from typing import Optional, Tuple, Union import torch import torch.nn as nn from transformers.activations import ACT2FN from transformers.models.gpt2.modeling_gpt2 import BaseModelOutputWithPastAndCrossAttentions, GPT2PreTrainedModel from transformers.pytorch_utils import Conv1D class GPT2MLP(nn.Module): def __init__(self, intermediate_size, config): super().__init__() embed_dim = config.hidden_size self.c_fc = Conv1D(intermediate_size, embed_dim) self.c_proj = Conv1D(embed_dim, intermediate_size) self.act = ACT2FN[config.activation_function] self.dropout = nn.Dropout(config.resid_pdrop) def forward(self, hidden_states: Optional[Tuple[torch.FloatTensor]]) -> torch.FloatTensor: hidden_states = self.c_fc(hidden_states) hidden_states = self.act(hidden_states) hidden_states = self.c_proj(hidden_states) return hidden_states # The reason Why we don't import GPT2Attention from transformers directly is that: # 1. The tracer will not work correctly when we feed meta_args and concrete_args at same time, # so we have to build the customized GPT2Attention class and remove the conditional branch manually. # 2. The order of split and view op has been changed in the customized GPT2Attention class, the new # order is same as megatron-lm gpt model. class GPT2Attention(nn.Module): def __init__(self, config, layer_idx=None): super().__init__() max_positions = config.max_position_embeddings self.register_buffer( "bias", torch.tril(torch.ones((max_positions, max_positions), dtype=torch.uint8)).view(1, 1, max_positions, max_positions), ) self.register_buffer("masked_bias", torch.tensor(-1e4)) self.embed_dim = config.hidden_size self.num_heads = config.num_attention_heads self.head_dim = self.embed_dim // self.num_heads self.split_size = self.embed_dim self.scale_attn_weights = config.scale_attn_weights # Layer-wise attention scaling, reordering, and upcasting self.scale_attn_by_inverse_layer_idx = config.scale_attn_by_inverse_layer_idx self.layer_idx = layer_idx self.c_attn = Conv1D(3 * self.embed_dim, self.embed_dim) self.c_proj = Conv1D(self.embed_dim, self.embed_dim) self.attn_dropout = nn.Dropout(config.attn_pdrop) self.resid_dropout = nn.Dropout(config.resid_pdrop) self.pruned_heads = set() def _attn(self, query, key, value, attention_mask=None, head_mask=None): attn_weights = torch.matmul(query, key.transpose(-1, -2)) if self.scale_attn_weights: attn_weights = attn_weights / (value.size(-1)**0.5) # Layer-wise attention scaling if self.scale_attn_by_inverse_layer_idx: attn_weights = attn_weights / float(self.layer_idx + 1) # if only "normal" attention layer implements causal mask query_length, key_length = query.size(-2), key.size(-2) causal_mask = self.bias[:, :, key_length - query_length:key_length, :key_length].to(torch.bool) attn_weights = torch.where(causal_mask, attn_weights, self.masked_bias.to(attn_weights.dtype)) if attention_mask is not None: # Apply the attention mask attn_weights = attn_weights + attention_mask attn_weights = nn.functional.softmax(attn_weights, dim=-1) attn_weights = attn_weights.type(value.dtype) # Mask heads if we want to if head_mask is not None: attn_weights = attn_weights * head_mask attn_output = torch.matmul(attn_weights, value) return attn_output, attn_weights def _split_heads(self, tensor, num_heads, attn_head_size): new_shape = tensor.size()[:-1] + (num_heads, attn_head_size) tensor = tensor.view(new_shape) return tensor.permute(0, 2, 1, 3) # (batch, head, seq_length, head_features) def _merge_heads(self, tensor, num_heads, attn_head_size): tensor = tensor.permute(0, 2, 1, 3).contiguous() new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,) return tensor.view(new_shape) def forward( self, hidden_states: Optional[Tuple[torch.FloatTensor]], attention_mask: Optional[torch.FloatTensor] = None, head_mask: Optional[torch.FloatTensor] = None, ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]: qkv = self.c_attn(hidden_states) query, key, value = self._split_heads(qkv, self.num_heads, 3 * self.head_dim).split(self.head_dim, dim=3) present = (key, value) attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask) attn_output = self._merge_heads(attn_output, self.num_heads, self.head_dim) attn_output = self.c_proj(attn_output) return attn_output class GPT2Block(nn.Module): def __init__(self, config, layer_idx=None): super().__init__() hidden_size = config.hidden_size inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon) self.attn = GPT2Attention(config, layer_idx=layer_idx) self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon) self.mlp = GPT2MLP(inner_dim, config) def forward( self, hidden_states: Optional[Tuple[torch.FloatTensor]], attention_mask: Optional[torch.FloatTensor] = None, head_mask: Optional[torch.FloatTensor] = None, ) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]: residual = hidden_states hidden_states = self.ln_1(hidden_states) attn_outputs = self.attn( hidden_states, attention_mask=attention_mask, head_mask=head_mask, ) # residual connection hidden_states = attn_outputs + residual residual = hidden_states hidden_states = self.ln_2(hidden_states) feed_forward_hidden_states = self.mlp(hidden_states) # residual connection hidden_states = residual + feed_forward_hidden_states return hidden_states class GPT2Model(GPT2PreTrainedModel): def __init__(self, config): super().__init__(config) self.embed_dim = config.hidden_size self.wte = nn.Embedding(config.vocab_size, self.embed_dim) self.wpe = nn.Embedding(config.max_position_embeddings, self.embed_dim) self.drop = nn.Dropout(config.embd_pdrop) self.h = nn.ModuleList([GPT2Block(config, layer_idx=i) for i in range(config.num_hidden_layers)]) self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon) # Initialize weights and apply final processing self.post_init() def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.FloatTensor] = None, head_mask: Optional[torch.FloatTensor] = None, ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]: input_shape = input_ids.size() input_ids = input_ids.view(-1, input_shape[-1]) batch_size = input_ids.shape[0] device = input_ids.device past_length = 0 past_key_values = tuple([None] * len(self.h)) position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device) position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1]) # GPT2Attention mask. attention_mask = attention_mask.view(batch_size, -1) attention_mask = attention_mask[:, None, None, :] attention_mask = attention_mask.to(dtype=self.dtype) # fp16 compatibility attention_mask = (1.0 - attention_mask) * -10000.0 encoder_attention_mask = None # Prepare head mask if needed # 1.0 in head_mask indicate we keep the head # attention_probs has shape bsz x n_heads x N x N # head_mask has shape n_layer x batch x n_heads x N x N head_mask = self.get_head_mask(head_mask, self.config.n_layer) inputs_embeds = self.wte(input_ids) position_embeds = self.wpe(position_ids) hidden_states = inputs_embeds + position_embeds output_shape = input_shape + (hidden_states.size(-1),) for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)): outputs = block(hidden_states, attention_mask=attention_mask, head_mask=head_mask[i]) hidden_states = outputs hidden_states = self.ln_f(hidden_states) hidden_states = hidden_states.view(output_shape) return hidden_states class GPT2LMHeadModel(GPT2PreTrainedModel): def __init__(self, config): super().__init__(config) self.transformer = GPT2Model(config) self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False) # Initialize weights and apply final processing self.post_init() def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.FloatTensor] = None, ): transformer_outputs = self.transformer( input_ids=input_ids, attention_mask=attention_mask, ) lm_logits = self.lm_head(transformer_outputs) return lm_logits class GPTLMLoss(nn.Module): def __init__(self): super().__init__() self.loss_fn = nn.CrossEntropyLoss() def forward(self, logits, labels): shift_logits = logits[..., :-1, :].contiguous() shift_labels = labels[..., 1:].contiguous() # Flatten the tokens return self.loss_fn(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))