""" Full definition of a GPT Language Model, all of it in this single file. References: 1) the official GPT-2 TensorFlow implementation released by OpenAI: https://github.com/openai/gpt-2/blob/master/src/model.py 2) huggingface/transformers PyTorch implementation: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py """ import math import inspect from dataclasses import dataclass import torch import torch.nn as nn from torch.nn import RMSNorm from torch.nn import functional as F from position import RotaryPosEncoding, PolarPosEncoding, AlibiPositionalBias, build_sinusoidal_pos_emb class LayerNorm(nn.Module): """ LayerNorm but with an optional bias. PyTorch doesn't support simply bias=False """ def __init__(self, ndim, bias): super().__init__() self.weight = nn.Parameter(torch.ones(ndim)) self.bias = nn.Parameter(torch.zeros(ndim)) if bias else None def forward(self, input): return F.layer_norm(input, self.weight.shape, self.weight, self.bias, 1e-5) class CausalSelfAttention(nn.Module): def __init__(self, config): super().__init__() assert config.n_embd % config.n_head == 0 # key, query, value projections for all heads, but in a batch self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias) # rotary positional encoding self.head_size = config.n_embd // config.n_head rot_embd = self.head_size if config.pos_type == 'rope': self.rotary_pos = RotaryPosEncoding(rot_embd, config.base_freq, pretrained_seq_len=config.block_size, extended_seq_len=config.block_size, rotate_fraction=config.rotate_fraction) elif config.pos_type == 'pope': self.rotary_pos = PolarPosEncoding(rot_embd, config.n_head, config.block_size, config.base_freq, rotate_fraction=config.rotate_fraction, use_theta_bias=config.use_theta_bias, thetab_init=config.thetab_init) elif config.pos_type == 'alibi': alibi_heads = config.alibi_n_heads if not (1 <= alibi_heads <= config.n_head): raise ValueError(f"alibi_n_heads must be in [1, n_head], got {alibi_heads} for n_head={config.n_head}") self.alibi_pos = AlibiPositionalBias(alibi_heads=alibi_heads, total_heads=config.n_head) # output projection self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias) # regularization self.attn_dropout = nn.Dropout(config.dropout) self.resid_dropout = nn.Dropout(config.dropout) self.n_head = config.n_head self.n_embd = config.n_embd self.dropout = config.dropout self.pos_type = config.pos_type # storage for analysis/visualization self.store_qk = False # flag to control storage of q/k self.stored_queries = None self.stored_keys = None self.store_attn_scores = False # flag to control storage of attn_scores self.stored_attn_scores = None # flash attention make GPU go brrrrr but support is only in PyTorch >= 2.0 self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention') # to use custom complex_flash_attention self.complex_flash = config.complex_flash self.complex_attention = None self.get_complex_dtype = None if self.complex_flash: from complex_flash_attention import complex_attention, get_dtype self.complex_attention = complex_attention self.get_complex_dtype = get_dtype if not self.flash: print("WARNING: using slow attention. Flash Attention requires PyTorch >= 2.0") # causal mask to ensure that attention is only applied to the left in the input sequence self.register_buffer("bias", torch.tril(torch.ones(config.block_size, config.block_size, dtype=torch.bool)) .view(1, 1, config.block_size, config.block_size)) def get_causal_mask(self, t: int) -> torch.Tensor: bias_len = self.bias.shape[-1] if bias_len < t: self.bias = torch.tril(torch.ones(t, t, device=self.c_proj.weight.device, dtype=torch.bool))[None, None] return self.bias[:, :, :t, :t] def enable_qk_storage(self): self.store_qk = True def disable_qk_storage(self): self.store_qk = False self.stored_queries = None self.stored_keys = None def get_stored_qk(self): return self.stored_queries, self.stored_keys def enable_attn_scores_storage(self): self.store_attn_scores = True def get_attn_scores(self): return self.stored_attn_scores def split_into_heads(self, x, B, T): if isinstance(x, torch.Tensor): return x.view(B, T, self.n_head, self.head_size).transpose(1, 2) # (B, nh, T, hs) elif isinstance(x, tuple): return tuple(map(lambda xi: xi.view(B, T, self.n_head, self.head_size).transpose(1, 2), x)) def forward(self, x): B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd) att_scale = math.pow(self.head_size, 0.25) att = None # default attention scores to None # calculate query, key, values for all heads in batch and move head forward to be the batch dim q, k, v = self.c_attn(x).split(self.n_embd, dim=2) # slice into heads before rotating (heads share the same frequency) q = self.split_into_heads(q, B, T) # (B, nh, T, hs) k = self.split_into_heads(k, B, T) # (B, nh, T, hs) v = self.split_into_heads(v, B, T) # (B, nh, T, hs) # apply rotary positional encoding if self.pos_type in ['rope', 'pope']: q, k = self.rotary_pos(q, k) # apply scaling to q/k q = tuple(map(lambda qi: qi / att_scale, q)) if isinstance(q, tuple) else q / att_scale k = tuple(map(lambda ki: ki / att_scale, k)) if isinstance(k, tuple) else k / att_scale # Store query and key vectors if requested if self.store_qk: if isinstance(q, tuple): # Detach and clone to avoid keeping gradients and allow independent access q_real, q_imag = q[0].detach().clone()[..., None], q[1].detach().clone()[..., None] k_real, k_imag = k[0].detach().clone()[..., None], k[1].detach().clone()[..., None] self.stored_queries = torch.view_as_complex(torch.cat((q_real, q_imag), dim=-1)) self.stored_keys = torch.view_as_complex(torch.cat((k_real, k_imag), dim=-1)) else: self.stored_queries = q.detach().clone() self.stored_keys = k.detach().clone() # causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T) if self.flash and self.pos_type not in ['pope', 'alibi'] and not self.store_attn_scores: # efficient attention using Flash Attention CUDA kernels y = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=self.dropout if self.training else 0, is_causal=True, scale=1.0) # custom complex flash attention implementation in triton elif self.pos_type == 'pope': if self.complex_flash: # TODO: the flash attention kernel requires contiguous tensors for now, but none of them are. Given this, # we might as well cast it to the AMP dtype to minimize bandwidth and saved memory. Still call contiguous() # just to be safe. However, it should already be contiguous after casting, so this should be free. amp_type = self.get_complex_dtype() # negative sign for q_imag due to conjugation y = self.complex_attention.apply(q[0].to(amp_type).contiguous(), k[0].to(amp_type).contiguous(), -q[1].to(amp_type).contiguous(), k[1].to(amp_type).contiguous(), v.to(amp_type).contiguous(), True, 1.0) # manual attention implementation else: # NOTE: real component has the cosine term # att = (q.conj() @ k.transpose(-2, -1)).real with torch.amp.autocast("cuda", enabled=False): q = tuple(map(lambda qi: qi.to(torch.float32), q)) k = tuple(map(lambda ki: ki.to(torch.float32), k)) att = q[0] @ k[0].transpose(-2, -1) + q[1] @ k[1].transpose(-2, -1) att = att.masked_fill(self.get_causal_mask(T) == 0, float('-inf')) att = F.softmax(att, dim=-1) att = self.attn_dropout(att) y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs) else: # Manual attention path used for ALiBi and as a fallback when Flash is disabled. with torch.amp.autocast("cuda", enabled=False): q = q.to(torch.float32) k = k.to(torch.float32) att = q @ k.transpose(-2, -1) if self.pos_type == 'alibi': att = att + self.alibi_pos(T, T, device=att.device).to(att.dtype) att = att.masked_fill(self.get_causal_mask(T) == 0, float('-inf')) att = F.softmax(att, dim=-1) att = self.attn_dropout(att) y = att @ v # store attention scores if self.store_attn_scores: self.stored_attn_scores = att.detach().clone() y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side # output projection y = self.resid_dropout(self.c_proj(y)) return y class MLP(nn.Module): def __init__(self, config): super().__init__() self.c_fc = nn.Linear(config.n_embd, 4 * config.n_embd, bias=config.bias) self.gelu = nn.GELU() self.c_proj = nn.Linear(4 * config.n_embd, config.n_embd, bias=config.bias) self.dropout = nn.Dropout(config.dropout) def forward(self, x): x = self.c_fc(x) x = self.gelu(x) x = self.c_proj(x) x = self.dropout(x) return x class Block(nn.Module): def __init__(self, config): super().__init__() if config.norm_type == 'layernorm': self.attn_norm = LayerNorm(config.n_embd, bias=config.bias) self.ff_norm = LayerNorm(config.n_embd, bias=config.bias) elif config.norm_type == 'rmsnorm': self.attn_norm = RMSNorm(config.n_embd) self.ff_norm = RMSNorm(config.n_embd) else: raise ValueError(f"unknown norm type {config.norm_type}") self.attn = CausalSelfAttention(config) self.mlp = MLP(config) def forward(self, x): x = x + self.attn(self.attn_norm(x)) x = x + self.mlp(self.ff_norm(x)) return x @dataclass class GPTConfig: dataset: str = 'shakespeare_char' # dataset name block_size: int = 1024 vocab_size: int = 50304 # GPT-2 vocab_size of 50257, padded up to nearest multiple of 64 for efficiency n_layer: int = 12 n_head: int = 12 n_embd: int = 768 norm_type: str = 'rmsnorm' # 'layernorm' or 'rmsnorm' pos_type: str = 'pope' # 'learnable', 'sinusoidal', 'rope', 'pope', 'alibi' alibi_n_heads: int = 12 # number of heads using ALiBi in [1, n_head] use_theta_bias: bool = True # whether to use theta bias in PoPE base_freq: int = 10000 # base frequency for rotary positional encoding thetab_init: str = 'two_pi' # 'two_pi' or 'zero', for PoPE_1b init of theta_bias rotate_fraction: float = 1.0 # fraction of the embedding dimension to rotate dropout: float = 0.0 bias: bool = True # True: bias in Linears and LayerNorms, like GPT-2. False: a bit better and faster complex_flash: bool = False # use custom complex_flash_attention class GPT(nn.Module): def __init__(self, config): super().__init__() assert config.vocab_size is not None assert config.block_size is not None self.config = config self.transformer = nn.ModuleDict(dict( wte = nn.Embedding(config.vocab_size, config.n_embd),)) # learned absolute positional embeddings if config.pos_type == 'learnable': self.transformer.update(dict( wpe = nn.Embedding(config.block_size, config.n_embd),)) # fixed sinusoidal positional embeddings elif config.pos_type == 'sinusoidal': self.register_buffer('sinusoidal_wpe', build_sinusoidal_pos_emb(config.block_size, config.n_embd)) self.transformer.update(dict( drop = nn.Dropout(config.dropout), h = nn.ModuleList([Block(config) for _ in range(config.n_layer)]),)) if config.norm_type == 'layernorm': self.transformer.update(dict( norm_f = LayerNorm(config.n_embd, bias=config.bias),)) elif config.norm_type == 'rmsnorm': self.transformer.update(dict( norm_f = RMSNorm(config.n_embd), )) self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False) # with weight tying when using torch.compile() some warnings get generated: # "UserWarning: functional_call was passed multiple values for tied weights. # This behavior is deprecated and will be an error in future versions" # not 100% sure what this is, so far seems to be harmless. TODO investigate self.transformer.wte.weight = self.lm_head.weight # https://paperswithcode.com/method/weight-tying # init all weights self.apply(self._init_weights) # apply special scaled init to the residual projections, per GPT-2 paper for pn, p in self.named_parameters(): if pn.endswith('c_proj.weight'): torch.nn.init.normal_(p, mean=0.0, std=0.02/math.sqrt(2 * config.n_layer)) # report number of parameters print("number of parameters: %.2fM" % (self.get_num_params()/1e6,)) def get_num_params(self, non_embedding=True): """ Return the number of parameters in the model. For non-embedding count (default), the position embeddings get subtracted. The token embeddings would too, except due to the parameter sharing these params are actually used as weights in the final layer, so we include them. """ n_params = sum(p.numel() for p in self.parameters()) if non_embedding and self.config.pos_type == 'learnable': n_params -= self.transformer.wpe.weight.numel() return n_params def _init_weights(self, module): if isinstance(module, nn.Linear): torch.nn.init.normal_(module.weight, mean=0.0, std=0.02) if module.bias is not None: torch.nn.init.zeros_(module.bias) elif isinstance(module, nn.Embedding): torch.nn.init.normal_(module.weight, mean=0.0, std=0.02) def forward(self, idx, targets=None, pad_token_id=None, target_mask=None): device = idx.device b, t = idx.size() # assert t <= self.config.block_size, f"Cannot forward sequence of length {t}, block size is only {self.config.block_size}" pos = torch.arange(0, t, dtype=torch.long, device=device) # shape (t) # forward the GPT model itself tok_emb = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd) if self.config.pos_type == 'learnable': pos_emb = self.transformer.wpe(pos) # position embeddings of shape (t, n_embd) x = self.transformer.drop(tok_emb + pos_emb) elif self.config.pos_type == 'sinusoidal': pos_emb = self.sinusoidal_wpe[:t].to(tok_emb.dtype) x = self.transformer.drop(tok_emb + pos_emb) else: x = self.transformer.drop(tok_emb) for block in self.transformer.h: x = block(x) x = self.transformer.norm_f(x) if targets is not None: # if we are given some desired targets also calculate the loss logits = self.lm_head(x) # apply loss on selected targets only (used by indirect_idx) if target_mask is not None: logits = logits.view(-1, logits.size(-1)) targets = targets.view(-1) target_mask = target_mask.view(-1) logits = logits[target_mask] targets = targets[target_mask] # mask out loss for padded tokens if pad_token_id is not None: logits = logits.view(-1, logits.size(-1)) targets = targets.view(-1) mask = targets != pad_token_id logits = logits[mask] targets = targets[mask] # compute CE loss loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1) else: # inference-time mini-optimization: only forward the lm_head on the very last position logits = self.lm_head(x[:, [-1], :]) # note: using list [-1] to preserve the time dim loss = None return logits, loss def crop_block_size(self, block_size): # model surgery to decrease the block size if necessary # e.g. we may load the GPT2 pretrained model checkpoint (block size 1024) # but want to use a smaller block size for some smaller, simpler model assert block_size <= self.config.block_size self.config.block_size = block_size if self.config.pos_type == 'learnable': self.transformer.wpe.weight = nn.Parameter(self.transformer.wpe.weight[:block_size]) elif self.config.pos_type == 'sinusoidal': self.sinusoidal_wpe = self.sinusoidal_wpe[:block_size] for block in self.transformer.h: if hasattr(block.attn, 'bias'): block.attn.bias = block.attn.bias[:,:,:block_size,:block_size] @classmethod def from_pretrained(cls, model_type, override_args=None): assert model_type in {'gpt2', 'gpt2-medium', 'gpt2-large', 'gpt2-xl'} override_args = override_args or {} # default to empty dict # only dropout can be overridden see more notes below assert all(k == 'dropout' for k in override_args) from transformers import GPT2LMHeadModel print("loading weights from pretrained gpt: %s" % model_type) # n_layer, n_head and n_embd are determined from model_type config_args = { 'gpt2': dict(n_layer=12, n_head=12, n_embd=768), # 124M params 'gpt2-medium': dict(n_layer=24, n_head=16, n_embd=1024), # 350M params 'gpt2-large': dict(n_layer=36, n_head=20, n_embd=1280), # 774M params 'gpt2-xl': dict(n_layer=48, n_head=25, n_embd=1600), # 1558M params }[model_type] print("forcing vocab_size=50257, block_size=1024, bias=True") config_args['vocab_size'] = 50257 # always 50257 for GPT model checkpoints config_args['block_size'] = 1024 # always 1024 for GPT model checkpoints config_args['bias'] = True # always True for GPT model checkpoints # we can override the dropout rate, if desired if 'dropout' in override_args: print(f"overriding dropout rate to {override_args['dropout']}") config_args['dropout'] = override_args['dropout'] # create a from-scratch initialized minGPT model config = GPTConfig(**config_args) model = GPT(config) sd = model.state_dict() sd_keys = sd.keys() sd_keys = [k for k in sd_keys if not k.endswith('.attn.bias')] # discard this mask / buffer, not a param # init a huggingface/transformers model model_hf = GPT2LMHeadModel.from_pretrained(model_type) sd_hf = model_hf.state_dict() # copy while ensuring all of the parameters are aligned and match in names and shapes sd_keys_hf = sd_hf.keys() sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.masked_bias')] # ignore these, just a buffer sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.bias')] # same, just the mask (buffer) transposed = ['attn.c_attn.weight', 'attn.c_proj.weight', 'mlp.c_fc.weight', 'mlp.c_proj.weight'] # basically the openai checkpoints use a "Conv1D" module, but we only want to use a vanilla Linear # this means that we have to transpose these weights when we import them assert len(sd_keys_hf) == len(sd_keys), f"mismatched keys: {len(sd_keys_hf)} != {len(sd_keys)}" for k in sd_keys_hf: if any(k.endswith(w) for w in transposed): # special treatment for the Conv1D weights we need to transpose assert sd_hf[k].shape[::-1] == sd[k].shape with torch.no_grad(): sd[k].copy_(sd_hf[k].t()) else: # vanilla copy over the other parameters assert sd_hf[k].shape == sd[k].shape with torch.no_grad(): sd[k].copy_(sd_hf[k]) return model def configure_optimizers(self, weight_decay, learning_rate, betas, pos_type, device_type): # start with all of the candidate parameters param_dict = {pn: p for pn, p in self.named_parameters()} # filter out those that do not require grad param_dict = {pn: p for pn, p in param_dict.items() if p.requires_grad} # create optim groups. Any parameters that is 2D will be weight decayed, otherwise no. # i.e. all weight tensors in matmuls + embeddings decay, all biases and layernorms don't. decay_params = [p for n, p in param_dict.items() if p.dim() >= 2] nodecay_params = [p for n, p in param_dict.items() if p.dim() < 2] optim_groups = [ {'params': decay_params, 'weight_decay': weight_decay}, {'params': nodecay_params, 'weight_decay': 0.0} ] num_decay_params = sum(p.numel() for p in decay_params) num_nodecay_params = sum(p.numel() for p in nodecay_params) print(f"num decayed parameter tensors: {len(decay_params)}, with {num_decay_params:,} parameters") print(f"num non-decayed parameter tensors: {len(nodecay_params)}, with {num_nodecay_params:,} parameters") # Create AdamW optimizer and use the fused version if it is available fused_available = 'fused' in inspect.signature(torch.optim.AdamW).parameters use_fused = fused_available and device_type == 'cuda' # NOTE: AdamW fused implementation is not supported for complex-valued tensors extra_args = dict(fused=True) if use_fused else dict() optimizer = torch.optim.AdamW(optim_groups, lr=learning_rate, betas=betas, **extra_args) print(f"using fused AdamW: {use_fused}") return optimizer def estimate_mfu(self, fwdbwd_per_iter, dt): """ estimate model flops utilization (MFU) in units of A100 bfloat16 peak FLOPS """ # first estimate the number of flops we do per iteration. # see PaLM paper Appendix B as ref: https://arxiv.org/abs/2204.02311 N = self.get_num_params() cfg = self.config L, H, Q, T = cfg.n_layer, cfg.n_head, cfg.n_embd//cfg.n_head, cfg.block_size flops_per_token = 6*N + 12*L*H*Q*T flops_per_fwdbwd = flops_per_token * T flops_per_iter = flops_per_fwdbwd * fwdbwd_per_iter # express our flops throughput as ratio of A100 bfloat16 peak flops flops_achieved = flops_per_iter * (1.0/dt) # per second flops_promised = 312e12 # A100 GPU bfloat16 peak flops is 312 TFLOPS mfu = flops_achieved / flops_promised return mfu @torch.no_grad() def generate(self, idx, max_new_tokens, temperature=1.0, top_k=None): """ Take a conditioning sequence of indices idx (LongTensor of shape (b,t)) and complete the sequence max_new_tokens times, feeding the predictions back into the model each time. Most likely you'll want to make sure to be in model.eval() mode of operation for this. """ for _ in range(max_new_tokens): # if the sequence context is growing too long we must crop it at block_size idx_cond = idx if idx.size(1) <= self.config.block_size else idx[:, -self.config.block_size:] # forward the model to get the logits for the index in the sequence logits, _ = self(idx_cond) # pluck the logits at the final step and scale by desired temperature logits = logits[:, -1, :] / temperature # optionally crop the logits to only the top k options if top_k is not None: v, _ = torch.topk(logits, min(top_k, logits.size(-1))) logits[logits < v[:, [-1]]] = -float('Inf') # apply softmax to convert logits to (normalized) probabilities probs = F.softmax(logits, dim=-1) # sample from the distribution idx_next = torch.multinomial(probs, num_samples=1) # append sampled index to the running sequence and continue idx = torch.cat((idx, idx_next), dim=1) return idx