talk-llama : sync llama.cpp

examples/talk-llama/llama-sampling.cpp

@@ -113,7 +113,7 @@ static void llama_sampler_softmax_impl(llama_token_data_array * cur_p) {
 }
 
 static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k) {
-    // TODO: move bucket sort to separate function so that top_p/tail_free/typical/softmax first is equally fast
+    // TODO: move bucket sort to separate function so that top_p/typical/softmax first is equally fast
     // if (k >= (int32_t)cur_p->size) {
     //     return;
     // }
@@ -733,101 +733,6 @@ struct llama_sampler * llama_sampler_init_min_p(float p, size_t min_keep) {
     };
 }
 
-// tail-free
-
-struct llama_sampler_tail_free {
-    const float  z;
-    const size_t min_keep;
-};
-
-static const char * llama_sampler_tail_free_name(const struct llama_sampler * /*smpl*/) {
-    return "tail-free";
-}
-
-static void llama_sampler_tail_free_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
-    const auto * ctx = (llama_sampler_tail_free *) smpl->ctx;
-
-    if (ctx->z >= 1.0f || cur_p->size <= 2) {
-        return;
-    }
-
-    llama_sampler_softmax_impl(cur_p);
-
-    // Compute the first and second derivatives
-    std::vector<float> first_derivatives(cur_p->size - 1);
-    std::vector<float> second_derivatives(cur_p->size - 2);
-
-    for (size_t i = 0; i < first_derivatives.size(); ++i) {
-        first_derivatives[i] = cur_p->data[i].p - cur_p->data[i + 1].p;
-    }
-    for (size_t i = 0; i < second_derivatives.size(); ++i) {
-        second_derivatives[i] = first_derivatives[i] - first_derivatives[i + 1];
-    }
-
-    // Calculate absolute value of second derivatives
-    for (size_t i = 0; i < second_derivatives.size(); ++i) {
-        second_derivatives[i] = std::abs(second_derivatives[i]);
-    }
-
-    // Normalize the second derivatives
-    {
-        const float second_derivatives_sum = std::accumulate(second_derivatives.begin(), second_derivatives.end(), 0.0f);
-
-        if (second_derivatives_sum > 1e-6f) {
-            for (float & value : second_derivatives) {
-                value /= second_derivatives_sum;
-            }
-        } else {
-            for (float & value : second_derivatives) {
-                value = 1.0f / second_derivatives.size();
-            }
-        }
-    }
-
-    float cum_sum = 0.0f;
-    size_t last_idx = cur_p->size;
-    for (size_t i = 0; i < second_derivatives.size(); ++i) {
-        cum_sum += second_derivatives[i];
-
-        // Check if the running sum is greater than z or if we have kept at least min_keep tokens
-        if (cum_sum > ctx->z && i >= ctx->min_keep) {
-            last_idx = i;
-            break;
-        }
-    }
-
-    // Resize the output vector to keep only the tokens above the tail location
-    cur_p->size = last_idx;
-}
-
-static struct llama_sampler * llama_sampler_tail_free_clone(const struct llama_sampler * smpl) {
-    const auto * ctx = (const llama_sampler_tail_free *) smpl->ctx;
-    return llama_sampler_init_tail_free(ctx->z, ctx->min_keep);
-}
-
-static void llama_sampler_tail_free_free(struct llama_sampler * smpl) {
-    delete (llama_sampler_tail_free *) smpl->ctx;
-}
-
-static struct llama_sampler_i llama_sampler_tail_free_i = {
-    /* .name   = */ llama_sampler_tail_free_name,
-    /* .accept = */ nullptr,
-    /* .apply  = */ llama_sampler_tail_free_apply,
-    /* .reset  = */ nullptr,
-    /* .clone  = */ llama_sampler_tail_free_clone,
-    /* .free   = */ llama_sampler_tail_free_free,
-};
-
-struct llama_sampler * llama_sampler_init_tail_free(float z, size_t min_keep) {
-    return new llama_sampler {
-        /* .iface = */ &llama_sampler_tail_free_i,
-        /* .ctx   = */ new llama_sampler_tail_free {
-            /* .z        = */ z,
-            /*. min_keep = */ min_keep,
-        },
-    };
-}
-
 // typical
 
 struct llama_sampler_typical {
@@ -1971,8 +1876,11 @@ static void llama_sampler_dry_reset(struct llama_sampler * smpl) {
 static struct llama_sampler * llama_sampler_dry_clone(const struct llama_sampler * smpl) {
     const auto * ctx = (llama_sampler_dry *) smpl->ctx;
 
-    // nullptr is passed as vocab because it is only needed for raw sequence breaker processing, which we have already done and will be copying
-    auto * result = llama_sampler_init_dry(nullptr, ctx->dry_multiplier, ctx->dry_base, ctx->dry_allowed_length, ctx->dry_penalty_last_n, NULL, 0);
+    llama_vocab dummy_vocab;
+
+    // dummy vocab is passed because it is only needed for raw sequence breaker processing, which we have already done and will simply be copying
+    auto * result = llama_sampler_init_dry_impl(dummy_vocab, ctx->total_context_size, ctx->dry_multiplier, ctx->dry_base, ctx->dry_allowed_length, ctx->dry_penalty_last_n, NULL, 0);
+
     // Copy the state, including the processed breakers
     {
         auto * result_ctx = (llama_sampler_dry *) result->ctx;

examples/talk-llama/llama.h

@@ -2,6 +2,7 @@
 #define LLAMA_H
 
 #include "ggml.h"
+#include "ggml-cpu.h"
 #include "ggml-backend.h"
 
 #include <stddef.h>
@@ -205,7 +206,7 @@ extern "C" {
     enum llama_split_mode {
         LLAMA_SPLIT_MODE_NONE  = 0, // single GPU
         LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs
-        LLAMA_SPLIT_MODE_ROW   = 2, // split rows across GPUs
+        LLAMA_SPLIT_MODE_ROW   = 2, // split layers and KV across GPUs, use tensor parallelism if supported
     };
 
     // TODO: simplify (https://github.com/ggerganov/llama.cpp/pull/9294#pullrequestreview-2286561979)
@@ -274,10 +275,7 @@ extern "C" {
         int32_t n_gpu_layers; // number of layers to store in VRAM
         enum llama_split_mode split_mode; // how to split the model across multiple GPUs
 
-        // main_gpu interpretation depends on split_mode:
-        // LLAMA_SPLIT_MODE_NONE: the GPU that is used for the entire model
-        // LLAMA_SPLIT_MODE_ROW: the GPU that is used for small tensors and intermediate results
-        // LLAMA_SPLIT_MODE_LAYER: ignored
+        // the GPU that is used for the entire model when split_mode is LLAMA_SPLIT_MODE_NONE
         int32_t main_gpu;
 
         // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
@@ -799,7 +797,7 @@ extern "C" {
     // Processes a batch of tokens with the ecoder part of the encoder-decoder model.
     // Stores the encoder output internally for later use by the decoder cross-attention layers.
     //   0 - success
-    // < 0 - error
+    // < 0 - error. the KV cache state is restored to the state before this call
     LLAMA_API int32_t llama_encode(
             struct llama_context * ctx,
               struct llama_batch   batch);
@@ -807,7 +805,7 @@ extern "C" {
     // Positive return values does not mean a fatal error, but rather a warning.
     //   0 - success
     //   1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
-    // < 0 - error
+    // < 0 - error. the KV cache state is restored to the state before this call
     LLAMA_API int32_t llama_decode(
             struct llama_context * ctx,
               struct llama_batch   batch);
@@ -1087,9 +1085,6 @@ extern "C" {
     /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
     LLAMA_API struct llama_sampler * llama_sampler_init_min_p      (float   p, size_t min_keep);
 
-    /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
-    LLAMA_API struct llama_sampler * llama_sampler_init_tail_free  (float   z, size_t min_keep);
-
     /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
     LLAMA_API struct llama_sampler * llama_sampler_init_typical    (float   p, size_t min_keep);