendim).zero(),  requiresgrad=True) return (h0, c0) def forward(self, sequences): lstmout, = ((1, self.batchlen, self.inputdim), ) ... = nn.LSTM(inputdim, hiddendim, numlayers, dropout = dropout, batchfirst=True) = self.inithidden() output = model(traininput) produces the following error although I defined weight and input dimensions as in the documentation: File \"/usr/local/lib/python2.7/dist-\", line 202, in call result = self.forward(input, kwargs) File \"/usr/local/lib/python2.7/dist-\", line 91, in forward output, hidden = func(input, self.allweights, hx) File \"/usr/local/lib/python2.7/dist-\", line 327, in forward return func(input, fargs, fkwargs) File \"/usr/local/lib/python2.7/dist-\", line 227, in forward nexth, output = func(input, hidden, weight) File \"/usr/local/lib/python2.7/dist-\", line 67, in forward hy, output = inner(input, hidden[l], weight[l]) File \"/usr/local/lib/python2.7/dist-\", line 96, in forward hidden = inner(input[i], hidden, *weight) File \"/usr/local/lib/python2.7/dist-\", line 22, in LSTMCell gates = F.linear(input, wih, bih) + F.linear(hx, whh, bhh) File \"/usr/local/lib/python2.7/dist-\", line 744, in add return (other) File \"/usr/local/lib/python2.7/dist-\", line 293, in add return (other, False) File \"/usr/local/lib/python2.7/dist-\", line 287, in add return Add(inplace)(self, other) File \"/usr/local/lib/python2.7/dist-\", line 20, in forward return a.add(b) RuntimeError: inconsistent tensor size at /data/users/soumith/builder/wheel/pytorch-src/torch/lib/TH/generic/THTensorMath.c:827\nI can't reproduce your problem. Could you include a short runnable example that exhibits this behavior?", "You also don't need to manually init hidden anymore, the following would work as well...\nThanks! After removing problem solved.", "Hi, I would like to implement a sharing weight 'alexnet' with two input batches of images, and the fully connected layers are concatenated together like this: However, when I run it, I come across the following error: Does anyone know about it? Thank you very much.\nYou forgot to call on the output of . In the future please remember that we're using the issues for bug reports and feature requests only. If you have questions or need help, please post ."], "neg": []}
{"query": "@wraps(fn) def wrapper(self): if self.rank == self.MANAGER_PROCESS_RANK: <del> self._join_and_reduce() </del> <ins> self._join_and_reduce(fn) </ins> else: fn(self) return wrapper", "pos": ["is causing a crash when we try to build pytorch. Crash looks like this: (all times in US/West) Our 7AM, 8AM builds and earlier were all fine but 9AM, 10AM, 11AM builds all crashed with that same traceback. The new import was in which I think went in between 8AM and 9AM today See command below: See here if needed for : <!-- A clear and concise description of what you expected to happen. --Please copy and paste the output from our (or fill out the checklist below manually). You can get the script and run it with: PyTorch Version (e.g., 1.0): building from master OS (e.g., Linux): linux How you installed PyTorch (, , source): source Build command you used (if compiling from source): see above under 'reproduce' Python version: 3.6\nmaybe easier to read for build command used:\nplease try in the CircleCi environment?\nin our build script seems to have fixed this. Thanks !"], "neg": []}
{"query": "def _initialize_backend(): from .._functions.thnn import _all_functions as _thnn_functions from .._functions.linear import Linear <del> from .._functions.batchnorm import BatchNorm </del> from .._functions.conv import ConvNd from .._functions.rnn import RNN,  RNNTanhCell, RNNReLUCell, GRUCell, LSTMCell", "pos": ["As Arthur Szlam reports, fb-internal cudnn is still lagging behind, and giving batch-size 1024 with batchnorm is raising an error from cudnn. Need to check for compile-time version and disable this codepath\nThe 1024 limitation was removed in 5.1.10"], "neg": []}
{"query": ".. automethod:: fmod_ .. automethod:: frac .. automethod:: frac_ <del> .. automethod:: free </del> .. automethod:: gather .. automethod:: ge .. automethod:: ge_", "pos": ["Dataloader call getitem as many times as indices in the current batch. In case datasets support a list of indices in one call, or a native python slice object, add a getbatch (optional) to pytorch dataset class. If dataloader sees the dataset has such a method implemented, make it fetch the batch corresping to the list of indices with one getbatch call rather than many getitem calls. This saves the trouble of calling getitem many times in cases where for example a disk file has to be opened then closed for every call and the underlying data structure supports a slicing or lists of indices for example hdf5 files.\nwill support something similar\nis the latest version of this."], "neg": []}
{"query": "else: sz = exp_max + 6 sz = max(min_sz, sz) <del> format = '{:' + str(sz) + '.4f}' </del> <ins> format = '{{:{}.{}f}}'.format(sz, prec) </ins> return format, scale, sz def _tensor_str(self): <ins> n = PRINT_OPTS.edgeitems has_hdots = self.size()[-1] > 2*n has_vdots = self.size()[-2] > 2*n print_full_mat = not has_hdots and not has_vdots formatter = _number_format(self, min_sz=3 if not print_full_mat else 0) print_dots = self.numel() >= PRINT_OPTS.threshold dim_sz = max(2, max(len(str(x)) for x in self.size())) dim_fmt = \"{:^\" + str(dim_sz) + \"}\" dot_fmt = u\"{:^\" + str(dim_sz+1) + \"}\" </ins> counter_dim = self.ndimension() - 2 counter = torch.LongStorage(counter_dim).fill_(0) <del> counter[0] = -1 </del> <ins> counter[counter.size()-1] = -1 </ins> finished = False strt = '' while True: <del> for i in _range(counter_dim): </del> <ins> nrestarted = [False for i in counter] nskipped = [False for i in counter] for i in _range(counter_dim - 1, -1, -1): </ins> counter[i] += 1 <ins> if print_dots and counter[i] == n and self.size(i) > 2*n: counter[i] = self.size(i) - n nskipped[i] = True </ins> if counter[i] == self.size(i): <del> if i == counter_dim-1: </del> <ins> if i == 0: </ins> finished = True counter[i] = 0 <ins> nrestarted[i] = True </ins> else: break if finished: break <ins> elif print_dots: if any(nskipped): for hdot in nskipped: strt += dot_fmt.format('...') if hdot  else dot_fmt.format('') strt += 'n' if any(nrestarted): strt += ' ' for vdot in nrestarted: strt += dot_fmt.format(u'u22EE' if vdot else '') strt += 'n' </ins> if strt != '': <del> strt += 'n' strt += '({},.,.) = n'.format(','.join(str(i) for i in counter)) submatrix = reduce(lambda t,i: t.select(0, i), counter, self) strt += _matrix_str(submatrix, ' ') </del> <ins> strt += 'n' strt += '({},.,.) = n'.format( ','.join(dim_fmt.format(i) for i in counter)) submatrix = reduce(lambda t, i: t.select(0, i), counter, self) strt += _matrix_str(submatrix, ' ', formatter, print_dots) </ins> return strt <del> def _matrix_str(self, indent=''): fmt, scale, sz = _number_format(self) nColumnPerLine = int(math.floor((80-len(indent))/(sz+1))) </del> <ins> def __repr_row(row, indent, fmt, scale, sz, truncate=None): if truncate is not None: dotfmt = \" {:^5} \" return (indent + ' '.join(fmt.format(val/scale) for val in row[:truncate]) + dotfmt.format('...') + ' '.join(fmt.format(val/scale) for val in row[-truncate:]) + 'n') else: return indent + ' '.join(fmt.format(val/scale) for val in row) + 'n' def _matrix_str(self, indent='', formatter=None, force_truncate=False): n = PRINT_OPTS.edgeitems has_hdots = self.size(1) > 2*n has_vdots = self.size(0) > 2*n print_full_mat = not has_hdots and not has_vdots if formatter is None: fmt, scale, sz = _number_format(self, min_sz=5 if not print_full_mat else 0) else: fmt, scale, sz = formatter nColumnPerLine = int(math.floor((PRINT_OPTS.linewidth-len(indent))/(sz+1))) </ins> strt = '' firstColumn = 0 <del> while firstColumn < self.size(1): lastColumn = min(firstColumn + nColumnPerLine - 1, self.size(1)-1) if nColumnPerLine < self.size(1): strt += 'n' if firstColumn != 1 else '' strt += 'Columns {} to {} n{}'.format(firstColumn, lastColumn, indent) </del> <ins> if not force_truncate and  (self.numel() < PRINT_OPTS.threshold or print_full_mat): while firstColumn < self.size(1): lastColumn = min(firstColumn + nColumnPerLine - 1, self.size(1)-1) if nColumnPerLine < self.size(1): strt += 'n' if firstColumn != 1 else '' strt += 'Columns {} to {} n{}'.format( firstColumn, lastColumn, indent) if scale != 1: strt += SCALE_FORMAT.format(scale) for l in _range(self.size(0)): strt += indent + (' ' if scale != 1 else '') row_slice = self[l, firstColumn:lastColumn+1] strt += ' '.join(fmt.format(val/scale) for val in row_slice) strt += 'n' firstColumn = lastColumn + 1 else: </ins> if scale != 1: strt += SCALE_FORMAT.format(scale) <del> for l in _range(self.size(0)): strt += indent + (' ' if scale != 1 else '') row_slice = self[l, firstColumn:lastColumn+1] strt += ' '.join(fmt.format(val/scale) for val in row_slice) + 'n' firstColumn = lastColumn + 1 </del> <ins> if has_vdots and has_hdots: vdotfmt = \"{:^\" + str((sz+1)*n-1) + \"}\" ddotfmt = u\"{:^5}\" for row in self[:n]: strt += __repr_row(row, indent, fmt, scale, sz, n) strt += indent + ' '.join([vdotfmt.format('...'), ddotfmt.format(u'u22F1'), vdotfmt.format('...')]) + \"n\" for row in self[-n:]: strt += __repr_row(row, indent, fmt, scale, sz, n) elif not has_vdots and has_hdots: for row in self: strt += __repr_row(row, indent, fmt, scale, sz, n) elif has_vdots and not has_hdots: vdotfmt = u\"{:^\" +  str(len(__repr_row(self[0], '', fmt, scale, sz))) +  \"}n\" for row in self[:n]: strt += __repr_row(row, indent, fmt, scale, sz) strt += vdotfmt.format(u'u22EE') for row in self[-n:]: strt += __repr_row(row, indent, fmt, scale, sz) else: for row in self: strt += __repr_row(row, indent, fmt, scale, sz) </ins> return strt def _vector_str(self): <del> fmt, scale, _ = _number_format(self) </del> <ins> fmt, scale, sz = _number_format(self) </ins> strt = '' ident = '' <ins> n = PRINT_OPTS.edgeitems dotfmt = u\"{:^\" + str(sz) + \"}n\" </ins> if scale != 1: strt += SCALE_FORMAT.format(scale) ident = ' ' <del> return strt + 'n'.join(ident + fmt.format(val/scale) for val in self) + 'n' </del> <ins> if self.numel() < PRINT_OPTS.threshold: return (strt + 'n'.join(ident + fmt.format(val/scale) for val in self) + 'n') else: return (strt + 'n'.join(ident + fmt.format(val/scale) for val in self[:n]) + 'n' + (ident + dotfmt.format(u\"u22EE\")) + 'n'.join(ident + fmt.format(val/scale) for val in self[-n:]) + 'n') </ins> def _str(self):", "pos": ["It seems that the following link that can be reached from here (check the screenshot) is broken: Page: Broken link: Screenshot: ! cc\nthis should be transferred to , but I don't have permissions to do that. Fix is available at pytorch/hub."], "neg": []}
{"query": "The NVIDIA driver on your system is too old (found version {}). Please update your GPU driver by downloading and installing a new version from the URL: http://www.nvidia.com/Download/index.aspx <del> Alternatively, go to: https://pytorch.org/binaries to install </del> <ins> Alternatively, go to: http://pytorch.org to install </ins> a PyTorch version that has been compiled with your version of the CUDA driver.\"\"\".format(str(torch._C._cuda_getDriverVersion()))) <ins> def _check_capability(): error_str = \"\"\" Found GPU%d %s which requires CUDA_VERSION >= %d for optimal performance and fast startup time, but your PyTorch was compiled with CUDA_VERSION %d. Please install the correct PyTorch binary using instructions from http://pytorch.org \"\"\" CUDA_VERSION = torch._C._cuda_getCompiledVersion() for d in range(device_count()): major = get_device_capability(d)[0] name = get_device_name(d) if CUDA_VERSION < 8000 and major >= 6: warnings.warn(error_str % (d, name, 8000, CUDA_VERSION)) elif CUDA_VERSION < 9000 and major >= 7: warnings.warn(error_str % (d, name, 8000, CUDA_VERSION)) </ins> def _lazy_call(callable): if _initialized: callable()", "pos": ["There's quite a few user-reports of installing the CUDA 7.5 binary of pytorch and then running on GTX 1080 / 1070, and there will be reports of installing 7.5 / 8.0 binaries and running on V100. Raising a user-warning in such situations and letting them know will fix the problem.\nThe version of CUDA used to compile the binary should be available thanks to this PR:\nI think is a more robust way of getting the CUDA used to compile PyTorch, as currently if there are multiple CUDA installed in the machine, the heuristics used in might not match those used by CMake FindCUDA. I think that we should modify the to return the cuda version used during compilation. I have the impression that it will not require changing much because doesn't require cuda to execute. What do you think?"], "neg": []}
{"query": "\"torch/csrc/cuda/Module.cpp\", \"torch/csrc/cuda/Storage.cpp\", \"torch/csrc/cuda/Stream.cpp\", <del> \"torch/csrc/cuda/Tensor.cpp\", </del> \"torch/csrc/cuda/AutoGPU.cpp\", \"torch/csrc/cuda/utils.cpp\", \"torch/csrc/cuda/serialization.cpp\", <ins> *split_types(\"torch/csrc/cuda/Tensor.cpp\"), </ins> ] if WITH_NCCL:", "pos": ["Trying to run on github actions via AWS instance gives me: I will of course set up now, but it shouldn't crash like that. May be it should use as a fallback? pt-2.0 cc\nIs this a multitenant setup? Also curious which distro are you running, I was under the impression that most linux distros set home by default\nthis is AWS image. one default user. I think it's user, but it wants you to run as root, which is weird. I have manually now set to overcome this.\nWe should update triton to use pathlib or something"], "neg": []}
{"query": "pip_packages=pip_list_output, conda_packages=get_conda_packages(run_lambda), os=get_os(run_lambda), <ins> libc_version=get_libc_version(), </ins> gcc_version=get_gcc_version(run_lambda), clang_version=get_clang_version(run_lambda), cmake_version=get_cmake_version(run_lambda),", "pos": ["TensorPipe uses some advanced features of the Linux kernel. It does so on a \"best effort\" basis (i.e., use them if available, fall back on something else if not). This logic is unfortunately hard to test because it would require specific kernel versions on our CI machines. So it occasionally happens that we find out about issues when users report them. To help investigate these reports we've found it useful to know the kernel version and the glibc version used by the user (both when building and when running, if they're different). It would be great if these could be to the report produced by collect_env.\nCC"], "neg": []}
{"query": "def cross(self, other, dim=-1): return Cross(dim)(self, other) <ins> def inverse(self): return Inverse.apply(self) </ins> def multinomial(self, num_samples=1, with_replacement=False): return Multinomial(num_samples, with_replacement)(self)", "pos": ["Command to install the latest nightly binary: installs : However, PyTorch itself reports : I guess this version tag might be defined during the build process from e.g. and then forwarded from , but also source builds should use based on the tag. Unsure, if this is just a reporting issue or if we are building the nightly binaries with and executing them with , so creating this issue to dig a bit further into it. CC cc\nThis is really weird, I've put up measure in place to validate against it... Will investigate today (and moving to pytorch repo)\ncan you please check that 2.2.0 reports (and is build) against the correct version of nccl library?\nConfirmed for nightlies : Same issue for RC:\nLooking at the build log Looks like we are compiling with: USESYSTEMNCCL\nOn Windows I see following exception: We compiling with should we return None rather then exception ?\nIn this case we are indeed using 2.18.1: Do you know why we are using the system-wide NCCL installation instead of the one? Should we delete the system NCCL and use by default?\nLooks like its set here: Looks like it was by: Let me propose the PR to unset this flag : USESYSTEM_NCCL\nI guess we are using the system NCCL installation to be able to during the runtime. If we would use the module I assume we would link NCCL into the PyTorch binaries. If so, we should make sure to update the NCCL version whenever is updated. This could break our dependency on the NCCL pip package.\nYes. Also when we update this nccl here:\nWe used following path  What would be good path for nccl 2.19.3 and cuda 12.1 ? As per this issue: Will be using 12.0 ncck 2.19.3\nInstalling system NCCL into the nightly build docker container feels great, but we need to figure out, how to keep it in sync with PyTorch. I wonder if one can, for example, just check nccl version from this requirements file and build from source?\nyes probably we should build and upload nccl for our supported cuda versions, same way as we build docker containers, by following details here:\nVerified with final rc"], "neg": []}
{"query": "if memo is None: memo = set() if self not in memo: <del> super(Container, self).modules(memo) </del> <ins> for m in super(Container, self).modules(memo): yield m </ins> for module in self.children(): for m in module.modules(memo): yield m", "pos": ["While using torch on a Macbook pro with M1 ship, using python 3.8.10 and torch 1.12.1, I came across a weird bug that happened because I forgot to put a tensor on MPS. The reproducible code is the following: As you can see, if is on MPS and on CPU, the line wrongly replaces the index in place of the index . I would expect to have an error saying both tensors should be on the same device instead. I get an error: cc\nThe error seems to happen somewhere within Also I'm getting a warning \"The operator 'aten::nonzero' is not currently supported on the MPS backend and will fall back to run on the CPU. This may have performance implications\" when printing tensors on MPS.\nHi - thanks for reporting this issue. This is actually an allowed usecase (copying CPU tensors into MPS tensors / MPS -CPU / MPS -MPS): The issue is on MPS side when copying a CPU tensor into a strided MPS tensor. I'll take a look."], "neg": []}
{"query": "#include \"THCTensorTypeUtils.cuh\" #include \"THCHalfAutoNumerics.cuh\" #include \"SharedMem.cuh\" <ins> #include <algorithm> </ins> //////////////////////////////////////////////////////////////////////////////// // Spatial kernel (fast with large inner_size and small dim_size) ////////////////////////////////////////////////////////////////////////////////", "pos": ["mac 10.12.6 xcode 8.3.3 clang Apple LLVM version 8.0.0 (clang-800.0.42.1)\nsame here. Adding fixes it.\nthanks! ! building now...\nsuccess!yeah! thanks again."], "neg": []}
{"query": "def backward(ctx, grad_output): inverse, = ctx.saved_variables return -torch.mm(inverse.t(), torch.mm(grad_output, inverse.t())) <ins> class Gesv(Function): @staticmethod def forward(ctx, b, a): # TODO see if one can backprop through LU X, LU = torch.gesv(b, a) ctx.save_for_backward(X, a) ctx.mark_non_differentiable(LU) return X, LU @staticmethod def backward(ctx, grad_output, grad_LU=None): X, a = ctx.saved_variables grad_b, _ = torch.gesv(grad_output, a.t()) grad_a = -torch.mm(grad_b, X.t()) return grad_b, grad_a </ins>", "pos": ["There seems to be regression in the dynamo export of the MONAI VNet model as regards . Export script: The versions I'm using is torch 2.5.0.dev20240625+cu121 Previous torch versions such as were able to export this model successfully. Perhaps dynamo was generating FX graph without the expm1 node previously? cc: PyTorch version: 2.5.0.dev20240625+cu121 Is debug build: False CUDA used to build PyTorch: 12.1 ROCM used to build PyTorch: N/A OS: Ubuntu 22.04.4 LTS (x8664) GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 Clang version: Could not collect CMake version: version 3.29.0 Libc version: glibc-2.35 Python version: 3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0] (64-bit runtime) Python platform: Linux-6.5.0-35-generic-x8664-with-glibc2.35 Is CUDA available: True CUDA runtime version: 12.4.131 CUDAMODULELOADING set to: LAZY GPU models and configuration: GPU 0: NVIDIA RTX A5000 Nvidia driver version: 535.171.04 cuDNN version: Probably one of the following: /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- /usr/lib/x8664-linux- HIP runtime version: N/A MIOpen runtime version: N/A Is XNNPACK available: True CPU: Architecture: x8664 CPU op-mode(s): 32-bit, 64-bit Address sizes: 46 bits physical, 48 bits virtual Byte Order: Little Endian CPU(s): 20 On-line CPU(s) list: 0-19 Vendor ID: GenuineIntel Model name: Intel(R) Core(TM) i9-9820X CPU @ 3.", "30GHz CPU family: 6 Model: 85 Thread(s) per core: 2 Core(s) per socket: 10 Socket(s): 1 Stepping: 4 CPU max MHz: 4200.0000 CPU min MHz: 1200.0000 BogoMIPS: 6599.98 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constanttsc art archperfmon pebs bts repgood nopl xtopology nonstoptsc cpuid aperfmperf pni pclmulqdq dtes64 monitor dscpl vmx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse41 sse42 x2apic movbe popcnt tscdeadlinetimer aes xsave avx f16c rdrand lahflm abm 3dnowprefetch cpuidfault epb catl3 cdpl3 invpcidsingle pti ssbd mba ibrs ibpb stibp tprshadow flexpriority ept vpid eptad fsgsbase tscadjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm mpx rdta avx512f avx512dq rdseed adx smap clflushopt clwb intelpt avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqmllc cqmoccupllc cqmmbmtotal cqmmbmlocal dtherm ida arat pln pts hwp hwpactwindow hwpepp hwppkgreq vnmi mdclear flushl1d archcapabilities Virtualization: VT-x L1d cache: 320 KiB (10 instances) L1i cache: 320 KiB (10 instances) L2 cache: 10 MiB (10 instances) L3 cache: 16.5 MiB (1 instance) NUMA node(s): 1 NUMA node0 CPU(s): 0-19 Vulnerability Gather data sampling: Mitigation;", "Microcode Vulnerability Itlb multihit: KVM: Mitigation: VMX disabled Vulnerability L1tf: Mitigation; PTE Inversion; VMX conditional cache flushes, SMT vulnerable Vulnerability Mds: Mitigation; Clear CPU buffers; SMT vulnerable Vulnerability Meltdown: Mitigation; PTI Vulnerability Mmio stale data: Mitigation; Clear CPU buffers; SMT vulnerable Vulnerability Retbleed: Mitigation; IBRS Vulnerability Spec rstack overflow: Not affected Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and _user pointer sanitization Vulnerability Spectre v2: Mitigation; IBRS; IBPB conditional; STIBP conditional; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Mitigation; Clear CPU buffers; SMT vulnerable Versions of relevant libraries: [pip3] numpy==1.24.4 [pip3] onnx==1.16.0 [pip3] onnxscript==0.1.0.dev20240625 [pip3] optree==0.11.0 [pip3] pytorch-quantization==2.1.2 [pip3] pytorch-triton==3.0.0+ [pip3] torch==2.5.0.dev20240625+cu121 [pip3] torch-tensorrt==2.3.0a0 [pip3] torchaudio==2.4.0.dev20240625+cu121 [pip3] torchdata==0.7.1a0 [pip3] torchtext==0.17.0a0 [pip3] torchvision==0.20.0.dev20240625+cu121 [conda] Could not collect\nI am not very sure, but this looks like the kind of ops that should get decomp'ed. do you know? could you help take a look as part of the op support work you are doing? Thanks!\nYes, it looks like there is a graph-changed in dynamo, which dynamo now decomposes elu for some reasons. Using can preserve elu and then successfully export though.", "Yes can run but we would need to be resolved first since the weights are still not exported properly.\nI can take a look at this.\nshould fix this.", "Here, I can't export trivial net with sort() using dynamoexport(). Legacy () works fine, dynamoexport errors out : Unsupported FX nodes: {'callfunction': ['']}. Pytorch nightly 06/03 cc\na missing op, could you take this?\nNot sure if related:\nHmmm. Ideally, ONNX should add a sort op. It doesn't have one. I wonder what the legacy exporter converts it to.\nshould fix this issue.", "Are there any plans to implement something like tensorflow's and ? exists but it doesn't have the backward step implemented.\nOne workaround for the moment would be to write as , which should support backprop since . Adding support for should be trivial, the gradient for it can be found in\nI'll add backward support for , which should be equivalent to . I'll try to do it later today.\nHow difficult would it be to add in batched versions (where only the final two dimensions are square) of GESV and inverse? Are there any plans for this?\nIf we have batched version of , then adapting the autograd part is very simple. There is an open issue already that discuss implementing batched gesv\nActually, there is an open PR for that, so it might be coming soon\nI'll rebase the batched version of gesv. It'll come in soon\nMay I ask why gesv only support . As I saw from , it support that A is a MxN matrix, and B is a NxK matrix, then solve a matrix X (NxK). If I want to use a non-square matrix A in gesv, would you mind to let me know how to implement in PyTorch?\nthis doesn't exactly answer your question but you can add zero-rows to the A matrix until it becomes square.\nthanks a lot, I find the solution, can solve the problem.\nAny plan to add gradient calculation for gels?\nI am also wondering if will have differentiation support. My current workaround is to use instead\nI actually implemented overdetermined case following tensorflow implementation. For backward pass I only did basic grad test. On Monday, March 18, 2019, Johnson Zhong <:\nThanks, seems to work fine!"], "neg": []}
{"query": "return grad_input <del> # TODO: gather </del> <ins> class Gather(Function): def __init__(self, dim): super(Gather, self).__init__() self.dim = dim def forward(self, input, index): assert not self.needs_input_grad[1], \"Gather can't differentiate \"  \"the index\" self.input_size = input.size() self.save_for_backward(index) return input.gather(self.dim, index) def backward(self, grad_output): index, = self.saved_tensors grad_input = grad_output.new(self.input_size).zero_() return grad_input.scatter_(self.dim, index, grad_output), None class Scatter(InplaceFunction): def __init__(self, dim, inplace=False): super(Scatter, self).__init__(inplace) self.dim = dim def forward(self, input, index, source): assert not self.needs_input_grad[1], \"Scatter can't differentiate \"  \"the index\" if self.inplace: self.mark_dirty(input) else: input = input.clone() self.save_for_backward(index) return input.scatter_(self.dim, index, source) def backward(self, grad_output): index, = self.saved_tensors grad_input = grad_source = None if self.needs_input_grad[0]: grad_input = grad_output.clone() grad_input.scatter_(self.dim, index, 0) if self.needs_input_grad[2]: grad_source = grad_output.gather(self.dim, index) return grad_input, None, grad_source </ins> # TODO: kthvalue # TODO: repeat # TODO: sort", "pos": ["<!-- A clear and concise description of what content in is an issue. If this has to do with the general website, please file an issue at instead. If this has to do with , please file an issue at --In the getting started section section the following quote appears: I guess this is from previous versions where CUDA support was not built-in and this can be removed, correct? cc\nThis seems to be easy fixable docs issue\nIf gets merged this is fixed.\nThanks for the PR Just merged it."], "neg": []}
{"query": "self.assertEqual(a[()], a) self.assertEqual(a[()].data_ptr(), a.data_ptr()) <del> # @unittest.skip('failing') </del> def test_empty_fancy_index(self): # Empty list index creates an empty array a = tensor([1, 2, 3])", "pos": ["In PyTorch master 0.4 , variables cannot be indexed. This used to work in PyTorch 0.2 :\nThis is breaking some of Pyro's inference algorithms on PyTorch master:"], "neg": []}
{"query": "def _test_sparse_mask_hybrid_fixed(self): i = self.IndexTensor([ <del> [1, 3, 3, 0, 4], [2, 1, 1, 2, 3], </del> <ins> [1, 3, 0, 4], [2, 1, 2, 3], </ins> ]) <del> v = self.ValueTensor([[1, 2], [2, 3], [3, 4], [4, 5], [5, 6]]) x = self.SparseTensor(i, v, torch.Size([5, 4, 2])) </del> <ins> v = self.ValueTensor([[1, 2], [2, 3], [3, 4], [4, 5]]) # TODO: This is also testing that, if coalesce is a no-op, # the indices don't get permuted. I don't know if we actually # want to give this invariant. x = self.SparseTensor(i, v, torch.Size([5, 4, 2])).coalesce() </ins> dense = self.ValueTensor([ [[1, 3], [2, 2], [3, 3], [4, 2]], [[5, 7], [6, 7], [7, 9], [8, 9]],", "pos": ["and I talked about the intended semantics of sparsemask when given a sparse tensor as an argument, and we came to the following conclusions: with an uncoalesced sparse tensor does not make sense, full stop. Unlike values, where your update might be linear and work correctly on an uncoalesced sparse tensor, there isn't any good reason why you want to pluck out values from the dense tensor multiple times, it is basically always a bug. the (eventual) documentation for sparsemask, we will recommend calling coalesce before doing any operations involving sparsemask (or nonlinear updates to indices/values). should add a test to sparsemask which checks if the flag of the input tensor is true, and raise an error if it is not. This is a cheap safety check which will prevent people from forgetting to call coalesce (as in (2)). There is a performance consequence to calling coalesce every iteration of training. If the gradient is producing coalesced vectors, we can alleviate this problem by adding a preprocessing step to coalesce() which tests if the input is already coalesced, and bugs out in that case. However, I think this is out of scope for this particular ticket for now."], "neg": []}
{"query": "# State initialization if len(state) == 0: state['step'] = 0 <del> state['square_avg'] = grad.new().resize_as_(grad).fill_(1) </del> <ins> state['square_avg'] = grad.new().resize_as_(grad).zero_() </ins> square_avg = state['square_avg'] alpha = group['alpha']", "pos": ["We should unify this with our normal list creating code for the reasonable cases cc\nSimilar issue to this, I believe:", "The hub tests are broken on macOS. I'm seeing SSL certificate related errors when trying to hit GitHub. Run the macOS tests on CI with TestHub enabled. I tried installing the latest version of from conda-forge (which pulled in ) but this didn't fix the issue. cc\nLooks like the same error I'm getting trying to use Ubuntu 18.04 with fresh install of conda and Pytorch 1.2\nI think this will all be solved with .", "Porting TH operators is essential for code simplicity and performance reasons. Porting guides and Q&A are available in umbrella issue: Feel free to add as a reviewer to get a prioritized review.", "This is of particular use in list comprehensions: This should work:\nI'll need to adapt list comprehensions to use Iterables. I'll take a look!", "I haven't looked into this in depth but this looks fishy: More here: cc\nis this still an issue in 1.3.1?\nI can repro this on master:\nThe problem is: removes the parameters from the module's parameter list and makes them normal class attributes: applies only to parameters, not to the weight attributes.  (which is called from acts on which includes the value of the weight attributes and the bias parameters. So, it contains a mix of CUDA and CPU tensors.\nThis is my workaround for right now: This applies weight norm to an rnn module and overrides the function to deal with the error.\nThis appears to be fixed. do you want to verify?\nI tested this on nightly and it worked for me, so closing.", "As mentioned in title, I want to make sure if pytorch can deal with the situation that each input in a batch has different shapes and thus different computation graphs.\nPlease don't open duplicated issue for discussion. Context will be missed. Try to discuss about it here:", "Be Careful What You Backpropagate: A Case For Linear Output Activations & Gradient Boosting I can work on this if this can be to pytorch. Please let me know. Thanks!\nThis should be fairly straightforward to add in user code using . Eventually it could be to core, but its utility should probably be validated in external repos first."], "neg": []}
{"query": "self.assertLessEqual(params.data.dist(solution), initial_dist) <ins> def _test_rosenbrock_sparse(self, constructor): params_t = torch.Tensor([1.5, 1.5]) params = Variable(torch.Tensor([1.5, 1.5]), requires_grad=True) params_c = Variable(torch.Tensor([1.5, 1.5]), requires_grad=True) optimizer = constructor([params]) optimizer_c = constructor([params_c]) solution = torch.Tensor([1, 1]) initial_dist = params.data.dist(solution) def eval(params, sparse_grad, w): optimizer.zero_grad() loss = rosenbrock(params) loss.backward() # params.grad.data now has the computed gradient. Turn # it into a sparse tensor params.grad.data.copy_(drosenbrock(params.data)) # This is really goofy if w: i = torch.LongTensor([[0]]) v = torch.DoubleTensor([params.grad.data[0]]) else: i = torch.LongTensor([[1]]) v = torch.DoubleTensor([params.grad.data[1]]) x = sparse.DoubleTensor(i, v, torch.Size([2])) if sparse_grad: params.grad.data = x else: params.grad.data = x.to_dense() return loss for i in range(2000): w = torch.rand(1)[0] > 0.5 optimizer.step(functools.partial(eval, params, True, w)) optimizer_c.step(functools.partial(eval, params_c, False, w)) self.assertEqual(params.data, params_c.data) self.assertLessEqual(params.data.dist(solution), initial_dist) </ins> def _test_basic_cases_template(self, weight, bias, input, constructor): weight = Variable(weight, requires_grad=True) bias = Variable(bias, requires_grad=True)", "pos": ["and I talked about the intended semantics of sparsemask when given a sparse tensor as an argument, and we came to the following conclusions: with an uncoalesced sparse tensor does not make sense, full stop. Unlike values, where your update might be linear and work correctly on an uncoalesced sparse tensor, there isn't any good reason why you want to pluck out values from the dense tensor multiple times, it is basically always a bug. the (eventual) documentation for sparsemask, we will recommend calling coalesce before doing any operations involving sparsemask (or nonlinear updates to indices/values). should add a test to sparsemask which checks if the flag of the input tensor is true, and raise an error if it is not. This is a cheap safety check which will prevent people from forgetting to call coalesce (as in (2)). There is a performance consequence to calling coalesce every iteration of training. If the gradient is producing coalesced vectors, we can alleviate this problem by adding a preprocessing step to coalesce() which tests if the input is already coalesced, and bugs out in that case. However, I think this is out of scope for this particular ticket for now."], "neg": []}
{"query": "example_value = tensor[invalid_value_mask.eq(0)][0] tensor[invalid_value_mask] = example_value if invalid_value_mask.any(): <del> min_sz = 3 </del> <ins> min_sz = max(min_sz, 3) </ins> int_mode = True # TODO: use fmod?", "pos": ["It seems that the following link that can be reached from here (check the screenshot) is broken: Page: Broken link: Screenshot: ! cc\nthis should be transferred to , but I don't have permissions to do that. Fix is available at pytorch/hub."], "neg": []}
{"query": "toDeviceTensor<THCIndex_t, 1>(state, target), toDeviceTensor<real, 1>(state, output), weights ? THCTensor_(data)(state, weights) : NULL, <ins> n_classes, </ins> ignore_index); <ins> THCudaCheck(cudaGetLastError()); </ins> if (weights) { THCTensor_(free)(state, weights); }", "pos": ["Claim: When I use an invalid label with F.crossentropy with reduce=False, I don't receive an error message. I do receive an error message when reduce=True. Code: Result: No error message received. Expected error message: Context: I was using , accidentally used an invalid label, and got extremely large losses and extremely small losses, and inconsistently so. Environment: OS: Ubuntu 16.04 on Docker and host PyTorch version: 0.3.0.post4 How you installed PyTorch: conda Python version: 3.6.3"], "neg": []}
{"query": "``` docker build -t pytorch . ``` <del> Alternatively, if you want a runtime image, build with ``` docker build -t pytorch . -f tools/docker/Dockerfile_runtime ``` and run with nvidia-docker: </del> <ins> Alternatively, if you want to use a runtime image, you can use the pre-built one from Docker Hub and run with nvidia-docker: </ins> ``` <del> nvidia-docker run --rm -ti --ipc=host pytorch </del> <ins> nvidia-docker run --rm -ti --ipc=host pytorch/pytorch:latest </ins> ``` Please note that PyTorch uses shared memory to share data between processes, so if torch multiprocessing is used (e.g. for multithreaded data loaders) the default shared memory segment size that container runs with is not enough, and you", "pos": ["Do you plan to publish an official entry on docker hub?\nany update? I think it's an important enhancement.\nI believe this would just be pushing the image from ? If so, I can add it to Docker Hub if you like?\nConfigure automated builds on Docker Hub\nThe dockerfile you referenced builds from current master, I don't know if that's desirable. There is a runtime dockerfile in tools/docker that installs current binary release, I'd think that's what should be hosted on the dockerhub.\nyup, that makes more sense. Thanks\nOops, looks like runtime dockerfile would be broken now (see discussion in #beginner, conda packages conflict with conda numpy).\nJust tested, the runtime dockerfile seems to work just fine. Pulls in numpy 1.13.1 which seems to not have any issues.\npre-built images are now available, you can run them with:"], "neg": []}
{"query": "return SymbolicCallArg(expr, tree.numel) def generate_workspace_allocation(self, nbytes, device, zero_fill): <ins> if isinstance(nbytes, sympy.Expr): nbytes = V.graph.sizevars.size_hint(nbytes) </ins> line = self.make_allocation( \"workspace\", device, torch.uint8, shape=(nbytes,), stride=(1,) )", "pos": ["Repro: Error: cc\nUn-assign myself to let work on it.\n, I can't reproduce the issue. The is no longer in the codebase. I gave a try of , which worked well. Could you double check if you can reproduce the issue?\nYes, I can still reproduce with your cmdline. Can you retry after deleting your local cache?"], "neg": []}
{"query": "T bern_uniform = bernoulli[idx]; int _mask = (int) THCNumerics<T>::lt(bern_uniform, q[rand_ind]); output[idx] = J[rand_ind]*(1 -_mask) + (rand_ind+1L) * _mask; <del> } </del> <ins> } </ins> } template <typename T>", "pos": ["Looking at the values returned from tn.multinomial(1) and counting them, it seems like it returns the values of the multinomial distribution taken from t, i.e. that it has ignored the signs of the values. However, it seem like always returns a tensor filled with zeros.\nlooking into this"], "neg": []}
{"query": "(e.g. BatchNorm stats) are broadcast form the module in process of rank 0, to all other replicas in the system in every iteration. <ins> .. warning:: Forward and backwrad hooks defined on :attr:`module` and its submodules won't be invoked anymore, unless the hooks are initialized in the :meth:`forward` method. </ins> Args: module: module to be parallelized device_ids: CUDA devices (default: all devices)", "pos": ["PyTorch version: 0.3.0 Python version: 3.5 GPU models and configuration: 8x NVIDIA Titan I found that requiresgrad setting in module is ignored when module is wrapped with nn.DataParallel. For example, This code should be executed because very small part of network activation (only fc layer) is stored for backward computation, but it results to run-time error: Internally, I found a crack in function which is in . In replicate function, it copies all parameter in module (# of replica times) with . In broadcasting code, it just defines new with default constructor parameter, which is always set to True. I think there can be some choices to fix this issue. is intended behavior for DataParallel, so we should use volatile (or ) to implement transfer learning in DataParallel model should be fixed to handle requires_grad when copying parameters. should be synchronizing in I think any of solution doesn't need lots of effort to fix it."], "neg": []}
{"query": "include_dirs = [] extra_link_args = [] <del> extra_compile_args = ['-std=c++11'] </del> <ins> extra_compile_args = ['-std=c++11', '-Wno-write-strings'] </ins> cwd = os.path.dirname(os.path.abspath(__file__)) lib_path = os.path.join(cwd, \"torch\", \"lib\")", "pos": ["<!-- A clear and concise description of what the bug is. --The target name is used by a test executable in mkldnn. Steps to reproduce the behavior: BUILDTEST=1 build <!-- If you have a code sample, error messages, stack traces, please provide it here as well --<!-- A clear and concise description of what you expected to happen. --CMake configure passes. Please copy and paste the output from our (or fill out the checklist below manually). You can get the script and run it with: PyTorch Version (e.g., 1.0): v1.5.0-rc2 OS (e.g., Linux): Windows How you installed PyTorch (, , source): source Build command you used (if compiling from source): python build Python version: 3.x CUDA/cuDNN version: 9.2 / 10.1 / 10.2 / NA GPU models and configuration: Tesla M60 Any other relevant information: <!-- Add any other context about the problem here. --We should not build tests for mkldnn.\nI have encountered the same error, how do I resolve it?\nIt is actually mkldnn that is out of sync. So running the following script should help.\nThanks a lot for this I posted this on the PyTorch Discussions but to no avail, and now I've finally understood the reason of this error. Thanks a lot. Surprisingly, for me the build works out if I build from the master repo, but changing to a different version always fails."], "neg": []}
{"query": "THLongStorage_free(topKSize); #define RUN_K(INDEX_T, DIM, DIR)                                         <del> gatherTopK<real, INDEX_T, DIM, DIR>                                          </del> <ins> gatherTopK<real, INDEX_T, DIM, DIR>                                    </ins> <<<grid, block, 0, THCState_getCurrentStream(state)>>>(              inputInfo,                                                         sliceSize,                                                        ", "pos": ["Repro: Expect two small tensors of small integers to be printed. Instead, only the first case works. In the second case, you get a series of failed assertions starting with this: I wonder if it is related to\nis this still an outstanding issue? I cannot reproduce this with an install from the master branch.\nThe size which causes failure doesn't seem deterministic. Using a pretty recent build (a few days old), I now have to increase the \"9\" to a larger number to make it fail. Sometimes it will fail at 10, sometimes at 11. Here's an updated case that seems to fail reliably at some size, for me:"], "neg": []}
{"query": "for(int i = 0; i < input->nDimension; i++){ no_elements *= input->size[i]; } <del> no_elements *= scale_factor * scale_factor; </del> <ins> no_elements *= scale_factor; </ins> int d1; int d2;", "pos": ["Original post at: Original post content: code: Tested on Linux with pytorch 0.2.0+ & GTX 750 and Mac OS X Sierra with pytorch 0.2.0+ & GT 750M It seems to be work well with mode=\u2018linear\u2019 and memory of gpu is larger than 1GB. I think it\u2019s something wrong. Comment: successfully reproduced using the above snippet on master."], "neg": []}
{"query": "y = x[2:] self.assertEqual(int(y), 3) <del> @unittest.skipIf(torch.set_flush_denormal(False), \"flush_denormal not supported\") </del> <ins> # skip this test for now as it affects all tests @unittest.skipIf(True, \"flush_denormal not supported\") </ins> def test_set_flush_denormal(self): tiny_float = 1e-42 tiny_double = 1e-320", "pos": ["hi, I just saw this error in the last few days. both for GPU and CPU only environment. Is the test precision assertion correct ?\nOn PPC64 it looks like denormal flags aren't being set correctly. This PR from 2 days ago denormals support, but likely assumed only AMD64 / x86\ncurrently the setflushdenormal call will always return False if on ppc64le. The unittest has a skipIf as below. I believe the intention is to skip this test if not running on x86. But since setflushdenormal returns a False on ppc64le , then it actually runs the test and not run on x86 \"flushdenormal not supported\") I believe is you just add a not as below in the torch.setflushdenormal(False), \"flush_denormal not supported\") the test should then run on x86 but skipped for ppc64le or other platforms.\nDoing calls that have side effects in skipIf is a bad idea. We should write a context manager that always resets the previous state + skips the test if the call returned False\nwhat can we do in the meantime to skip over the denormal test, which is invalid, on ppc64 ?\nSkip this test everywhere. These checks will affect all tests and we don\u2019t want that\nclosed via"], "neg": []}
{"query": "THD_LIB = os.path.join(lib_path, 'libTHD.1.dylib') NCCL_LIB = os.path.join(lib_path, 'libnccl.1.dylib') <del> if WITH_NCCL and subprocess.call('ldconfig -p | grep libnccl >/dev/null', shell=True) == 0: </del> <ins> if WITH_NCCL and (subprocess.call('ldconfig -p | grep libnccl >/dev/null', shell=True) == 0 or subprocess.call('/sbin/ldconfig -p | grep libnccl >/dev/null', shell=True) == 0): </ins> SYSTEM_NCCL = True main_compile_args = ['-D_THP_CORE']", "pos": ["Hello, on Debian (and possibly Ubuntu and other systems), is in and not in non-root users' search paths. Thus this line: at does not do what it should when compiling not as root. Either writing or adding a second check with the directory would help building pytorch on these systems. Would such a change be acceptable? Thank you & best regards Thomas\nOr fix this by exporting ?\nsuch a chance would be absolutely acceptable.\nfixed in PR"], "neg": []}
{"query": "THCAssertSameGPU(THCTensor_(checkGPU)(state, 2, self_, src_)); THArgCheck(src_->nDimension == 2, 1, \"expected a matrix\"); <del> THCTensor *src = src_; if (self_ == src_) src = THCTensor_(newContiguous)(state, src_); </del> <ins> if (self_ != src_) THCTensor_(resizeAs)(state, self_, src_); </ins> <del> int64_t stride0 = src->stride[0]; int64_t stride1 = src->stride[1]; real *start = THCTensor_(data)(state, src); </del> <ins> int64_t stride0 = self_->stride[0]; int64_t stride1 = self_->stride[1]; real *start = THCTensor_(data)(state, self_); </ins> TensorTriOp<real, 1> op(start, stride0, stride1, k); if (self_ == src_) { <del> if (!THC_pointwiseApply1(state, src, op)) { </del> <ins> if (!THC_pointwiseApply1(state, src_, op)) { </ins> THArgCheck(false, 2, CUTORCH_DIM_WARNING); } } else { <del> THCTensor_(resizeAs)(state, self_, src); if (!THC_pointwiseApply2(state, self_, src, op)) { </del> <ins> if (!THC_pointwiseApply2(state, self_, src_, op)) { </ins> THArgCheck(false, 2, CUTORCH_DIM_WARNING); } } <del> if (self_ == src_) THCTensor_(freeCopyTo)(state, src, src_); </del> THCudaCheck(cudaGetLastError()); }", "pos": ["On recent PyTorch master (built from source; 0.4.0a0+) on a P100, CUDA 9:\nI think the results on CPU meets the defination of triu (upper triangular part of the matrix). It is really hard to handle with a tensor from expand operation with multi-thread not only on GPU but also on CPU. The unexpected result will also happen on CPU if we apply openmp to it. I think it is necessary to eleminate the difference between the implementation and defination.\nIt doesn\u2019t matter if it\u2019s cpu or not. triu doesn\u2019t have an underscore which means it doesn\u2019t operate in-place and should always give the same result (in new memory)"], "neg": []}
{"query": "self._data_ptrs = [] return <del> # This is quite ugly, but it allows us to reuse the cuDNN code without larger # modifications. It's really a low-level API that doesn't belong in here, but # let's make this exception. from torch.backends.cudnn import rnn from torch.backends import cudnn from torch.nn._functions.rnn import CudnnRNN handle = cudnn.get_handle() with warnings.catch_warnings(record=True): fn = CudnnRNN( self.mode, self.input_size, self.hidden_size, num_layers=self.num_layers, batch_first=self.batch_first, dropout=self.dropout, train=self.training, bidirectional=self.bidirectional, dropout_state=self.dropout_state, ) # Initialize descriptors fn.datatype = cudnn._typemap[any_param.type()] fn.x_descs = cudnn.descriptor(any_param.new(1, self.input_size), 1) fn.rnn_desc = rnn.init_rnn_descriptor(fn, handle) # Allocate buffer to hold the weights num_weights = rnn.get_num_weights(handle, fn.rnn_desc, fn.x_descs[0], fn.datatype) fn.weight_buf = any_param.new(num_weights).zero_() fn.w_desc = rnn.init_weight_descriptor(fn, fn.weight_buf) # Slice off views into weight_buf params = rnn.get_parameters(fn, handle, fn.weight_buf) all_weights = [[p.data for p in l] for l in self.all_weights] # Copy weights and update their storage rnn._copyParams(all_weights, params) for orig_layer_param, new_layer_param in zip(all_weights, params): for orig_param, new_param in zip(orig_layer_param, new_layer_param): orig_param.set_(new_param.view_as(orig_param)) self._data_ptrs = list(p.data.data_ptr() for p in self.parameters()) </del> <ins> with torch.cuda.device_of(any_param): # This is quite ugly, but it allows us to reuse the cuDNN code without larger # modifications. It's really a low-level API that doesn't belong in here, but # let's make this exception. from torch.backends.cudnn import rnn from torch.backends import cudnn from torch.nn._functions.rnn import CudnnRNN handle = cudnn.get_handle() with warnings.catch_warnings(record=True): fn = CudnnRNN( self.mode, self.input_size, self.hidden_size, num_layers=self.num_layers, batch_first=self.batch_first, dropout=self.dropout, train=self.training, bidirectional=self.bidirectional, dropout_state=self.dropout_state, ) # Initialize descriptors fn.datatype = cudnn._typemap[any_param.type()] fn.x_descs = cudnn.descriptor(any_param.new(1, self.input_size), 1) fn.rnn_desc = rnn.init_rnn_descriptor(fn, handle) # Allocate buffer to hold the weights self._param_buf_size = rnn.get_num_weights(handle, fn.rnn_desc, fn.x_descs[0], fn.datatype) fn.weight_buf = any_param.new(self._param_buf_size).zero_() fn.w_desc = rnn.init_weight_descriptor(fn, fn.weight_buf) # Slice off views into weight_buf params = rnn.get_parameters(fn, handle, fn.weight_buf) all_weights = [[p.data for p in l] for l in self.all_weights] # Copy weights and update their storage rnn._copyParams(all_weights, params) for orig_layer_param, new_layer_param in zip(all_weights, params): for orig_param, new_param in zip(orig_layer_param, new_layer_param): orig_param.set_(new_param.view_as(orig_param)) self._data_ptrs = list(p.data.data_ptr() for p in self.parameters()) </ins> def _apply(self, fn): ret = super(RNNBase, self)._apply(fn)", "pos": ["Since the new pytorch version, I have a warning when I that says: When I do as suggested, and then try to encode again, I get the following error though: Any idea?\ndo you have code the reproduces the issue?\nWhat was the fix? I still see the error on Pytorch Version: 0.2.03 AttributeError: 'LSTM' object has no attribute 'parambufsize'"], "neg": []}
{"query": "other types will be a shallow copy and can be corrupted if written to in the model's forward pass. <ins> .. warning:: Forward and backwrad hooks defined on :attr:`module` and its submodules won't be invoked anymore, unless the hooks are initialized in the :meth:`forward` method. </ins> Args: module: module to be parallelized device_ids: CUDA devices (default: all devices)", "pos": ["PyTorch version: 0.3.0 Python version: 3.5 GPU models and configuration: 8x NVIDIA Titan I found that requiresgrad setting in module is ignored when module is wrapped with nn.DataParallel. For example, This code should be executed because very small part of network activation (only fc layer) is stored for backward computation, but it results to run-time error: Internally, I found a crack in function which is in . In replicate function, it copies all parameter in module (# of replica times) with . In broadcasting code, it just defines new with default constructor parameter, which is always set to True. I think there can be some choices to fix this issue. is intended behavior for DataParallel, so we should use volatile (or ) to implement transfer learning in DataParallel model should be fixed to handle requires_grad when copying parameters. should be synchronizing in I think any of solution doesn't need lots of effort to fix it."], "neg": []}
{"query": "if (weight_dim != k) { std::stringstream ss; <del> ss << \"Expected \" << weight_dim << \"-dimensional input for \" << weight_dim << \"-dimensional weight \" << weight.sizes() << \", but got input of size \" << input.sizes() << \" instead\"; </del> <ins> ss << \"Expected \" << k << \"-dimensional weight for \" << k << \"-dimensional input \" << input.sizes() << \", but got weight of size \" << weight.sizes() << \" instead\"; </ins> throw std::runtime_error(ss.str()); } if (weight.size(0) < groups) {", "pos": ["OS: Ubuntu 16.04 PyTorch version: 0.3.0.post4 How you installed PyTorch (conda, pip, source): pip Python version: 3.5.2 CUDA/cuDNN version: CUDA 8 The below script produces an incorrect error message () The below script produces an incorrect error message ()"], "neg": []}
{"query": "(4.53, 3.83, -6.64, 2.06))).t() b = torch.Tensor(((8.58, 8.26, 8.48, -5.28), (9.35, -4.43, -0.70, -0.26))).t() <del> _test(a, b, expectedNorm) </del> <ins> _test_underdetermined(a, b, expectedNorm) </ins> # test overderemined expectedNorm = 17.390200628863", "pos": ["b = (600,1) A = (600,2) X, = (b,A) X.size() #torch.Size([2, 1]) Y, = (b.cuda(),A.cuda()) Y.size() #torch.Size([600, 1])\nthat is interesting. says that the first rows returned ( here) include the solution of the least-squares problem. I've verified that the first 2 rows of X and Y are the same so you can probably use that for now while we figure out what the rest of the output is."], "neg": []}
{"query": "print('Distributed not available, skipping tests') sys.exit(0) <ins> SKIP_IF_NO_CUDA_EXIT_CODE = 75 def skip_if_no_cuda_distributed(func): func.skip_if_no_cuda_distributed = True @wraps(func) def wrapper(*args, **kwargs): if not torch.cuda.is_available(): sys.exit(SKIP_IF_NO_CUDA_EXIT_CODE) return func(*args, **kwargs) return wrapper </ins> @contextmanager def _lock():", "pos": ["is causing a crash when we try to build pytorch. Crash looks like this: (all times in US/West) Our 7AM, 8AM builds and earlier were all fine but 9AM, 10AM, 11AM builds all crashed with that same traceback. The new import was in which I think went in between 8AM and 9AM today See command below: See here if needed for : <!-- A clear and concise description of what you expected to happen. --Please copy and paste the output from our (or fill out the checklist below manually). You can get the script and run it with: PyTorch Version (e.g., 1.0): building from master OS (e.g., Linux): linux How you installed PyTorch (, , source): source Build command you used (if compiling from source): see above under 'reproduce' Python version: 3.6\nmaybe easier to read for build command used:\nplease try in the CircleCi environment?\nin our build script seems to have fixed this. Thanks !"], "neg": []}
{"query": "log_scale = (self.temperature.new(self.temperature.shape).fill_(K).lgamma() - self.temperature.log().mul(-(K - 1))) score = logits - value.mul(self.temperature) <del> score = (score - log_sum_exp(score)).sum(-1) </del> <ins> score = (score - _log_sum_exp(score)).sum(-1) </ins> return score + log_scale", "pos": ["Leaving this issue here for higher visibility. Originally reported by here: Also mentioned by Piggybacking off of this issue, has strange behavior (reported here: )\nI have fixed this locally. Could I send a PR?\nYes, go ahead!"], "neg": []}
{"query": "toDeviceTensor<real, 1>(state, gradOutput), toDeviceTensor<real, 2>(state, gradInput), weights ? THCTensor_(data)(state, weights) : NULL, <ins> n_classes, </ins> ignore_index); <ins> THCudaCheck(cudaGetLastError()); </ins> if (weights) { THCTensor_(free)(state, weights); }", "pos": ["Claim: When I use an invalid label with F.crossentropy with reduce=False, I don't receive an error message. I do receive an error message when reduce=True. Code: Result: No error message received. Expected error message: Context: I was using , accidentally used an invalid label, and got extremely large losses and extremely small losses, and inconsistently so. Environment: OS: Ubuntu 16.04 on Docker and host PyTorch version: 0.3.0.post4 How you installed PyTorch: conda Python version: 3.6.3"], "neg": []}
{"query": "if hasattr(torch, name) and name not in EXCLUDE_FUNCTIONAL: def fn(*inputs): output = getattr(torch, name)(*inputs) <del> if test_output_indices is None: return output else: return tuple(output[i] for i in test_output_indices) </del> <ins> return output_process_fn(output) </ins> f_args_variable = (self_variable,) + args_variable f_args_tensor = (self_tensor,) + args_tensor", "pos": ["I am having problem in running the backward propagation with svd. In particular, it seems working when the first dimension of the matrix x (below) is greater than the second. But I get the following error if the first is smaller than the second. PyTorch version: 2018.04.20 Is debug build: No CUDA used to build PyTorch: 8.0.61 OS: Ubuntu 16.04.4 LTS GCC version: (Ubuntu 5.4.0-6ubuntu1~16.04.9) 5.4.0 CMake version: Could not collect Python version: 3.5 Is CUDA available: No CUDA runtime version: No CUDA GPU models and configuration: No CUDA Nvidia driver version: No CUDA cuDNN version: No CUDA Versions of relevant libraries: [pip3] numpy (1.13.3) [pip3] torch (2018.4.20) [pip3] torchvision (0.2.0) [conda] pytorch 0.3.1 py35cuda8.0.61cudnn7.0.52 pytorch [conda] pytorch-nightly 2018.04.20 py35cuda8.0.61cudnn7.1.21 pytorch [conda] torchvision 0.2.0 py35heaa392f_1 pytorch\nRepro'ed on 0.4 branch. works btw.\nWhat does \"Repro'ed\" means? Should I install from that branch? I need to compute the kernel of a matrix, therefore I need the last columns of the matrix V (which is why I have to pass ).\nIt's a shorthand for \"reproduced\", which means that the bug still hasn't been fixed. Sorry, we'll try to get to it soon."], "neg": []}
{"query": "IF(C_AVX2_FOUND) IF(MSVC) <del> SET_SOURCE_FILES_PROPERTIES(${PROJECT_SOURCE_DIR}/src/TH/vector/AVX2.c PROPERTIES COMPILE_FLAGS \"/Ox /arch:AVX2 ${C_AVX2_FLAGS}\") </del> <ins> SET_SOURCE_FILES_PROPERTIES(${PROJECT_SOURCE_DIR}/src/TH/vector/AVX2.c PROPERTIES COMPILE_FLAGS \"${MSVC_OPT_FLAG}/arch:AVX2 ${C_AVX2_FLAGS}\") </ins> ELSE(MSVC) SET_SOURCE_FILES_PROPERTIES(${PROJECT_SOURCE_DIR}/src/TH/vector/AVX2.c PROPERTIES COMPILE_FLAGS \"-O3 ${C_AVX2_FLAGS}\") ENDIF(MSVC)", "pos": ["The current master triggers the following error messages from Visual Studio 2015 (Debug build): 2------ Build started: Project: ATen, Configuration: Debug x64 ------ 2cl : Command line error D8016: '/Ox' and '/RTC1' command-line options are incompatible ... 14LINK : fatal error LNK1104: cannot open file '....srcATenDebug' It appears that they are triggered by *.c files in the ATen project, like \"srcTHvectorAVX.c\"\nWhy do you need Debug build?\nIsn't ATen meant to support both Debug and Release builds? FYI I'm involved with the project at which does support both build configurations, and we would like to be able to link our project with ATen.\nWell, you can remove the flags in as a workground.\nWe should patch this file to ignore this flag when building the library in debug mode. It shouldn't be too hard\nI catched similar errors , when I Tryed to build Debug Version I try to build Release and catched next errors e:pytorchatensrcatencpuvec256vec256int.h(184): note: see reference to function template instantiation 'at::vec256::anonymous-namespace'::Vec256<int16t::blend<1(at::vec256::anonymous-namespace'::Vec256<int16t)' being compiled 28e:pytorchatensrcatencpuvec256vec256int.h(147): error C3861: 'mm256extract_epi16': identifier not found What I should do?\nThere're two issues in your description. The first question is easy. You'll also have to specify the flag to 1. The second question is about your compiler. I guess you are using VS2015 or earlier versions of VS2017. Make sure the version in the output is 19.11 after your enter in CMD.\nYes I am using VS2015. My Compiler is 19.0.2. Must I use VS 2017 in any case?\nSorry, but the answer is yes. The AVX2 functions was in VS2017 and they are used in PyTorch. So VS2017 19.11 is a must.\nNo problem:) Thank you."], "neg": []}
{"query": "parameters[2] = make_param() param_list[2] = parameters[2] check() <ins> self.assertEqual(param_list[1:], nn.ParameterList(parameters[1:])) self.assertEqual(param_list[3:], nn.ParameterList(parameters[3:])) self.assertEqual(param_list[:-1], nn.ParameterList(parameters[:-1])) self.assertEqual(param_list[:-3], nn.ParameterList(parameters[:-3])) self.assertEqual(param_list[::-1], nn.ParameterList(parameters[::-1])) </ins> with self.assertRaises(TypeError): param_list += make_param()", "pos": ["it'd be great to pick out elements of of a sequential with array slicing:\nI agree this idea!\nHowever,\nthis has come up again (last in ). Shall we reconsider?\nI attempted getting a to work. Is this fine? If this is fine, I can issue a PR too. Let me know what you think. I believe this modification needs to be done in the class.\ncc: This issue has been open for quite sometime."], "neg": []}
{"query": ".. autofunction:: linear <ins> :hidden:`bilinear` ~~~~~~~~~~~~~~~~ .. autofunction:: bilinear </ins> Dropout functions -----------------", "pos": ["is not available in the docs page, but the doc string has been in . This was mentioned earlier in", "Oh, bilinear, I missed that last time I looked at it. I'll add a docstring.", "returns nothing. Might be useful to have some linter in the build process to find Python functions missing", "Pytorch 0.4.0 gives a segmentation fault with dilated 2D convolutions over a specific size. import torch import as nn model = nn.Conv2d(inchannels=64, outchannels=64, kernelsize=5, dilation=2) model(([1, 64, 1166, 1166])) # This runs model(([1, 64, 1167, 1167])) # This crashes PyTorch version: 0.4.0 Is debug build: No CUDA used to build PyTorch: 9.1.85 OS: CentOS Linux 7 (Core) GCC version: (GCC) 4.8.5 (Red Hat 4.8.5-28) CMake version: version 2.8.12.2 Python version: 3.6 Is CUDA available: Yes CUDA runtime version: 8.0.61 GPU models and configuration: GPU 0: GeForce GTX 1080 Ti GPU 1: TITAN X (Pascal) GPU 2: Quadro K2200 Nvidia driver version: 390.59 cuDNN version: Probably one of the following: /usr/local/cuda- /usr/local/cuda- /usr/local/cuda-7.5/lib64/libcudnnstatic.a /usr/local/cuda- /usr/local/cuda-8.0/lib64/libcudnn_static.a\ndo you think this is the same as the segfault you showed me a week ago on convolution?\nah yes. int overflow. I only fixed conv3d because I didn\u2019t want to hurt perf. I\u2019ll fix all of them.", "This should be an easy one.. The function is implemented, just no sign of it in the docs."], "neg": []}
{"query": "for (int i = 0; i < required_args; ++i) { if (!inputs[i].defined()) { std::stringstream ss; <del> ss << name << \": expected Variable at argument \" << i << \" (got None)\"; </del> <ins> ss << name << \": expected Tensor at argument \" << i << \" (got None)\"; </ins> throw std::runtime_error(ss.str()); } }", "pos": ["Considering the merge of and , and the \"removal\" of from the API, there are still some messages that reference . This might prove to be confusing for new users starting off from , as they probably might not have heard of . I did a code search to find out multiple instances of this in . PyTorch or Caffe2: PyTorch How you installed PyTorch (conda, pip, source): source (@ ) Build command you used (if compiling from source): OS: Ubuntu 16.04.2 PyTorch version: Python version: 3.5 CUDA/cuDNN version: 8.0.44 GPU models and configuration: GeForce 940M GCC version (if compiling from source): 5.4.0 CMake version: 3.5.1"], "neg": []}
{"query": "set(CUDNN_ROOT_DIR \"\" CACHE PATH \"Folder contains NVIDIA cuDNN\") <del> find_path(CUDNN_INCLUDE_DIR cudnn.h </del> <ins> if($ENV{CUDNN_INCLUDE_DIR}) SET(CUDNN_INCLUDE_DIR $ENV{CUDNN_INCLUDE_DIR}) else($ENV{CUDNN_INCLUDE_DIR}) find_path(CUDNN_INCLUDE_DIR cudnn.h </ins> HINTS ${CUDNN_ROOT_DIR} ${CUDA_TOOLKIT_ROOT_DIR} PATH_SUFFIXES cuda/include include) <ins> endif($ENV{CUDNN_INCLUDE_DIR}) </ins> <del> find_library(CUDNN_LIBRARY cudnn </del> <ins> if($ENV{CUDNN_LIBRARY}) SET(CUDNN_LIBRARY $ENV{CUDNN_LIBRARY}) else($ENV{CUDNN_LIBRARY}) find_library(CUDNN_LIBRARY cudnn </ins> HINTS ${CUDNN_LIB_DIR} ${CUDNN_ROOT_DIR} ${CUDA_TOOLKIT_ROOT_DIR} PATH_SUFFIXES lib lib64 cuda/lib cuda/lib64 lib/x64) <ins> endif($ENV{CUDNN_LIBRARY}) </ins> find_package_handle_standard_args( CUDNN DEFAULT_MSG CUDNN_INCLUDE_DIR CUDNN_LIBRARY)", "pos": ["Traceback (most recent call last): File \"\", line 3, in <moduleimport model File \"/home/abc/github/chinese-\", line 5, in <modulefrom import crnnOcr File \"/home/abc/github/chinese-\", line 72, in <modulemodel,converter = crnnSource() File \"/home/abc/github/chinese-\", line 64, in crnnSource model = crnn.CRNN(32, 1, len(alphabet)+1, 256, 1).cuda() File \"/usr/local/lib/python2.7/dist-\", line 216, in cuda return (lambda t: t.cuda(device)) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 123, in apply self.flattenparameters() File \"/usr/local/lib/python2.7/dist-\", line 65, in flattenparameters if not or not (anyparam): File \"/usr/local/lib/python2.7/dist-\", line 82, in isacceptable if libcudnn() is None: File \"/usr/local/lib/python2.7/dist-\", line 58, in libcudnn 'but linked against {}'.format(compileversion, _cudnnversion)) RuntimeError: cuDNN version mismatch: PyTorch was compiled against 6021 but linked against 5110 How to solve the problem?\ncould you pull the latest master branch and recompile. I pushed a change yesterday that fixes this:\nI remove the old code and pull new code, then compile, install.It still fails. /usr/local/lib/python2.7/dist- UserWarning: No training configuration found in save file: the model was not compiled. Compile it manually.", "('No training configuration found in save file: ' Traceback (most recent call last): File \"\", line 3, in <moduleimport model File \"/home/qianbb/github/chinese-\", line 5, in <modulefrom import crnnOcr File \"/home/qianbb/github/chinese-\", line 72, in <modulemodel,converter = crnnSource() File \"/home/qianbb/github/chinese-\", line 64, in crnnSource model = crnn.CRNN(32, 1, len(alphabet)+1, 256, 1).cuda() File \"/usr/local/lib/python2.7/dist-\", line 216, in cuda return (lambda t: t.cuda(device)) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 100, in apply self.flattenparameters() File \"/usr/local/lib/python2.7/dist-\", line 66, in flattenparameters if not or not (anyparam): File \"/usr/local/lib/python2.7/dist-\", line 87, in isacceptable if libcudnn() is None: File \"/usr/local/lib/python2.7/dist-\", line 58, in libcudnn 'but linked against {}'.format(compileversion, _cudnnversion)) RuntimeError: cuDNN version mismatch: PyTorch was compiled against 6021 but linked against 5110 The library of cuda is shown as following screenshot. !\nI found the issue, I fixed it in which we will merge shortly.\nThanksDoes it have been merged?\nyes. it is merged.\nIt seems the fix doesn't work.", "The log is shown as following Traceback (most recent call last): File \"\", line 3, in <moduleimport model File \"/home/qianbb/github/chinese-\", line 5, in <modulefrom import crnnOcr File \"/home/qianbb/github/chinese-\", line 72, in <modulemodel,converter = crnnSource() File \"/home/qianbb/github/chinese-\", line 64, in crnnSource model = crnn.CRNN(32, 1, len(alphabet)+1, 256, 1).cuda() File \"/usr/local/lib/python2.7/dist-\", line 222, in cuda return (lambda t: t.cuda(device)) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 100, in apply self.flattenparameters() File \"/usr/local/lib/python2.7/dist-\", line 66, in flattenparameters if not or not (anyparam): File \"/usr/local/lib/python2.7/dist-\", line 87, in isacceptable if libcudnn() is None: File \"/usr/local/lib/python2.7/dist-\", line 58, in libcudnn 'but linked against {}'.format(compileversion, _cudnnversion)) RuntimeError: cuDNN version mismatch: PyTorch was compiled against 6021 but linked against 5110\nthat is weird. Can I see the full log of in pytorch directory. Can you make sure you have latest commits with .\nI am sure the code is up-to-date. First I uninstall the code using the command pip unisntall torch. Then I remove the old code using the command rm -r pytorch and git clone again. The log of install is\nSorry to trouble you. Is there any new information on this issue?\nI spent some time resolving this issue and I don't see too much about it online. Try ensuring that your Pytorch build is correct for your specific versions of Cuda and Cudnn. While I had the correct version of Pytorch, I resolved the issue by installing a different build of the same version.", "The build I installed is . Here are the list of versions and builds: You can download these files locally and install with conda: (assumes command is being run in a directory containing , where this file will be specific to your Cuda and Cudnn version.\nfeedback from this code PackagesNotFoundError\nthank you - I have edited my comment to reflect that I installed from a local zip."], "neg": []}
{"query": "if (cur_target == ignore_index) { continue; } <ins> assert(cur_target  >= 0 && cur_target  < n_classes); </ins> Dtype weight = weights ? weights[cur_target] : ScalarConvert<int, Dtype>::to(1); gradInput[index][cur_target] = -weight * gradOutput[index];", "pos": ["Claim: When I use an invalid label with F.crossentropy with reduce=False, I don't receive an error message. I do receive an error message when reduce=True. Code: Result: No error message received. Expected error message: Context: I was using , accidentally used an invalid label, and got extremely large losses and extremely small losses, and inconsistently so. Environment: OS: Ubuntu 16.04 on Docker and host PyTorch version: 0.3.0.post4 How you installed PyTorch: conda Python version: 3.6.3"], "neg": []}
{"query": "- THTensor* self - long k - CONSTANT __last_dim <del> - CONSTANT true - CONSTANT true </del> <ins> - arg: bool largest default: \"true\" kwarg_only: True - arg: bool sorted default: \"true\" kwarg_only: True </ins> - arguments: - arg: THTensor* values output: True", "pos": ["Hi.. I think this is a bug: So, it doesn't matter if descending is True or False, just specifying it explicitly requires dimension to be specified.. It would be better (I think) if descending could be specified alone..\nSame issue for if is used:\nsort was fixed in topk is still a problem."], "neg": []}
{"query": "if arg['name'] in ['self', 'state', 'dataType', 'handle']: arg['ignore_check'] = True declaration['options'] = self.filter_unique_options(declaration['options']) <del> return declarations </del> <ins> return [d for d in declarations if not d.get('only_register', False)] </ins> def filter_unique_options(self, options): def signature(option):", "pos": ["We should check that the version from cudnn.h matches the version from the loaded shared library. Really weird things happen if they don't match."], "neg": []}
{"query": "self.ceil_mode = ceil_mode def extra_repr(self): <del> return 'norm_type={norm_type}, kernel_size{kernel_size}, stride={stride}, '  </del> <ins> return 'norm_type={norm_type}, kernel_size={kernel_size}, stride={stride}, '  </ins> 'ceil_mode={ceil_mode}'.format(**self.__dict__)", "pos": ["is not available in the docs page, but the doc string has been in . This was mentioned earlier in", "Oh, bilinear, I missed that last time I looked at it. I'll add a docstring.", "returns nothing. Might be useful to have some linter in the build process to find Python functions missing", "Pytorch 0.4.0 gives a segmentation fault with dilated 2D convolutions over a specific size. import torch import as nn model = nn.Conv2d(inchannels=64, outchannels=64, kernelsize=5, dilation=2) model(([1, 64, 1166, 1166])) # This runs model(([1, 64, 1167, 1167])) # This crashes PyTorch version: 0.4.0 Is debug build: No CUDA used to build PyTorch: 9.1.85 OS: CentOS Linux 7 (Core) GCC version: (GCC) 4.8.5 (Red Hat 4.8.5-28) CMake version: version 2.8.12.2 Python version: 3.6 Is CUDA available: Yes CUDA runtime version: 8.0.61 GPU models and configuration: GPU 0: GeForce GTX 1080 Ti GPU 1: TITAN X (Pascal) GPU 2: Quadro K2200 Nvidia driver version: 390.59 cuDNN version: Probably one of the following: /usr/local/cuda- /usr/local/cuda- /usr/local/cuda-7.5/lib64/libcudnnstatic.a /usr/local/cuda- /usr/local/cuda-8.0/lib64/libcudnn_static.a\ndo you think this is the same as the segfault you showed me a week ago on convolution?\nah yes. int overflow. I only fixed conv3d because I didn\u2019t want to hurt perf. I\u2019ll fix all of them.", "This should be an easy one.. The function is implemented, just no sign of it in the docs."], "neg": []}
{"query": "self.assertEqual(output[0][0].sum().data[0], 0) self.assertEqual(output[1][2].sum().data[0], 0) <ins> embedding = nn.Embedding(10, 20, padding_idx=0, sparse=True) input = Variable(torch.LongTensor([[0, 2, 4, 5], [4, 3, 0, 9]])) output = embedding(input) self.assertEqual(output[0][0].sum().data[0], 0) self.assertEqual(output[1][2].sum().data[0], 0) </ins> def test_embedding_max_norm(self): embedding = nn.Embedding(22, 5, max_norm=1.0) input = Variable(torch.LongTensor([2, 8, 8, 6]))", "pos": ["As reported here with repro. When sparse=True, paddingidx should probably error out (as the easiest solution). Alternatively, the gradients for paddingidx indices should be zeroed out explicitly in the sparse=True backward codepath."], "neg": []}
{"query": "torch.randn(4, 4).cuda()) @unittest.skipIf(not TEST_MULTIGPU, \"multi-GPU not supported\") <ins> def test_broadcast_not_requiring_grad(self): variables = [ Variable(torch.randn(1, 2).cuda(), requires_grad=True), Variable(torch.randn(1, 2).cuda(), requires_grad=False), Variable(torch.randn(1, 2).cuda(), requires_grad=False), Variable(torch.randn(1, 2).cuda(), requires_grad=True), Variable(torch.randn(1, 2).cuda(), requires_grad=True), ] broadcasted_variables = Broadcast.apply((0, 1), *variables) for output_idx, broadcasted_var in enumerate(broadcasted_variables): input_var = variables[output_idx % len(variables)] self.assertEqual(input_var.requires_grad, broadcasted_var.requires_grad) @unittest.skipIf(not TEST_MULTIGPU, \"multi-GPU not supported\") </ins> def test_replicate(self): module = nn.Linear(10, 5).float().cuda() input = Variable(torch.randn(2, 10).float().cuda())", "pos": ["PyTorch version: 0.3.0 Python version: 3.5 GPU models and configuration: 8x NVIDIA Titan I found that requiresgrad setting in module is ignored when module is wrapped with nn.DataParallel. For example, This code should be executed because very small part of network activation (only fc layer) is stored for backward computation, but it results to run-time error: Internally, I found a crack in function which is in . In replicate function, it copies all parameter in module (# of replica times) with . In broadcasting code, it just defines new with default constructor parameter, which is always set to True. I think there can be some choices to fix this issue. is intended behavior for DataParallel, so we should use volatile (or ) to implement transfer learning in DataParallel model should be fixed to handle requires_grad when copying parameters. should be synchronizing in I think any of solution doesn't need lots of effort to fix it."], "neg": []}
{"query": "\"torch/csrc/autograd/functions/convolution.cpp\", \"torch/csrc/autograd/functions/init.cpp\", \"torch/csrc/nn/THNN_generic.cpp\", <ins> *split_types(\"torch/csrc/Tensor.cpp\") </ins> ] try:", "pos": ["Trying to run on github actions via AWS instance gives me: I will of course set up now, but it shouldn't crash like that. May be it should use as a fallback? pt-2.0 cc\nIs this a multitenant setup? Also curious which distro are you running, I was under the impression that most linux distros set home by default\nthis is AWS image. one default user. I think it's user, but it wants you to run as root, which is weird. I have manually now set to overcome this.\nWe should update triton to use pathlib or something"], "neg": []}
{"query": "#include \"torch/csrc/autograd/engine.h\" #include \"torch/csrc/autograd/functions/basic_ops.h\" <ins> #include \"torch/csrc/utils/auto_gpu.h\" </ins> #include <atomic> #include <condition_variable>", "pos": ["Consider following code, and run it with multi gpus (e.g. 4): It will output: That is, the default device would always be zero even in DataParallel. My pytorch version is 0.1.122. I think this is not the desired behavior. It would cause some troubles. I tried to insert my own cuda kernel into backward to calculate the gradients, it become very slow, and I fixed it by (grad.getdevice()). Anyway, I think currentdevice in forward and backward should be the same, or could anyone explain to me why they are different?\nthis seems like something we should fix: ?\n<!-- drci-comment-start --:pagefacingup: Preview :pagefacingup: Preview :question: Need help or want to give feedback on the CI? Visit our Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. <!-- drci-comment-end --"], "neg": []}
{"query": "modules[2] = nn.Conv2d(5, 3, 2) module_list[2] = modules[2] check() <ins> self.assertEqual(module_list[1:], nn.ModuleList(modules[1:])) self.assertEqual(module_list[3:], nn.ModuleList(modules[3:])) self.assertEqual(module_list[:-1], nn.ModuleList(modules[:-1])) self.assertEqual(module_list[:-3], nn.ModuleList(modules[:-3])) self.assertEqual(module_list[::-1], nn.ModuleList(modules[::-1])) </ins> with self.assertRaises(TypeError): module_list += nn.ReLU()", "pos": ["it'd be great to pick out elements of of a sequential with array slicing:\nI agree this idea!\nHowever,\nthis has come up again (last in ). Shall we reconsider?\nI attempted getting a to work. Is this fine? If this is fine, I can issue a PR too. Let me know what you think. I believe this modification needs to be done in the class.\ncc: This issue has been open for quite sometime."], "neg": []}
{"query": "-DNO_CUDA=%NO_CUDA% ^ -DNO_NNPACK=%NO_NNPACK% ^ -Dnanopb_BUILD_GENERATOR=0 ^ <del> -DCMAKE_BUILD_TYPE=Release </del> <ins> -DCMAKE_BUILD_TYPE=%BUILD_TYPE% </ins> %MAKE_COMMAND% cd ../..", "pos": ["The current master triggers the following error messages from Visual Studio 2015 (Debug build): 2------ Build started: Project: ATen, Configuration: Debug x64 ------ 2cl : Command line error D8016: '/Ox' and '/RTC1' command-line options are incompatible ... 14LINK : fatal error LNK1104: cannot open file '....srcATenDebug' It appears that they are triggered by *.c files in the ATen project, like \"srcTHvectorAVX.c\"\nWhy do you need Debug build?\nIsn't ATen meant to support both Debug and Release builds? FYI I'm involved with the project at which does support both build configurations, and we would like to be able to link our project with ATen.\nWell, you can remove the flags in as a workground.\nWe should patch this file to ignore this flag when building the library in debug mode. It shouldn't be too hard\nI catched similar errors , when I Tryed to build Debug Version I try to build Release and catched next errors e:pytorchatensrcatencpuvec256vec256int.h(184): note: see reference to function template instantiation 'at::vec256::anonymous-namespace'::Vec256<int16t::blend<1(at::vec256::anonymous-namespace'::Vec256<int16t)' being compiled 28e:pytorchatensrcatencpuvec256vec256int.h(147): error C3861: 'mm256extract_epi16': identifier not found What I should do?\nThere're two issues in your description. The first question is easy. You'll also have to specify the flag to 1. The second question is about your compiler. I guess you are using VS2015 or earlier versions of VS2017. Make sure the version in the output is 19.11 after your enter in CMD.\nYes I am using VS2015. My Compiler is 19.0.2. Must I use VS 2017 in any case?\nSorry, but the answer is yes. The AVX2 functions was in VS2017 and they are used in PyTorch. So VS2017 19.11 is a must.\nNo problem:) Thank you."], "neg": []}
{"query": ".. autofunction:: cumprod .. autofunction:: cumsum .. autofunction:: dist <ins> .. autofunction:: logsumexp </ins> .. autofunction:: mean .. autofunction:: median .. autofunction:: mode", "pos": ["Leaving this issue here for higher visibility. Originally reported by here: Also mentioned by Piggybacking off of this issue, has strange behavior (reported here: )\nI have fixed this locally. Could I send a PR?\nYes, go ahead!"], "neg": []}
{"query": "{ PyObject_GC_UnTrack(self); THPFunction_clear(self); <ins> self->cdata_ptr.~weak_ptr(); </ins> self->cdata.~PyFunction(); Py_TYPE(self)->tp_free((PyObject*)self); }", "pos": ["there is a bug. the code is from The code for installing cuda10.2 should be corrected to The code for installing cuda10.2 should be corrected to cc\nThank you for reporting the typo", "This is how I implement the decoder of a sequence to sequence model I'm not sure about the new autograd mechanics but this worked in the previous version. If I didn't make the unrolling codes a function it will work. It will also work with CPU. I compiled from source with Python 2.7, CUDA 8.0 and Cudnn 6\nThanks for the minimal reproduction! It broke OpenNMT and a big internal model but we weren't looking forward to isolating what parts of those had the issue.\nThat's great, I'll take a look today. Thanks!\nI've pushed a fix to .\nHow does the inference work in this case? I mean, once you train it how do you test it in the test phase?"], "neg": []}
{"query": "<ins> #pragma once // RAII structs to acquire and release Python's global interpreter lock (GIL) #include <Python.h> // Acquires the GIL on construction struct AutoGIL { AutoGIL() : gstate(PyGILState_Ensure()) { } ~AutoGIL() { PyGILState_Release(gstate); } PyGILState_STATE gstate; }; // Releases the GIL on construction struct AutoNoGIL { AutoNoGIL() : save(PyEval_SaveThread()) { } ~AutoNoGIL() { PyEval_RestoreThread(save); } PyThreadState* save; }; </ins>", "pos": ["Go to Look at the html. It contains the following: ! This problem manifests in google search. Searching for 1.8.1 docs brings up the \"1.8.0 docs\": ! We can probably fix this by doing a quick sed on the website but we should develop a solution for future releases. I think this has happened before in the past and am not sure what causes it. cc"], "neg": []}
{"query": "c = nn.Conv2d(3, 3, 3) self.assertRaises(RuntimeError, lambda: c(None)) <ins> def test_Conv2d_backward_twice(self): input = Variable(torch.randn(2, 3, 5, 5)) c = nn.Conv2d(3, 3, 3) o1 = c(input) o1.sum().backward() self.assertRaisesRegex(RuntimeError, 'Specify retain_variables=True', lambda: o1.sum().backward()) </ins> @unittest.skipIf(not TEST_CUDA, 'CUDA not available') def test_Conv2d_large_workspace(self): # These sizes require huge cuDNN workspaces. Make sure we choose a", "pos": ["Usually calling backwards twice gives this message: But conv2d just segfaults. $ pt A B Segmentation fault (core dumped)"], "neg": []}
{"query": "// For Convolution strategies that don't implicitly handle grad_bias, we add a helper // function here to perform it using simple Tensor operators static at::Tensor compute_grad_bias(const at::Tensor& grad_output) { <del> // grad_output is in N, C, H, W, we re-shape and reduce over spatial dims and batches </del> <ins> // grad_output is in N, C, H, W, we re-shape and reduce over spatial dims and batches </ins> return grad_output.contiguous().view({grad_output.size(0), grad_output.size(1), -1}).sum(0).sum(1); }", "pos": ["The following code: produces this error Also, removing the ReLU from the network: causes the error to change to: However, when the second convolution is removed as well: then the code runs error free I'm on a ppc64le architecture with a P100 GPU, CUDA 8.0 and PyTorch version '0.4.0a0+'\nI'm also having this problem using branch v0.3.0\ni could reproduce this as well. we'll look into this.\nI'm having the same error :/\nsame error with gradient penalty for WGAN\nWhen you execute , are you cause memory leak\nno there shouldn't be a memory leak.\nI am having an out of memory error as well, only when applying backward to the output of this method. I am using 3 titans GPUs so memory is not a problem I would assume. def computeGP(netD, realdata, realembed, fakedata, LAMBDA): BATCHSIZE = (0) alpha = (BATCHSIZE, 1) alpha = (BATCHSIZE, int(realdata.nelement() / BATCHSIZE)).contiguous().view(BATCHSIZE, 3, 64, 64) alpha = () interpolates = alpha realdata + ((1 - alpha) fakedata) interpolates = () interpolates = autograd.Variable(interpolates, requiresgrad=True) discinterpolates, = netD(interpolates, realembed) gradients = (outputs=discinterpolates, inputs=interpolates, (()).cuda(), creategraph=True, retaingraph=True, onlyinputs=True)[0] gradientpenalty = (((2, dim=1) - 1) * 2).mean() LAMBDA return gradient_penalty `\nI am also using it with gradient penalty for WGAN, But When I use gradient_penalty.backward(), the memory leak is happen, My pytorch is 0.4.0a0+.\nI can reproduce the memory leak with this: However the leak disappears when only one convolution is used: Also, if nn.Linear is used, there is no leak. Torch version: 0.4.0a0+\nI've opened a separate issue for that memory leak:"], "neg": []}
{"query": "\"torch/csrc/cuda/python_nccl.cpp\", ] if WITH_CUDNN: <del> main_libraries += ['cudnn'] library_dirs.insert(0, CUDNN_LIB_DIR) </del> <ins> main_libraries += [CUDNN_LIBRARY] </ins> # NOTE: these are at the front, in case there's another cuDNN in CUDA path include_dirs.insert(0, CUDNN_INCLUDE_DIR) if not IS_WINDOWS:", "pos": ["Traceback (most recent call last): File \"\", line 3, in <moduleimport model File \"/home/abc/github/chinese-\", line 5, in <modulefrom import crnnOcr File \"/home/abc/github/chinese-\", line 72, in <modulemodel,converter = crnnSource() File \"/home/abc/github/chinese-\", line 64, in crnnSource model = crnn.CRNN(32, 1, len(alphabet)+1, 256, 1).cuda() File \"/usr/local/lib/python2.7/dist-\", line 216, in cuda return (lambda t: t.cuda(device)) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 123, in apply self.flattenparameters() File \"/usr/local/lib/python2.7/dist-\", line 65, in flattenparameters if not or not (anyparam): File \"/usr/local/lib/python2.7/dist-\", line 82, in isacceptable if libcudnn() is None: File \"/usr/local/lib/python2.7/dist-\", line 58, in libcudnn 'but linked against {}'.format(compileversion, _cudnnversion)) RuntimeError: cuDNN version mismatch: PyTorch was compiled against 6021 but linked against 5110 How to solve the problem?\ncould you pull the latest master branch and recompile. I pushed a change yesterday that fixes this:\nI remove the old code and pull new code, then compile, install.It still fails. /usr/local/lib/python2.7/dist- UserWarning: No training configuration found in save file: the model was not compiled. Compile it manually.", "('No training configuration found in save file: ' Traceback (most recent call last): File \"\", line 3, in <moduleimport model File \"/home/qianbb/github/chinese-\", line 5, in <modulefrom import crnnOcr File \"/home/qianbb/github/chinese-\", line 72, in <modulemodel,converter = crnnSource() File \"/home/qianbb/github/chinese-\", line 64, in crnnSource model = crnn.CRNN(32, 1, len(alphabet)+1, 256, 1).cuda() File \"/usr/local/lib/python2.7/dist-\", line 216, in cuda return (lambda t: t.cuda(device)) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 100, in apply self.flattenparameters() File \"/usr/local/lib/python2.7/dist-\", line 66, in flattenparameters if not or not (anyparam): File \"/usr/local/lib/python2.7/dist-\", line 87, in isacceptable if libcudnn() is None: File \"/usr/local/lib/python2.7/dist-\", line 58, in libcudnn 'but linked against {}'.format(compileversion, _cudnnversion)) RuntimeError: cuDNN version mismatch: PyTorch was compiled against 6021 but linked against 5110 The library of cuda is shown as following screenshot. !\nI found the issue, I fixed it in which we will merge shortly.\nThanksDoes it have been merged?\nyes. it is merged.\nIt seems the fix doesn't work.", "The log is shown as following Traceback (most recent call last): File \"\", line 3, in <moduleimport model File \"/home/qianbb/github/chinese-\", line 5, in <modulefrom import crnnOcr File \"/home/qianbb/github/chinese-\", line 72, in <modulemodel,converter = crnnSource() File \"/home/qianbb/github/chinese-\", line 64, in crnnSource model = crnn.CRNN(32, 1, len(alphabet)+1, 256, 1).cuda() File \"/usr/local/lib/python2.7/dist-\", line 222, in cuda return (lambda t: t.cuda(device)) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 146, in apply (fn) File \"/usr/local/lib/python2.7/dist-\", line 100, in apply self.flattenparameters() File \"/usr/local/lib/python2.7/dist-\", line 66, in flattenparameters if not or not (anyparam): File \"/usr/local/lib/python2.7/dist-\", line 87, in isacceptable if libcudnn() is None: File \"/usr/local/lib/python2.7/dist-\", line 58, in libcudnn 'but linked against {}'.format(compileversion, _cudnnversion)) RuntimeError: cuDNN version mismatch: PyTorch was compiled against 6021 but linked against 5110\nthat is weird. Can I see the full log of in pytorch directory. Can you make sure you have latest commits with .\nI am sure the code is up-to-date. First I uninstall the code using the command pip unisntall torch. Then I remove the old code using the command rm -r pytorch and git clone again. The log of install is\nSorry to trouble you. Is there any new information on this issue?\nI spent some time resolving this issue and I don't see too much about it online. Try ensuring that your Pytorch build is correct for your specific versions of Cuda and Cudnn. While I had the correct version of Pytorch, I resolved the issue by installing a different build of the same version.", "The build I installed is . Here are the list of versions and builds: You can download these files locally and install with conda: (assumes command is being run in a directory containing , where this file will be specific to your Cuda and Cudnn version.\nfeedback from this code PackagesNotFoundError\nthank you - I have edited my comment to reflect that I installed from a local zip."], "neg": []}
{"query": "#  binary_linux_libtorch_3.6m_cpu_test: #    environment: #      BUILD_ENVIRONMENT: \"libtorch 3.6m cpu\" <del> #    resource_class: gpu.medium </del> <ins> #    resource_class: gpu.nvidia.small </ins> #    <<: *binary_linux_test # #  binary_linux_libtorch_3.6m_cu90_test: #    environment: #      BUILD_ENVIRONMENT: \"libtorch 3.6m cu90\" <del> #    resource_class: gpu.medium </del> <ins> #    resource_class: gpu.nvidia.small </ins> #    <<: *binary_linux_test # docker_build_job:", "pos": ["As it still using deprecated resource class, see See example of the failure in\nFollowup: we need to plan better --we knew it was gonna happen and none of us fixed it early enough. Action item: go through tasks about deprecation in sprint planning"], "neg": []}
{"query": "} } <del> return F::detail::batch_norm( </del> <ins> return torch::nn::functional::detail::batch_norm( </ins> input, this->running_mean, this->running_var,", "pos": ["As part of pull request , F is redefined globally in nn/modules/batchnorm.h F is then used once in batchnorm.h and again in instancenorm.h not a bug, but a problem for anyone using F in some other way. it's solvable, but I don't see this pattern anywhere else in libtorch, was wondering if this use is an anomaly, and if we could use thanks cc\nThanks a lot for reporting! Yes this is an anomaly, and will be fixed by"], "neg": []}
{"query": "(Trace, (), ((S, S),)), (Cross, (), ((S, 3), (S, 3))), (Cross, (1,), ((S, 3, S), (S, 3, S)), 'dim'), <ins> (Inverse, (), ((S, S),), '', (), [skipIfNoLapack]), </ins> (Clone, (), ((S, M, S),)), (Squeeze, (), ((S, 1, M, 1),)), # TODO: enable neg dim checks", "pos": ["Command to install the latest nightly binary: installs : However, PyTorch itself reports : I guess this version tag might be defined during the build process from e.g. and then forwarded from , but also source builds should use based on the tag. Unsure, if this is just a reporting issue or if we are building the nightly binaries with and executing them with , so creating this issue to dig a bit further into it. CC cc\nThis is really weird, I've put up measure in place to validate against it... Will investigate today (and moving to pytorch repo)\ncan you please check that 2.2.0 reports (and is build) against the correct version of nccl library?\nConfirmed for nightlies : Same issue for RC:\nLooking at the build log Looks like we are compiling with: USESYSTEMNCCL\nOn Windows I see following exception: We compiling with should we return None rather then exception ?\nIn this case we are indeed using 2.18.1: Do you know why we are using the system-wide NCCL installation instead of the one? Should we delete the system NCCL and use by default?\nLooks like its set here: Looks like it was by: Let me propose the PR to unset this flag : USESYSTEM_NCCL\nI guess we are using the system NCCL installation to be able to during the runtime. If we would use the module I assume we would link NCCL into the PyTorch binaries. If so, we should make sure to update the NCCL version whenever is updated. This could break our dependency on the NCCL pip package.\nYes. Also when we update this nccl here:\nWe used following path  What would be good path for nccl 2.19.3 and cuda 12.1 ? As per this issue: Will be using 12.0 ncck 2.19.3\nInstalling system NCCL into the nightly build docker container feels great, but we need to figure out, how to keep it in sync with PyTorch. I wonder if one can, for example, just check nccl version from this requirements file and build from source?\nyes probably we should build and upload nccl for our supported cuda versions, same way as we build docker containers, by following details here:\nVerified with final rc"], "neg": []}
{"query": "gate_size = hidden_size self._all_weights = [] <del> self._param_buf_size = 0 </del> for layer in range(num_layers): for direction in range(num_directions): layer_input_size = input_size if layer == 0 else hidden_size * num_directions", "pos": ["Since the new pytorch version, I have a warning when I that says: When I do as suggested, and then try to encode again, I get the following error though: Any idea?\ndo you have code the reproduces the issue?\nWhat was the fix? I still see the error on Pytorch Version: 0.2.03 AttributeError: 'LSTM' object has no attribute 'parambufsize'"], "neg": []}
{"query": "def backward(self, grad_output): tensors = self.saved_tensors input, params = tensors[0], tensors[1:] <del> grad_input = (input.new().resize_as_(input).zero_() </del> <ins> grad_input = (input.new().resize_as_(input) </ins> if self.needs_input_grad[0] else None,) grad_param = tuple(p.new().resize_as_(p).zero_() if self.needs_input_grad[i+1] else None for i, p in enumerate(params))", "pos": ["THNN convolutions are slower due to unnecessary .zero() calls on gradients: is not needed same for bias, prob worth adding a constant in C-side that controls whether to override grad or not\nSame for batch_norm and auto thnn functions, not sure if all of them though. Prob the speedups will be quite significant.\nI think they are needed. THNN functions accumulate the gradient instead of overwriting the buffer, don't they?\nonly grad wrt weight, inputs are overriden\nWe should reopen this, still many functions have unnecessary .zero() calls on gradinput and gradparam, e.g.: : : most of the functions don't need it\nI think the ones are currently necessary.\nfor nn yes, but cudnn has: and on cudnnConvolutionBackwardFilter and on cudnnBatchNormalizationBackward which control addition of gradients and should be set to 1 and 0 in pytorch cudnn (it is and 1 in )\nYeah, we can remove it on BatchNorm. If you look at Conv, it already only zeros out grad_weight for cunn.\nthis is now fixed in master:"], "neg": []}
{"query": "} void ObjectPtrAllocator::free(void* ptr) { <del> object = nullptr; </del> <ins> { AutoGIL gil; object = nullptr; } </ins> allocator->free(allocatorContext, ptr); delete this; } void StorageWeakRefAllocator::free(void* ptr) { <del> PyGILState_STATE gstate = PyGILState_Ensure(); PyObject_SetAttrString(object.get(), \"cdata\", Py_None); object = nullptr; PyGILState_Release(gstate); </del> <ins> { AutoGIL gil; PyObject_SetAttrString(object.get(), \"cdata\", Py_None); object = nullptr; } </ins> allocator->free(allocatorContext, ptr); delete this; }", "pos": ["When both PyTorch and Pandas are imported, calling results in a segfault. The issue was originally reported by It looks like calling returns a null pointer here: Examining in GDB: I suspect it's both PyTorch and Pandas doing threading magic and somehow getting in each other's way. I haven't yet been able to provide a minimal example that triggers this behaviour and does not use my package; it's likely that I'm doing something odd that causes this. More documentation here: Possibly related issues: and Python version: 3.6 PyTorch version: v0.2.0 Pandas version: 0.20.1"], "neg": []}
{"query": "global Storage Tensor = _import_dotted_name(t) Storage = _import_dotted_name(t.replace('Tensor', 'Storage')) <ins> if 'cuda' in t: import torch.cuda torch.cuda.init() </ins> _C._set_default_tensor_type(Tensor)", "pos": ["When I was working on Vision and Audio CI, I often find that the dowloaded nightly PyTorch is CPU version instead of expected CUDA version. It will break the domain library CIs. It's often caused by nightly build failed or that day's nightly build isn't finished. There are 2 open issues for it. If this day's nightly PyTorch cuda build isn't available, I think it's better to provide last day's nightly builds rather than CPU version. cc\nIt looks that all nightly builds would be finished before 4 AM at PST, so vision CI failure probability might be higher in the morning of France.", "Hi, I've updated pytorch to the lastest master () from github, but got an error when I use : What to do for workaround? Thanks! Jin\nHere's a workaround. I think the problem is that pytorch hasn't initialized CUDA yet. At any rate, the following works for me:\nThanks but your workaround doesn't work for me. The error message still shows. Jin\nSorry, I meant the following: Could you please try if that works for you?\nYep. Just what I thought and was going to add comment here :smile: It works. Thank you! Jin"], "neg": []}
{"query": "<del> CMAKE_MINIMUM_REQUIRED(VERSION 3.0) </del> <ins> CMAKE_MINIMUM_REQUIRED(VERSION 2.8) </ins> SET(CMAKE_MODULE_PATH ${CMAKE_CURRENT_SOURCE_DIR}/cmake ${CMAKE_MODULE_PATH}) ################################################################################", "pos": ["Currently, CMake 3.0 or higher is required for THPP. This is annoying for building from source on Ubuntu 14.04 systems, as default CMake is 2.8, and updating it requires either compiling from source or adding some random repos:\nit should work fine if you change it to 2.8. I'm testing this out in a sec, and will push it into master\nNope, it does not for me."], "neg": []}
{"query": "} } <del> PyObject * THSPTensor_(NewEmpty)() { auto r = THSTensor_(new)(LIBRARY_STATE_NOARGS); THSPTensor_(initStorage)(r); return THSPTensor_(New)(r); } </del> PyObject * THSPTensor_(New)(THSTensor *tensor) { THSTensorPtr ptr(tensor);", "pos": ["This causes to not work properly on newly created empty sparse tensors. In the example below, both should print 1 instead."], "neg": []}
{"query": "if hasattr(lib, 'cudnnGetErrorString'): lib.cudnnGetErrorString.restype = ctypes.c_char_p __cudnn_version = lib.cudnnGetVersion() <ins> compile_version = torch._C._cudnn_version() # Check that cuDNN major and minor versions match if (__cudnn_version // 100) != (compile_version // 100): raise RuntimeError( 'cuDNN version mismatch: PyTorch was compiled against {} ' 'but linked against {}'.format(compile_version, __cudnn_version)) </ins> else: lib = None return lib", "pos": ["We should check that the version from cudnn.h matches the version from the loaded shared library. Really weird things happen if they don't match."], "neg": []}
{"query": "THPUtils_setError(error.c_str()); } <del> #if defined(NUMPY_TYPE_ENUM) || defined(THC_GENERIC_FILE) </del> <ins> #if defined(NUMPY_TYPE_ENUM) || (defined(WITH_NUMPY) && defined(THC_GENERIC_FILE)) </ins> #ifndef THC_REAL_IS_HALF #define load_real real", "pos": ["When I don't have a working numpy dependency detected (due to unrelated shenanigans), the build fails with this: And then a big pile of related errors. CC\nThis is most likely the problem:\nCC\nOk, should probably read\nShould I go ahead and make the change? I'll also remove the at line 193. The rest looks ok to me.\ngo for it.\nActually, the change I proposed is incorrect (I misread the second defined), but I'm on it."], "neg": []}
{"query": "SparseTensor = getattr(torch.cuda.sparse, tensor_type) else: SparseTensor = getattr(torch.sparse, tensor_type) <ins> padding_idx = ctx.padding_idx indices = indices.view(1, -1) grad_output = grad_output.view(-1, ctx._weight_size[1]) if padding_idx is not None: nonpadding_indices_indices = (indices.view(-1) != padding_idx).nonzero().view(-1) indices = indices.index_select(1, nonpadding_indices_indices) grad_output = grad_output.index_select(0, nonpadding_indices_indices) </ins> grad_weight = SparseTensor( <del> indices.view(1, -1), grad_output.view(-1, ctx._weight_size[1]), </del> <ins> indices, grad_output, </ins> ctx._weight_size, ) return None, grad_weight, None, None, None, None, None", "pos": ["As reported here with repro. When sparse=True, paddingidx should probably error out (as the easiest solution). Alternatively, the gradients for paddingidx indices should be zeroed out explicitly in the sparse=True backward codepath."], "neg": []}
{"query": "ensure_attr('ClassNLLCriterion', 'weights') ensure_attr('ParallelCriterion', 'repeatTarget') ensure_attr('MultiMarginCriterion', 'weights') <del> ensure_attr('SpatialConvolution', 'bias', 'finput', 'fgradInput', 'gradWeight', 'gradBias') </del> <ins> ensure_attr('SpatialConvolution', 'bias', 'finput', 'fgradInput', 'gradWeight', 'gradBias', '_gradOutput') </ins> attr_map('ReLU', {'val': 'value'}) attr_map('Threshold', {'val': 'value'}) attr_map('Unsqueeze', {'pos': 'dim'})", "pos": ["On Windows, torch.slogdet reproducibly fails when computing torch.slogdet on larger tensors. Here is the error: Steps to reproduce the behavior: Expected this op to succeed on Windows. This does not fail on Linux. cc\nWhich GPU card are you using on Windows?\nI have an RTX 3080 and an RTX 3090 and see the issue on both. If this is a CUDA issue, could you point me to the right place to report it?\nI can reproduce this with PyTorch 1.7.0 + RTX 3080.\nIt seems that the current implementation of this function is based on MAGMA. Will try to rebuild with newer versions.\nIt seems that the current windows MAGMA binaries are built without cuda arch 8.0 support. I will fix it.\nAwesome! Thank you so much for the quick turnaround on this. Please let me know if you want me to test a fix.\nWait for the merge of"], "neg": []}
{"query": "return: self cname: resize cpu_half: True <ins> before_call: THPUtils_assert(arg_self->storage->flag & TH_STORAGE_RESIZABLE, \"calling resize_ on a tensor that has non-resizable storage. Clone it first \" \"or create a new tensor instead.\"); </ins> arguments: - THTensor* self - arg: THSize* size", "pos": ["When both PyTorch and Pandas are imported, calling results in a segfault. The issue was originally reported by It looks like calling returns a null pointer here: Examining in GDB: I suspect it's both PyTorch and Pandas doing threading magic and somehow getting in each other's way. I haven't yet been able to provide a minimal example that triggers this behaviour and does not use my package; it's likely that I'm doing something odd that causes this. More documentation here: Possibly related issues: and Python version: 3.6 PyTorch version: v0.2.0 Pandas version: 0.20.1"], "neg": []}
{"query": "return type(self), (self.tolist(),) def __repr__(self): <del> return str(self) </del> <ins> return repr(str(self)) </ins> def __str__(self): <del> return _tensor_str._str(self) </del> <ins> # All strings are unicode in Python 3, while we have to encode unicode # strings in Python2. If we can't, let python decide the best # characters to replace unicode characters with. if sys.version_info > (3,): return _tensor_str._str(self) else: return _tensor_str._str(self).encode( sys.stdout.encoding or 'UTF-8', 'replace') </ins> def __iter__(self): return iter(map(lambda i: self.select(0, i), _range(self.size(0))))", "pos": ["It seems that the following link that can be reached from here (check the screenshot) is broken: Page: Broken link: Screenshot: ! cc\nthis should be transferred to , but I don't have permissions to do that. Fix is available at pytorch/hub."], "neg": []}
{"query": "import math from bisect import bisect_right <ins> from functools import partial </ins> from .optimizer import Optimizer", "pos": ["I'm working on checkpointing in our nn trainer. When saving the state objects, I get an error: PicklingError: Can't pickle <type 'function': attribute lookup builtin.function failed I tracked it down to trying to save the scheduler ReduceLROnPlateau. I think the issue is due to the scheduler referencing the optimizer, which has to be saved with statedict(). I presume the same error will occur with other schedulers too. It seems like the problem could be solved by providing a statedict() and loadstatedict() in the scheduler classes. pytorch master rev (Jan 22) installed from source x86_64 on centos 7.4 cuda 9.0 cudnn 7 k80 gcc 5.4 python 2.7.14\nThe following script reproduces the crash: import torch import from import ReduceLROnPlateau if name == 'main': model = (10,10) optim = (model.parameters()) sched = ReduceLROnPlateau(optim, mode='max', factor=0.5) with open('','wb') as f: # Works # ((model.statedict(), optim.statedict()), f) # Crashes ((model.statedict(), optim.state_dict(), sched), f)"], "neg": []}
{"query": "ctx.num_inputs = len(inputs) ctx.input_device = inputs[0].get_device() outputs = comm.broadcast_coalesced(inputs, ctx.target_gpus) <ins> non_differentiables = [] for idx, input_requires_grad in enumerate(ctx.needs_input_grad[1:]): if not input_requires_grad: for output in outputs: non_differentiables.append(output[idx]) ctx.mark_non_differentiable(*non_differentiables) </ins> return tuple([t for tensors in outputs for t in tensors]) @staticmethod", "pos": ["PyTorch version: 0.3.0 Python version: 3.5 GPU models and configuration: 8x NVIDIA Titan I found that requiresgrad setting in module is ignored when module is wrapped with nn.DataParallel. For example, This code should be executed because very small part of network activation (only fc layer) is stored for backward computation, but it results to run-time error: Internally, I found a crack in function which is in . In replicate function, it copies all parameter in module (# of replica times) with . In broadcasting code, it just defines new with default constructor parameter, which is always set to True. I think there can be some choices to fix this issue. is intended behavior for DataParallel, so we should use volatile (or ) to implement transfer learning in DataParallel model should be fixed to handle requires_grad when copying parameters. should be synchronizing in I think any of solution doesn't need lots of effort to fix it."], "neg": []}
{"query": "THCDeviceTensor<THCIndex_t, 1> target, THCDeviceTensor<Dtype, 1> output, Dtype *weights, <ins> int n_classes, </ins> int ignore_index) { CUDA_KERNEL_LOOP(index, batch_size) {", "pos": ["Claim: When I use an invalid label with F.crossentropy with reduce=False, I don't receive an error message. I do receive an error message when reduce=True. Code: Result: No error message received. Expected error message: Context: I was using , accidentally used an invalid label, and got extremely large losses and extremely small losses, and inconsistently so. Environment: OS: Ubuntu 16.04 on Docker and host PyTorch version: 0.3.0.post4 How you installed PyTorch: conda Python version: 3.6.3"], "neg": []}
{"query": "skipTestIf = test[5] if len(test) >= 6 else [] <del> test_output_indices = test[6] if len(test) >= 7 else None </del> <ins> output_process_fn = test[6] if len(test) >= 7 else lambda x: x </ins> for dim_perm in product([-1, 1], repeat=len(dim_args_idx)): test_name = basic_test_name", "pos": ["I am having problem in running the backward propagation with svd. In particular, it seems working when the first dimension of the matrix x (below) is greater than the second. But I get the following error if the first is smaller than the second. PyTorch version: 2018.04.20 Is debug build: No CUDA used to build PyTorch: 8.0.61 OS: Ubuntu 16.04.4 LTS GCC version: (Ubuntu 5.4.0-6ubuntu1~16.04.9) 5.4.0 CMake version: Could not collect Python version: 3.5 Is CUDA available: No CUDA runtime version: No CUDA GPU models and configuration: No CUDA Nvidia driver version: No CUDA cuDNN version: No CUDA Versions of relevant libraries: [pip3] numpy (1.13.3) [pip3] torch (2018.4.20) [pip3] torchvision (0.2.0) [conda] pytorch 0.3.1 py35cuda8.0.61cudnn7.0.52 pytorch [conda] pytorch-nightly 2018.04.20 py35cuda8.0.61cudnn7.1.21 pytorch [conda] torchvision 0.2.0 py35heaa392f_1 pytorch\nRepro'ed on 0.4 branch. works btw.\nWhat does \"Repro'ed\" means? Should I install from that branch? I need to compute the kernel of a matrix, therefore I need the last columns of the matrix V (which is why I have to pass ).\nIt's a shorthand for \"reproduced\", which means that the bug still hasn't been fixed. Sorry, we'll try to get to it soon."], "neg": []}
{"query": "y.sum().backward() self.assertEqual(x.grad.data, torch.ones(5, 5) * 2) <ins> @unittest.skipIf(torch.cuda.device_count() < 2, \"no multi-GPU\") def test_backward_device(self): # check that current device matches the variable's device device = [None] class Identity(torch.autograd.Function): @staticmethod def forward(ctx, x): return x.clone() @staticmethod def backward(ctx, grad_output): device[0] = torch.cuda.current_device() return grad_output.clone() v = Variable(torch.randn(1).cuda(1), requires_grad=True) Identity.apply(v).backward() self.assertEqual(device[0], 1) </ins> def test_detach(self): x = Variable(torch.randn(10, 10), requires_grad=True) y = x + 2", "pos": ["Consider following code, and run it with multi gpus (e.g. 4): It will output: That is, the default device would always be zero even in DataParallel. My pytorch version is 0.1.122. I think this is not the desired behavior. It would cause some troubles. I tried to insert my own cuda kernel into backward to calculate the gradients, it become very slow, and I fixed it by (grad.getdevice()). Anyway, I think currentdevice in forward and backward should be the same, or could anyone explain to me why they are different?\nthis seems like something we should fix: ?\n<!-- drci-comment-start --:pagefacingup: Preview :pagefacingup: Preview :question: Need help or want to give feedback on the CI? Visit our Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. <!-- drci-comment-end --"], "neg": []}
{"query": "with values from the interval ``[start, end)`` taken with common difference :attr:`step` beginning from `start`. <ins> Note that non-integer `step` is subject to floating point rounding errors when comparing against `end`; to avoid inconsistency, we advise adding a small epsilon to `end` in such cases. </ins> .. math:: out_{i+1} = out_{i} + mathrm{step}", "pos": ["! OS: Ubuntu 17.10 PyTorch version: 0.3.1 How you installed PyTorch (conda, pip, source): conda Python version: 3.6.4 CUDA/cuDNN version: - GPU models and configuration: -\nThis is a floating point problem that has been mitigated in master with this commit: But even with that fix, arange doesn't always give the expected result. You can avoid this issue by doing\nIt\u2019s impossible to have arange work in all kinda of possible conditions due to rounding errors in floating point math. What we did in master was to make it consistent with numpy. An alternate solution would be to avoid having the length of the interval be exactly divisible by its length (you can make it a tiny bit larger).\nThe numpy docs have a note saying \"When using a non-integer step, such as 0.1, the results will often not be consistent.\" Maybe this could be to the doc as well?"], "neg": []}
{"query": "torch/csrc/nn/THNN_generic.cwrap torch/csrc/nn/THNN_generic.cpp torch/csrc/nn/THNN_generic.h <ins> torch/csrc/generated </ins> docs/src/**/* test/data/legacy_modules.t7 test/data/gpu_tensors.pt", "pos": ["Trying to run on github actions via AWS instance gives me: I will of course set up now, but it shouldn't crash like that. May be it should use as a fallback? pt-2.0 cc\nIs this a multitenant setup? Also curious which distro are you running, I was under the impression that most linux distros set home by default\nthis is AWS image. one default user. I think it's user, but it wants you to run as root, which is weird. I have manually now set to overcome this.\nWe should update triton to use pathlib or something"], "neg": []}
{"query": ":math:`p_n > 1` increases the recall, :math:`p_n < 1` increases the precision. For example, if a dataset contains 100 positive and 300 negative examples of a single class, <del> then `pos_weight` for the class should be equal to math:`frac{300}{100}=3`. </del> <ins> then `pos_weight` for the class should be equal to :math:`frac{300}{100}=3`. </ins> The loss would act as if the dataset contains math:`3times 100=300` positive examples. Args:", "pos": ["is not available in the docs page, but the doc string has been in . This was mentioned earlier in", "Oh, bilinear, I missed that last time I looked at it. I'll add a docstring.", "returns nothing. Might be useful to have some linter in the build process to find Python functions missing", "Pytorch 0.4.0 gives a segmentation fault with dilated 2D convolutions over a specific size. import torch import as nn model = nn.Conv2d(inchannels=64, outchannels=64, kernelsize=5, dilation=2) model(([1, 64, 1166, 1166])) # This runs model(([1, 64, 1167, 1167])) # This crashes PyTorch version: 0.4.0 Is debug build: No CUDA used to build PyTorch: 9.1.85 OS: CentOS Linux 7 (Core) GCC version: (GCC) 4.8.5 (Red Hat 4.8.5-28) CMake version: version 2.8.12.2 Python version: 3.6 Is CUDA available: Yes CUDA runtime version: 8.0.61 GPU models and configuration: GPU 0: GeForce GTX 1080 Ti GPU 1: TITAN X (Pascal) GPU 2: Quadro K2200 Nvidia driver version: 390.59 cuDNN version: Probably one of the following: /usr/local/cuda- /usr/local/cuda- /usr/local/cuda-7.5/lib64/libcudnnstatic.a /usr/local/cuda- /usr/local/cuda-8.0/lib64/libcudnn_static.a\ndo you think this is the same as the segfault you showed me a week ago on convolution?\nah yes. int overflow. I only fixed conv3d because I didn\u2019t want to hurt perf. I\u2019ll fix all of them.", "This should be an easy one.. The function is implemented, just no sign of it in the docs."], "neg": []}
{"query": "{ transpose_m1 = (transpose_r == 'n' ? 't' : 'n'); m1_ = THTensor_(newContiguous)(m1); <ins> free_m1 = 1; </ins> } /* m2 */", "pos": ["Here is an example of what seems to be a memory leak in pytorch, although it could be numpy? The leak was only occurring when using the or array unsqueezing methods in numpy. Also worth noting that the issue is resolved by removing the from and casting the torch tensor to type. An easy workaround is to reshape the array after turning it into a torch Tensor."], "neg": []}
{"query": "l.param_attr = None self.assertEqual(num_params(), 3) <ins> def test_assignment(self): l = nn.Module() a = nn.Parameter(torch.randn(2)) b = nn.Parameter(torch.randn(3)) c = nn.Parameter(torch.randn(4)) q = nn.Linear(4, 4) r = nn.Linear(5, 5) w = nn.Linear(6, 6) def test_assignments(get_list, a, b, c): # Check that None can be shadowed l.a = None self.assertIsNone(l.a) self.assertIn('a', l.__dict__) l.a = a self.assertIs(l.a, a) self.assertEqual(get_list(), [a]) self.assertNotIn('a', l.__dict__) # Assign second object l.b = None self.assertIsNone(l.b) self.assertIn('b', l.__dict__) l.b = b self.assertIs(l.b, b) self.assertEqual(get_list(), [a, b]) self.assertNotIn('b', l.__dict__) # Remove and add the object back. Order should be unchanged. l.a = None self.assertIsNone(l.a) self.assertEqual(get_list(), [b]) l.a = a self.assertIs(l.a, a) self.assertEqual(get_list(), [a, b]) # Replace object with another one. Order should be unchanged. l.a = c self.assertIs(l.a, c) self.assertEqual(get_list(), [c, b]) # Remove and reassign an attribute. It should appear at the end of the list now. del l.a self.assertFalse(hasattr(l, 'a')) l.a = a self.assertIs(l.a, a) self.assertEqual(get_list(), [b, a]) test_assignments(lambda: list(l.parameters()), a, b, c) del l.a, l.b self.assertEqual(list(l.parameters()), []) test_assignments(lambda: list(l.children()), q, r, w) del l.a, l.b self.assertEqual(list(l.children()), []) buf = torch.randn(10) l.register_buffer('buf', buf) self.assertIs(l.buf, buf) l.buf = None self.assertIs(l.buf, None) self.assertNotIn('buf', l.__dict__)  # should be stored in l._buffers l.buf = buf self.assertIn('buf', l.state_dict()) self.assertIs(l.state_dict()['buf'], buf) </ins> @unittest.skipIf(not TEST_CUDA, 'CUDA not available') def test_Conv2d_large_workspace(self): # These sizes require huge cuDNN workspaces. Make sure we choose a", "pos": ["They do one pair of these calls for each slicing operation.\nDoes this actually have a performance impact? These calls tend to be very quick\nYes, it's a major reason why small unrolled LSTMs (which involve in addition to two GEMMs: a chunk into four gates, for 4 pairs of calls, five unary pointwise ops, for 5 more, and five binary pointwise ops, for another 5) are much slower than they are in Theano and TensorFlow.\nAlso, there are excessive cudaGetDevice calls for each pointwise operation (9 cudaGetDevice) even without chunking, as in the following repro code: As for small unrolled LSTMs it does show up.\nThis should no longer be an issue, now that is implemented as a native ATen function. I think we're still using the Python version for tensors, but this will be resolved soon when we stop using tensors altogether."], "neg": []}
{"query": "<ins> import os this_file = os.path.dirname(os.path.abspath(__file__)) generated_dir = os.path.abspath(os.path.join(this_file, '..', '..', 'torch', 'csrc', 'generated')) line_start = '//generic_include ' types = [ 'Double', 'Float', 'Half', 'Long', 'Int', 'Short', 'Char', 'Byte' ] generic_include = '#define {lib}_GENERIC_FILE \"{path}\"' generate_include = '#include \"{lib}/{lib}Generate{type}Type.h\"' def split_types(file_name): assert file_name.startswith('torch/csrc/') if not os.path.exists(generated_dir): os.makedirs(generated_dir) with open(file_name, 'r') as f: lines = f.read().split('n') # Find //generic_include for i, l in enumerate(lines): if l.startswith(line_start): args = l[len(line_start):] lib_prefix, generic_file = filter(bool, args.split()) break else: raise RuntimeError(\"generic include not found\") gen_name_prefix = file_name[len('torch/csrc/'):].replace('/', '_').replace('.cpp', '') gen_path_prefix = os.path.join(generated_dir, gen_name_prefix) prefix = 'n'.join(lines[:i]) suffix = 'n'.join(lines[i + 1:]) to_build = [] g_include = generic_include.format(lib=lib_prefix, path=generic_file) for t in types: t_include = generate_include.format(lib=lib_prefix, type=t) gen_path = gen_path_prefix + t + '.cpp' to_build.append(gen_path) with open(gen_path, 'w') as f: f.write(prefix + 'n' + g_include + 'n' + t_include + 'n' + suffix) return to_build </ins>", "pos": ["Trying to run on github actions via AWS instance gives me: I will of course set up now, but it shouldn't crash like that. May be it should use as a fallback? pt-2.0 cc\nIs this a multitenant setup? Also curious which distro are you running, I was under the impression that most linux distros set home by default\nthis is AWS image. one default user. I think it's user, but it wants you to run as root, which is weird. I have manually now set to overcome this.\nWe should update triton to use pathlib or something"], "neg": []}
{"query": "self.add_module(str(idx), module) def __getitem__(self, idx): <del> if not (-len(self) <= idx < len(self)): raise IndexError('index {} is out of range'.format(idx)) if idx < 0: idx += len(self) it = iter(self._modules.values()) for i in range(idx): next(it) return next(it) </del> <ins> if isinstance(idx, slice): return Sequential(OrderedDict(list(self._modules.items())[idx])) else: if not (-len(self) <= idx < len(self)): raise IndexError('index {} is out of range'.format(idx)) if idx < 0: idx += len(self) it = iter(self._modules.values()) for i in range(idx): next(it) return next(it) </ins> def __len__(self): return len(self._modules)", "pos": ["it'd be great to pick out elements of of a sequential with array slicing:\nI agree this idea!\nHowever,\nthis has come up again (last in ). Shall we reconsider?\nI attempted getting a to work. Is this fine? If this is fine, I can issue a PR too. Let me know what you think. I believe this modification needs to be done in the class.\ncc: This issue has been open for quite sometime."], "neg": []}
{"query": "cudnn=True, ), dict( <ins> fullname='Conv2d_groups_thnn', constructor=lambda: nn.Conv2d(4, 6, (3, 2), groups=2), input_size=(2, 4, 6, 5), ), dict( </ins> module_name='ConvTranspose2d', constructor_args=(3, 4, 3, (3, 2), 1, (1, 1)), cudnn=True,", "pos": ["x = Variable(((16, 64, 8, 8))) res = nn.Conv2d(16, 16, kernelsize=3, groups=2, bias=False)(x) results in: return ('updateoutput', input, weight, bias) File \"\", line 158, in thnn groupedargs += [group(t) for t in args] File \"\", line 153, in group dim = 0 if () == 1 else 1 AttributeError: 'NoneType' object has no attribute 'dim' The in line 158 is because there is no bias. I am running on CPU so this is in the code."], "neg": []}
{"query": ":members: :hidden:`InstanceNorm1d` <del> ~~~~~~~~~~~~~~~~~~~~~ </del> <ins> ~~~~~~~~~~~~~~~~~~~~~~~~ </ins> .. autoclass:: InstanceNorm1d :members: :hidden:`InstanceNorm2d` <del> ~~~~~~~~~~~~~~~~~~~~~ </del> <ins> ~~~~~~~~~~~~~~~~~~~~~~~~ </ins> .. autoclass:: InstanceNorm2d :members: :hidden:`InstanceNorm3d` <del> ~~~~~~~~~~~~~~~~~~~~~ </del> <ins> ~~~~~~~~~~~~~~~~~~~~~~~~ </ins> .. autoclass:: InstanceNorm3d :members:", "pos": ["why 1.21.6? because that's the oldest supported numpy on windows/python3.10: all cc\nHere is the test for pinning numpy to 1.21.6 The reason its was not pinned to this version is that conda can't properly resolve all the dependencies at this step here:\nIt looks like you are correct, I will put up a PR to pin lower then already pinned:\nHere is the PR:"], "neg": []}
{"query": "for t in types: tensor = t() gpu_tensor = get_gpu_type(t)() <ins> # Default values desc = '' type_subset = types no_inplace = False </ins> if len(decl) == 3: name, constr, arg_constr = decl <del> desc = '' </del> elif len(decl) == 4: name, constr, arg_constr, desc = decl elif len(decl) == 5: name, constr, arg_constr, desc, type_subset = decl <del> if t not in type_subset: continue </del> <ins> elif len(decl) == 6: name, constr, arg_constr, desc, type_subset, no_inplace = decl if t not in type_subset: continue </ins> precision = custom_precision.get(name, TestCuda.precision) for inplace in (True, False): <ins> if inplace and no_inplace: continue </ins> if inplace: name_inner = name + '_' else:", "pos": ["On recent PyTorch master (built from source; 0.4.0a0+) on a P100, CUDA 9:\nI think the results on CPU meets the defination of triu (upper triangular part of the matrix). It is really hard to handle with a tensor from expand operation with multi-thread not only on GPU but also on CPU. The unexpected result will also happen on CPU if we apply openmp to it. I think it is necessary to eleminate the difference between the implementation and defination.\nIt doesn\u2019t matter if it\u2019s cpu or not. triu doesn\u2019t have an underscore which means it doesn\u2019t operate in-place and should always give the same result (in new memory)"], "neg": []}
{"query": "| :attr:`stride` controls the stride for the cross-correlation. | If :attr:`padding` is non-zero, then the input is implicitly zero-padded on both sides <del> for :attr:`padding` number of points | :attr:`dilation` controls the spacing between the kernel points. It is harder to describe, but this `link`_ has a nice visualization of what :attr:`dilation` does. | :attr:`groups` controls the connections between inputs and outputs. </del> <ins> for :attr:`padding` number of points. | :attr:`dilation` controls the spacing between the kernel points; also known as the \u00e0 trous algorithm. It is harder to describe, but this `link`_ has a nice visualization of what :attr:`dilation` does. | :attr:`groups` controls the connections between inputs and outputs. `in_channels` and `out_channels` must both be divisible by `groups`. </ins> |       At groups=1, all inputs are convolved to all outputs. |       At groups=2, the operation becomes equivalent to having two conv layers side by side, each seeing half the input channels, and producing half the output channels, and both subsequently concatenated. <ins> At groups=`in_channels`, each input channel is convolved with its own set of filters (of size `out_channels // in_channels`). </ins> The parameters :attr:`kernel_size`, :attr:`stride`, :attr:`padding`, :attr:`dilation` can either be:", "pos": ["Hi, I have installed two versions of CUDA. But I want to use one of it, can I specify it during setup process?\nTry running\nI have tried it, () still returns False. Thx for your reply!\nDo you have the cuda driver installed?\nOf course I do. It raise assertion error which shows \"The NVIDIA driver on your system is too old(found version 8000)\".\nthat's the answer to your problem. upgrade your driver.\nThx!", "When building from source, I got this error. My system is OSX 10.12.4, conda, CUDA 8.0. Any thoughts? Thanks!\ncan you try adding to the file and see if that fixes things? if so i'll commit it.\nactually, that file already has memory included. Are you following instructions from Especially: This part is important:\nworks nicely! Thanks a lot!", "The suggest that takes a as a parameter, but it actually must be wrapped in a Variable first.\nyou are looking at the Tensor docs. When operating on Variable, it is implied from the tutorials that all functions operating on Variables have to use Variables. We didn't write separate docs for Variable itself.", "something wrong with import torchvision Traceback (most recent call last): File \"\", line 3, in <moduleimport torchvision File \"\", line 4, in <moduleimport torchvision.transforms as transforms ImportError: No module named transforms\nyou have a file called for some reason. delete it or change your pythonpath to ignore it.\nHey, I have the same problem persisting when I import torchvision. I have no other file by the same name. ImportError Traceback (most recent call last) <ipython-input-6-3baf6837d8f1in <module() 8 from import Dataset 9 from import DataLoader ---10 from torchvision import transforms ImportError: No module named torchvision Any idea why this is happening?\nDo you have torchvision installed (it is a separate package)?\npip Install torchvision try this\nfrom data import BaseTransform,VOC_CLASSES as labelmap this is not working\npip install torchvision", "Hi, I need to pass input through one nn.Module, then argmax the output, and then pass that output to a second nn.Module. I want to backprop through the argmax back to the weights of the first module. That's impossible right now, though, it seems. When I loop through the gradient of the weights of my first module (with the following code): firstmodel.zerograd() secondmodel.zerograd() loss.backward() # retainvariables=True for param in secondmodel.parameters(): print( 'data: ', ) print( 'grad data: ', ) for param in firstmodel.parameters(): print( 'data: ', ) print( 'grad data: ', ) I get the following error, stating that those gradients can no longer be viewed/exist: -main()\nargmax is not differentiable (and even if you try to think about it as a function differentiable almost everywhere, its derivative is always 0, so you won't get anything meaningful).", "def inithidden(self): h0 = Variable(FloatTensor(self.numlayers, self.batchlen, self.hiddendim).zero(),  requiresgrad=True) c0 = Variable(FloatTensor(self.numlayers, self.batchlen, self.hiddendim).zero(),  requiresgrad=True) return (h0, c0) def forward(self, sequences): lstmout, = ((1, self.batchlen, self.inputdim), ) ... = nn.LSTM(inputdim, hiddendim, numlayers, dropout = dropout, batchfirst=True) = self.inithidden() output = model(traininput) produces the following error although I defined weight and input dimensions as in the documentation: File \"/usr/local/lib/python2.7/dist-\", line 202, in call result = self.forward(input, kwargs) File \"/usr/local/lib/python2.7/dist-\", line 91, in forward output, hidden = func(input, self.allweights, hx) File \"/usr/local/lib/python2.7/dist-\", line 327, in forward return func(input, fargs, fkwargs) File \"/usr/local/lib/python2.7/dist-\", line 227, in forward nexth, output = func(input, hidden, weight) File \"/usr/local/lib/python2.7/dist-\", line 67, in forward hy, output = inner(input, hidden[l], weight[l]) File \"/usr/local/lib/python2.7/dist-\", line 96, in forward hidden = inner(input[i], hidden, *weight) File \"/usr/local/lib/python2.7/dist-\", line 22, in LSTMCell gates = F.linear(input, wih, bih) + F.linear(hx, whh, bhh) File \"/usr/local/lib/python2.7/dist-\", line 744, in add return (other) File \"/usr/local/lib/python2.7/dist-\", line 293, in add return (other, False) File \"/usr/local/lib/python2.7/dist-\", line 287, in add return Add(inplace)(self, other) File \"/usr/local/lib/python2.7/dist-\", line 20, in forward return a.add(b) RuntimeError: inconsistent tensor size at /data/users/soumith/builder/wheel/pytorch-src/torch/lib/TH/generic/THTensorMath.c:827\nI can't reproduce your problem. Could you include a short runnable example that exhibits this behavior?", "You also don't need to manually init hidden anymore, the following would work as well...\nThanks! After removing problem solved.", "Hi, I would like to implement a sharing weight 'alexnet' with two input batches of images, and the fully connected layers are concatenated together like this: However, when I run it, I come across the following error: Does anyone know about it? Thank you very much.\nYou forgot to call on the output of . In the future please remember that we're using the issues for bug reports and feature requests only. If you have questions or need help, please post ."], "neg": []}
{"query": "RECORD_TRACE = CodeTemplate(\"\"\" if (jit::tracer::isTracing({ ${tensor_args} })) { <del> jit::Node *n = jit::tracer::recordTrace( \"${api_name}\", { ${tensor_args} }, ${return_name} ); </del> <ins> jit::Node *n = jit::tracer::recordTrace( \"${api_name}\", ${trace_inputs}, ${trace_outputs} ); </ins> ${record_attributes} } \"\"\")", "pos": ["when executing:\nI'm seeing this as well as of last night. Its preceded by many warnings about : But the first error is the one about . I went back, and I get a good install as late as commit"], "neg": []}
{"query": "#include \"tuple_parser.h\" <ins> #include <string> </ins> #include \"torch/csrc/DynamicTypes.h\" #include \"torch/csrc/autograd/python_variable.h\" #include \"python_numbers.h\"", "pos": ["Version include <stringin some cpp files: define order in The pre-defined struct is not passed by clang in macOS, so we should reorder them.\nSimple fix"], "neg": []}
{"query": "#ifndef TH_GENERIC_FILE #define TH_GENERIC_FILE \"generic/Tensor.cpp\" #else", "pos": ["When I don't have a working numpy dependency detected (due to unrelated shenanigans), the build fails with this: And then a big pile of related errors. CC\nThis is most likely the problem:\nCC\nOk, should probably read\nShould I go ahead and make the change? I'll also remove the at line 193. The rest looks ok to me.\ngo for it.\nActually, the change I proposed is incorrect (I misread the second defined), but I'm on it."], "neg": []}
{"query": "if (free_b) THTensor_(free)(b);), \"gels\", info,\"\"); <del> /* rb__ is currently ldb by nrhs; resize it to n by nrhs */ rb__->size[0] = n; if (rb__ != rb_) </del> <ins> /* * In the m < n case, if the input b is used as the result (so b == _rb), * then rb_ was originally m by nrhs but now should be n by nrhs. * This is larger than before, so we need to expose the new rows by resizing. */ if (m < n && b == rb_) { </ins> THTensor_(resize2d)(rb_, n, nrhs); <ins> } </ins> THTensor_(freeCopyTo)(ra__, ra_); THTensor_(freeCopyTo)(rb__, rb_);", "pos": ["b = (600,1) A = (600,2) X, = (b,A) X.size() #torch.Size([2, 1]) Y, = (b.cuda(),A.cuda()) Y.size() #torch.Size([600, 1])\nthat is interesting. says that the first rows returned ( here) include the solution of the least-squares problem. I've verified that the first 2 rows of X and Y are the same so you can probably use that for now while we figure out what the rest of the output is."], "neg": []}
{"query": "gtest_reports_dir=\"${TEST_DIR}/cpp\" pytest_reports_dir=\"${TEST_DIR}/python\" <ins> # This is needed to work around ROCm using old docker images until # the transition to new images is complete. # TODO: Remove once ROCm CI is using new images. if [[ $BUILD_ENVIRONMENT == py3.6-devtoolset7-rocmrpm-centos* ]]; then # This file is sourced multiple times, only install conda the first time. # We must install conda where we have write access. CONDA_DIR=\"$ROOT_DIR/conda\" if [[ ! -d $CONDA_DIR ]]; then ANACONDA_PYTHON_VERSION=3.6 BASE_URL=\"https://repo.anaconda.com/miniconda\" CONDA_FILE=\"Miniconda3-latest-Linux-x86_64.sh\" mkdir $CONDA_DIR pushd /tmp wget -q \"${BASE_URL}/${CONDA_FILE}\" chmod +x \"${CONDA_FILE}\" ./\"${CONDA_FILE}\" -b -f -p \"$CONDA_DIR\" popd export PATH=\"$CONDA_DIR/bin:$PATH\" # Ensure we run conda in a directory that jenkins has write access to pushd $CONDA_DIR # Track latest conda update conda update -n base conda # Install correct Python version conda install python=\"$ANACONDA_PYTHON_VERSION\" conda_install() { # Ensure that the install command don't upgrade/downgrade Python # This should be called as #   conda_install pkg1 pkg2 ... [-c channel] conda install -q -y python=\"$ANACONDA_PYTHON_VERSION\" $* } # Install PyTorch conda deps, as per https://github.com/pytorch/pytorch README conda_install numpy pyyaml mkl mkl-include setuptools cffi typing future six # TODO: This isn't working atm conda_install nnpack -c killeent # Install some other packages # Need networkx 2.0 because bellmand_ford was moved in 2.1 . Scikit-image by # defaults installs the most recent networkx version, so we install this lower # version explicitly before scikit-image pulls it in as a dependency pip install networkx==2.0 # TODO: Why is scipy pinned # numba & llvmlite is pinned because of https://github.com/numba/numba/issues/4368 # scikit-learn is pinned because of # https://github.com/scikit-learn/scikit-learn/issues/14485 (affects gcc 5.5 # only) pip install --progress-bar off pytest scipy==1.1.0 scikit-learn==0.20.3 scikit-image librosa>=0.6.2 psutil numba==0.46.0 llvmlite==0.30.0 # click - onnx # hypothesis - tests # jupyter - for tutorials pip install --progress-bar off click hypothesis jupyter protobuf tabulate virtualenv mock typing-extensions popd else export PATH=\"$CONDA_DIR/bin:$PATH\" fi fi </ins> # Figure out which Python to use PYTHON=\"$(which python)\" if [[ \"${BUILD_ENVIRONMENT}\" =~ py((2|3).?[0-9]?.?[0-9]?) ]]; then", "pos": ["See e.g.: cc\nCan we close it?\nThe build is passing now, although the tests might be failing (although it might have been an environment failure)."], "neg": []}
{"query": "compare_cpu_gpu(outputs_cpu, outputs_gpu) @unittest.skipIf(not TEST_CUDNN, \"needs cudnn\") <ins> @default_tensor_type(torch.FloatTensor)  # FIXME: just until torch.cuda.DoubleTensor.sum() implemented def test_RNN_cpu_vs_cudnn_no_dropout(self): self._test_RNN_cpu_vs_cudnn(0) @unittest.skipIf(not (TEST_CUDNN and TEST_CUDNN_VERSION >= 5103), \"needs cudnn >= 5.1\") @default_tensor_type(torch.FloatTensor)  # FIXME: just until torch.cuda.DoubleTensor.sum() implemented def test_RNN_cpu_vs_cudnn_with_dropout(self): # Because of dropout randomness, can only compare dropout=0 and dropout=1 self._test_RNN_cpu_vs_cudnn(1) @unittest.skipIf(not (TEST_CUDNN and TEST_CUDNN_VERSION >= 5103), \"needs cudnn >= 5.1\") </ins> def test_RNN_dropout(self): # checking the assumption that cuDNN sticks dropout in between # RNN layers", "pos": ["cuDNN backend for RNN stabilizes at large mem usage. The following code stabilizes for me at 14+GiB reported by nvidia-smi. The same RNN without cuDNN uses <2.5GiB.\nOops... any idea what's going on with the garbage collector in this case?\nare you running this on a P100 GPU?\nYes, I observe similar behavior on Titan X Pascal\nhmmm, do cudnn RNNs have weird memory-oversubscribing properties? If it wasn't for CuDNN RNN, then it's not unreasonable. we use a custom CUDA allocator that will oversubscribe if needed, but will free the memory blocks it's holding onto if it hits OOM / failed allocations.\nTried the same RNN in torch and am getting ~2.6GiB on nvidia-smi.\nWith cuDNN I'm getting a constant usage of 2241MB on my machine using PyTorch (M40).\nare you using cudnnv5, or v6?\nv5\nno, cudnn does not do memory oversubscribing, and already tried the same network in lua torch with cudnn, where it gets expected memory usage. Also, he is getting OOMs on slightly bigger networks, where apparently caching allocator should have tried to free previous blocks. Interesting that can not repro this.\nJust to make sure did you set when you ran? I left it as False in the repro.\nI can repro in the docker image built from the current version on P100, and I go OOM after 8 batches with the same image on 1080, which generally should not happen - if there is enough memory to go through 1 batch, memory usage should not increase afterwards.\nNo, I just copied your script and ran it. Weird thing is that it goes OOM after 18 batches once I enable . Without it memory remains stable.\nA lot of cudaMalloc allocations go not through THCudaMalloc, and never get freed:\nSo not going through THCudaMalloc is actually expected for storages allocated with THCStorage_(newWithSize), but what's not expected is to every time cudaMalloc bytes of rnn workspace, instead of reusing earlier allocations. (~784 MB allocations are for workspace, ~156MB allocations happen somewhere in criterion).\nMy guess is that the bug is at the Python level.", "Either: Something is holding on to a Tensor longer than it should Something is creating a circular reference which prevents quick clean-up\nThe issue is that in creates cyclic references. It's made worse by the fact that it holds on to more memory than it should (like the workspace). You can see this by adding a as the first line in your for loop. With the explicit call to the GC every iteration, the memory usage doesn't explode.\nhow can these cyclic references be fixed?\nNot sure yet, but I will fix it tomorrow with"], "neg": []}
{"query": ">>> m = nn.Dropout(p=0.2) >>> input = autograd.Variable(torch.randn(20, 16)) >>> output = m(input) <ins> .. _Improving neural networks by preventing co-adaptation of feature detectors: https://arxiv.org/abs/1207.0580 </ins> \"\"\" def __init__(self, p=0.5, inplace=False):", "pos": ["\"Randomly zeroes some of the elements of the input tensor. The elements to zero are randomized on every forward call.\" This is incorrect; the function also scales up by 1/(1-p), which the implementation correctly does."], "neg": []}
{"query": "torch.Size([3, 2, 2, 1]) >>> torch.unflatten(torch.randn(3, 4, 1), 1, (-1, 2)).shape torch.Size([3, 2, 2, 1]) <del> >>> torch.unflatten(torch.randn(5, 12, 3), -1, (2, 2, 3, 1, 1)).shape </del> <ins> >>> torch.unflatten(torch.randn(5, 12, 3), -2, (2, 2, 3, 1, 1)).shape </ins> torch.Size([5, 2, 2, 3, 1, 1, 3]) \"\"\".format( **common_args", "pos": ["Link In the third example the doc shows torch.unflatten((5, 12, 3), -1, (2, 2, 3, 1, 1)).shape torch.Size([5, 2, 2, 3, 1, 1, 3]) which should be incorrect, the correct code should be torch.unflatten((5, 12, 3), -2, (2, 2, 3, 1, 1)).shape torch.Size([5, 2, 2, 3, 1, 1, 3]) No response cc\nThank you for reporting. Next time you spot a problem like that please do not hesitate to propose a PR with the fix."], "neg": []}
{"query": "from torch.autograd import Variable <ins> def get_a_var(obj): if isinstance(obj, Variable): return obj if isinstance(obj, list) or isinstance(obj, tuple): results = map(get_a_var, obj) for result in results: if isinstance(result, Variable): return result if isinstance(obj, dict): results = map(get_a_var, obj.items()) for result in results: if isinstance(result, Variable): return result return None </ins> def parallel_apply(modules, inputs, kwargs_tup=None): assert len(modules) == len(inputs) if kwargs_tup:", "pos": ["is the culprit. Assumes that the first input is a variable or a sequence containing a variable in index 0 (recursively indexed). Workaround is to pass a dummy variable as first input, as suggested in\nCould you provide a short reproducer for testing? Think this shouldn't be too hard to fix.\nSure, here's a gist along with commands/outputs:\nfixed in master via"], "neg": []}
{"query": "scale = 1 exp_max = int(exp_max) <ins> prec = PRINT_OPTS.precision </ins> if int_mode: <del> if exp_max > 9: format = '{:11.4e}' sz = max(min_sz, 11) </del> <ins> if exp_max > prec + 1: format = '{{:11.{}e}}'.format(prec) sz = max(min_sz, 7 + prec) </ins> else: sz = max(min_sz, exp_max + 1) format = '{:' + str(sz) + '.0f}' else: <del> if exp_max - exp_min > 4: sz = 11 </del> <ins> if exp_max - exp_min > prec: sz = 7 + prec </ins> if abs(exp_max) > 99 or abs(exp_min) > 99: sz = sz + 1 sz = max(min_sz, sz) <del> format = '{:' + str(sz) + '.4e}' </del> <ins> format = '{{:{}.{}e}}'.format(sz, prec) </ins> else: <del> if exp_max > 5 or exp_max < 0: </del> <ins> if exp_max > prec + 1 or exp_max < 0: </ins> sz = max(min_sz, 7) scale = math.pow(10, exp_max-1) else:", "pos": ["It seems that the following link that can be reached from here (check the screenshot) is broken: Page: Broken link: Screenshot: ! cc\nthis should be transferred to , but I don't have permissions to do that. Fix is available at pytorch/hub."], "neg": []}
{"query": "for (int idx = 0; idx < size; idx++) { PyObject* obj = tuple ? PyTuple_GET_ITEM(arg, idx) : PyList_GET_ITEM(arg, idx); if (!THPVariable_Check(obj)) { <del> throw TypeError(\"expected Variable as element %d in argument %d, but got %s\", </del> <ins> throw TypeError(\"expected Tensor as element %d in argument %d, but got %s\", </ins> idx, i, Py_TYPE(args[i])->tp_name); } res[idx] = reinterpret_cast<THPVariable*>(obj)->cdata;", "pos": ["Considering the merge of and , and the \"removal\" of from the API, there are still some messages that reference . This might prove to be confusing for new users starting off from , as they probably might not have heard of . I did a code search to find out multiple instances of this in . PyTorch or Caffe2: PyTorch How you installed PyTorch (conda, pip, source): source (@ ) Build command you used (if compiling from source): OS: Ubuntu 16.04.2 PyTorch version: Python version: 3.5 CUDA/cuDNN version: 8.0.44 GPU models and configuration: GeForce 940M GCC version (if compiling from source): 5.4.0 CMake version: 3.5.1"], "neg": []}
{"query": "- bool training - double epsilon ]] <ins> [[ name: cudnn_version only_register: True ]] static PyObject * THCUDNN_cudnn_version(PyObject *self, PyObject *args) { return PyLong_FromLong(CUDNN_VERSION); } </ins>", "pos": ["We should check that the version from cudnn.h matches the version from the loaded shared library. Really weird things happen if they don't match."], "neg": []}
{"query": "{ transpose_m2 = (transpose_r == 'n' ? 't' : 'n'); m2_ = THTensor_(newContiguous)(m2); <ins> free_m2 = 1; </ins> } #pragma omp critical(blasgemm)", "pos": ["Here is an example of what seems to be a memory leak in pytorch, although it could be numpy? The leak was only occurring when using the or array unsqueezing methods in numpy. Also worth noting that the issue is resolved by removing the from and casting the torch tensor to type. An easy workaround is to reshape the array after turning it into a torch Tensor."], "neg": []}
{"query": "setattr(self, name, param) self._all_weights.append(param_names) <del> self._param_buf_size += sum(p.numel() for p in layer_params) </del> self.flatten_parameters() self.reset_parameters()", "pos": ["Since the new pytorch version, I have a warning when I that says: When I do as suggested, and then try to encode again, I get the following error though: Any idea?\ndo you have code the reproduces the issue?\nWhat was the fix? I still see the error on Pytorch Version: 0.2.03 AttributeError: 'LSTM' object has no attribute 'parambufsize'"], "neg": []}
{"query": "from copy import deepcopy from collections import OrderedDict from itertools import product <del> from operator import mul </del> <ins> from operator import mul, itemgetter </ins> from functools import reduce, wraps from torch.autograd.gradcheck import gradgradcheck, gradcheck from torch.autograd.function import once_differentiable", "pos": ["I am having problem in running the backward propagation with svd. In particular, it seems working when the first dimension of the matrix x (below) is greater than the second. But I get the following error if the first is smaller than the second. PyTorch version: 2018.04.20 Is debug build: No CUDA used to build PyTorch: 8.0.61 OS: Ubuntu 16.04.4 LTS GCC version: (Ubuntu 5.4.0-6ubuntu1~16.04.9) 5.4.0 CMake version: Could not collect Python version: 3.5 Is CUDA available: No CUDA runtime version: No CUDA GPU models and configuration: No CUDA Nvidia driver version: No CUDA cuDNN version: No CUDA Versions of relevant libraries: [pip3] numpy (1.13.3) [pip3] torch (2018.4.20) [pip3] torchvision (0.2.0) [conda] pytorch 0.3.1 py35cuda8.0.61cudnn7.0.52 pytorch [conda] pytorch-nightly 2018.04.20 py35cuda8.0.61cudnn7.1.21 pytorch [conda] torchvision 0.2.0 py35heaa392f_1 pytorch\nRepro'ed on 0.4 branch. works btw.\nWhat does \"Repro'ed\" means? Should I install from that branch? I need to compute the kernel of a matrix, therefore I need the last columns of the matrix V (which is why I have to pass ).\nIt's a shorthand for \"reproduced\", which means that the bug still hasn't been fixed. Sorry, we'll try to get to it soon."], "neg": []}
{"query": "output[index] = ScalarConvert<int, Dtype>::to(0); continue; } <ins> assert(cur_target  >= 0 && cur_target  < n_classes); </ins> Dtype weight = weights ? weights[cur_target] : ScalarConvert<int, Dtype>::to(1); output[index] = -weight * input[index][cur_target];", "pos": ["Claim: When I use an invalid label with F.crossentropy with reduce=False, I don't receive an error message. I do receive an error message when reduce=True. Code: Result: No error message received. Expected error message: Context: I was using , accidentally used an invalid label, and got extremely large losses and extremely small losses, and inconsistently so. Environment: OS: Ubuntu 16.04 on Docker and host PyTorch version: 0.3.0.post4 How you installed PyTorch: conda Python version: 3.6.3"], "neg": []}
{"query": "Writes all values from the tensor :attr:`src` into :attr:`self` at the indices specified in the :attr:`index` tensor. For each value in :attr:`src`, its output <del> index is specified by its index in :attr:`src` for dimension != :attr:`dim` and by the corresponding value in :attr:`index` for dimension = :attr:`dim`. </del> <ins> index is specified by its index in :attr:`src` for ``dimension != dim`` and by the corresponding value in :attr:`index` for ``dimension = dim``. </ins> For a 3-D tensor, :attr:`self` is updated as::", "pos": ["is not available in the docs page, but the doc string has been in . This was mentioned earlier in", "Oh, bilinear, I missed that last time I looked at it. I'll add a docstring.", "returns nothing. Might be useful to have some linter in the build process to find Python functions missing", "Pytorch 0.4.0 gives a segmentation fault with dilated 2D convolutions over a specific size. import torch import as nn model = nn.Conv2d(inchannels=64, outchannels=64, kernelsize=5, dilation=2) model(([1, 64, 1166, 1166])) # This runs model(([1, 64, 1167, 1167])) # This crashes PyTorch version: 0.4.0 Is debug build: No CUDA used to build PyTorch: 9.1.85 OS: CentOS Linux 7 (Core) GCC version: (GCC) 4.8.5 (Red Hat 4.8.5-28) CMake version: version 2.8.12.2 Python version: 3.6 Is CUDA available: Yes CUDA runtime version: 8.0.61 GPU models and configuration: GPU 0: GeForce GTX 1080 Ti GPU 1: TITAN X (Pascal) GPU 2: Quadro K2200 Nvidia driver version: 390.59 cuDNN version: Probably one of the following: /usr/local/cuda- /usr/local/cuda- /usr/local/cuda-7.5/lib64/libcudnnstatic.a /usr/local/cuda- /usr/local/cuda-8.0/lib64/libcudnn_static.a\ndo you think this is the same as the segfault you showed me a week ago on convolution?\nah yes. int overflow. I only fixed conv3d because I didn\u2019t want to hurt perf. I\u2019ll fix all of them.", "This should be an easy one.. The function is implemented, just no sign of it in the docs."], "neg": []}
{"query": "Args: input (Tensor): the input 2-D tensor <del> some (bool, optional): controls the number of singular values to be computed </del> <ins> some (bool, optional): controls the shape of returned `U` and `V` </ins> out (tuple, optional): the output tuple of tensors Example::", "pos": ["The documentation of the parameter to is one of the strangest I've seen... what does (the default) do???\nHa it is somewhat weird indeed. When , the returned U and V matrices are square orthonormal matrices. I will send out a PR to clarify this."], "neg": []}