import copy import torch import torch.nn as nn from torch.nn import CrossEntropyLoss from torch.nn import functional as F from typing import Optional, Tuple, Union from transformers import T5Config, T5Tokenizer, T5ForConditionalGeneration from transformers.models.t5.modeling_t5 import ( T5LayerNorm, T5LayerFF, T5Attention, T5LayerSelfAttention, T5LayerCrossAttention, T5Block, T5Stack, ) from transformers.modeling_outputs import BaseModelOutput, Seq2SeqLMOutput class BottleneckCrossAttentionGate(nn.Module): def __init__(self, config): super().__init__() self.gate = nn.Linear(2 * config.d_model, config.d_model, bias=False) self.act = nn.Sigmoid() def forward(self, query_states, latents): latents = latents.unsqueeze(1).expand(query_states.shape) query_latents = torch.cat([query_states, latents], dim=-1) return 2 * self.act(self.gate(query_latents)) class BottleneckT5Attention(T5Attention): def __init__(self, config: T5Config, has_relative_attention_bias=False): super(T5Attention, self).__init__() self.is_decoder = config.is_decoder self.has_relative_attention_bias = has_relative_attention_bias self.relative_attention_num_buckets = config.relative_attention_num_buckets self.relative_attention_max_distance = config.relative_attention_max_distance self.d_model = config.d_model self.key_value_proj_dim = config.d_kv self.n_heads = config.num_heads self.dropout = config.dropout_rate self.inner_dim = self.n_heads * self.key_value_proj_dim # Mesh TensorFlow initialization to avoid scaling before softmax self.v = nn.Linear(self.d_model, self.inner_dim, bias=False) self.o = nn.Linear(self.inner_dim, self.d_model, bias=False) if self.has_relative_attention_bias: self.relative_attention_bias = nn.Embedding(self.relative_attention_num_buckets, self.n_heads) self.pruned_heads = set() self.gradient_checkpointing = False def prune_heads(self, heads): if len(heads) == 0: return heads, index = find_pruneable_heads_and_indices( heads, self.n_heads, self.key_value_proj_dim, self.pruned_heads ) # Prune linear layers self.v = prune_linear_layer(self.v, index) self.o = prune_linear_layer(self.o, index, dim=1) # Update hyper params self.n_heads = self.n_heads - len(heads) self.inner_dim = self.key_value_proj_dim * self.n_heads self.pruned_heads = self.pruned_heads.union(heads) def forward( self, hidden_states, mask=None, key_value_states=None, position_bias=None, past_key_value=None, layer_head_mask=None, query_length=None, use_cache=False, output_attentions=False, ): """ Self-attention (if key_value_states is None) or attention over source sentence (provided by key_value_states). """ # Input is (batch_size, seq_length, dim) # Mask is (batch_size, key_length) (non-causal) or (batch_size, key_length, key_length) # past_key_value[0] is (batch_size, n_heads, q_len - 1, dim_per_head) batch_size, seq_length = hidden_states.shape[:2] real_seq_length = seq_length if past_key_value is not None: assert ( len(past_key_value) == 2 ), f"past_key_value should have 2 past states: keys and values. Got { len(past_key_value)} past states" real_seq_length += past_key_value[0].shape[2] if query_length is None else query_length key_length = real_seq_length if key_value_states is None else key_value_states.shape[1] def shape(states): """projection""" return states.view(batch_size, -1, self.n_heads, self.key_value_proj_dim).transpose(1, 2) def unshape(states): """reshape""" return states.transpose(1, 2).contiguous().view(batch_size, -1, self.inner_dim) def project(hidden_states, proj_layer, key_value_states, past_key_value): """projects hidden states correctly to key/query states""" if key_value_states is None: # self-attn # (batch_size, n_heads, seq_length, dim_per_head) hidden_states = shape(proj_layer(hidden_states)) elif past_key_value is None: # cross-attn # (batch_size, n_heads, seq_length, dim_per_head) hidden_states = shape(proj_layer(key_value_states)) if past_key_value is not None: if key_value_states is None: # self-attn # (batch_size, n_heads, key_length, dim_per_head) hidden_states = torch.cat([past_key_value, hidden_states], dim=2) else: # cross-attn hidden_states = past_key_value return hidden_states # key/value states key_states = torch.zeros((batch_size, self.n_heads, seq_length, key_length), device=hidden_states.device) value_states = project( hidden_states, self.v, key_value_states, past_key_value[1] if past_key_value is not None else None ) # compute scores scores = torch.ones((batch_size, self.n_heads, seq_length, key_length), device=hidden_states.device) if position_bias is None: if not self.has_relative_attention_bias: position_bias = torch.zeros( (1, self.n_heads, real_seq_length, key_length), device=scores.device, dtype=scores.dtype ) if self.gradient_checkpointing and self.training: position_bias.requires_grad = True else: position_bias = self.compute_bias(real_seq_length, key_length, device=scores.device) # if key and values are already calculated # we want only the last query position bias if past_key_value is not None: position_bias = position_bias[:, :, -hidden_states.size(1) :, :] if mask is not None: position_bias = position_bias + mask # (batch_size, n_heads, seq_length, key_length) if self.pruned_heads: mask = torch.ones(position_bias.shape[1]) mask[list(self.pruned_heads)] = 0 position_bias_masked = position_bias[:, mask.bool()] else: position_bias_masked = position_bias scores += position_bias_masked attn_weights = nn.functional.softmax(scores.float(), dim=-1).type_as( scores ) # (batch_size, n_heads, seq_length, key_length) attn_weights = nn.functional.dropout( attn_weights, p=self.dropout, training=self.training ) # (batch_size, n_heads, seq_length, key_length) # Mask heads if we want to if layer_head_mask is not None: attn_weights = attn_weights * layer_head_mask attn_output = unshape(torch.matmul(attn_weights, value_states)) # (batch_size, seq_length, dim) attn_output = self.o(attn_output) present_key_value_state = (key_states, value_states) if (self.is_decoder and use_cache) else None outputs = (attn_output,) + (present_key_value_state,) + (position_bias,) if output_attentions: outputs = outputs + (attn_weights,) return outputs class BottleneckT5LayerCrossAttention(T5LayerCrossAttention): def __init__(self, config): super(T5LayerCrossAttention, self).__init__() self.EncDecAttention = BottleneckT5Attention(config, has_relative_attention_bias=False) self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon) self.gate = BottleneckCrossAttentionGate(config) self.dropout = nn.Dropout(config.dropout_rate) def forward( self, hidden_states, key_value_states, attention_mask=None, position_bias=None, layer_head_mask=None, past_key_value=None, use_cache=False, query_length=None, output_attentions=False, ): normed_hidden_states = self.layer_norm(hidden_states) attention_output = self.EncDecAttention( normed_hidden_states, mask=attention_mask, key_value_states=key_value_states, position_bias=position_bias, layer_head_mask=layer_head_mask, past_key_value=past_key_value, use_cache=use_cache, query_length=query_length, output_attentions=output_attentions, ) latents = key_value_states[:, 0] layer_output = hidden_states + self.dropout(self.gate(normed_hidden_states, latents) * attention_output[0]) outputs = (layer_output,) + attention_output[1:] # add attentions if we output them return outputs class BottleneckT5Block(T5Block): def __init__(self, config, has_relative_attention_bias=False): super(T5Block, self).__init__() self.is_decoder = config.is_decoder self.layer = nn.ModuleList() self.layer.append(T5LayerSelfAttention(config, has_relative_attention_bias=has_relative_attention_bias)) if self.is_decoder: self.layer.append(BottleneckT5LayerCrossAttention(config)) self.layer.append(T5LayerFF(config)) class BottleneckT5Stack(T5Stack): def __init__(self, config, embed_tokens=None): super(T5Stack, self).__init__(config) self.embed_tokens = embed_tokens self.is_decoder = config.is_decoder self.block = nn.ModuleList( [BottleneckT5Block(config, has_relative_attention_bias=bool(i == 0)) for i in range(config.num_layers)] ) self.final_layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon) self.dropout = nn.Dropout(config.dropout_rate) # Initialize weights and apply final processing self.post_init() # Model parallel self.model_parallel = False self.device_map = None self.gradient_checkpointing = False class BottleneckT5LMWithPerturb(T5ForConditionalGeneration): def __init__(self, config: T5Config): super(T5ForConditionalGeneration, self).__init__(config) self.model_dim = config.d_model self.shared = nn.Embedding(config.vocab_size, config.d_model) encoder_config = copy.deepcopy(config) encoder_config.is_decoder = False encoder_config.use_cache = False encoder_config.is_encoder_decoder = False self.encoder = T5Stack(encoder_config, self.shared) # New in Contra: MHA bottleneck block self.num_heads = 8 self.bottleneck = nn.MultiheadAttention(config.d_model, num_heads=self.num_heads, dropout=config.dropout_rate, bias=False, batch_first=True) self.bottleneck_scale = nn.Parameter(torch.ones(1)) self.dec_emb = nn.Embedding(config.vocab_size, config.d_model) decoder_config = copy.deepcopy(config) decoder_config.is_decoder = True decoder_config.is_encoder_decoder = False decoder_config.num_layers = config.num_decoder_layers self.decoder = BottleneckT5Stack(decoder_config, self.dec_emb) self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False) # Initialize weights and apply final processing self.post_init() # Model parallel self.model_parallel = False self.device_map = None def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.FloatTensor] = None, decoder_input_ids: Optional[torch.LongTensor] = None, decoder_attention_mask: Optional[torch.BoolTensor] = None, head_mask: Optional[torch.FloatTensor] = None, decoder_head_mask: Optional[torch.FloatTensor] = None, cross_attn_head_mask: Optional[torch.Tensor] = None, encoder_outputs: Optional[Tuple[Tuple[torch.Tensor]]] = None, past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, decoder_inputs_embeds: Optional[torch.FloatTensor] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, perturb_vector: Optional[torch.FloatTensor] = None, encode_only: Optional[bool] = None, ) -> Union[Tuple[torch.FloatTensor], Seq2SeqLMOutput]: use_cache = use_cache if use_cache is not None else self.config.use_cache return_dict = return_dict if return_dict is not None else self.config.use_return_dict # FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask if head_mask is not None and decoder_head_mask is None: if self.config.num_layers == self.config.num_decoder_layers: warnings.warn(__HEAD_MASK_WARNING_MSG, FutureWarning) decoder_head_mask = head_mask # Encode if needed (training, first prediction pass) if encoder_outputs is None: # Convert encoder inputs in embeddings if needed encoder_outputs = self.encoder( input_ids=input_ids, attention_mask=attention_mask, inputs_embeds=inputs_embeds, head_mask=head_mask, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) elif return_dict and not isinstance(encoder_outputs, BaseModelOutput): encoder_outputs = BaseModelOutput( last_hidden_state=encoder_outputs[0], hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None, attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None, ) hidden_states = encoder_outputs[0] # MHA across token embeddings + embedding normalization + broadcast hidden_states = hidden_states.repeat( attention_mask.shape[0] // hidden_states.shape[0], 1, 1) # during contrastive search, attn mask can have higher batch size than hidden_state mask_expanded = attention_mask.float().unsqueeze(-1).expand(hidden_states.shape) mean_pooled_embedding = torch.sum(hidden_states * mask_expanded, 1) / torch.clamp(mask_expanded.sum(1), min=1e-9) unscaled_latent, attn_weights = self.bottleneck(mean_pooled_embedding.unsqueeze(1), hidden_states, hidden_states, need_weights=False, # torch MHA attn_mask has opposite signs to HF T5 masks... sigh attn_mask=attention_mask.float().unsqueeze(1).repeat_interleave(self.num_heads, dim=0)) latent = self.bottleneck_scale * F.normalize(unscaled_latent, p=2, dim=2) if encode_only: return latent.squeeze(1) hidden_states = latent.expand(hidden_states.shape) if hasattr(self, 'perturb_vector'): hidden_states = self.bottleneck_scale * F.normalize(hidden_states + self.perturb_vector, p=2, dim=2) if self.model_parallel: torch.cuda.set_device(self.decoder.first_device) if labels is not None and decoder_input_ids is None and decoder_inputs_embeds is None: # get decoder inputs from shifting lm labels to the right decoder_input_ids = self._shift_right(labels) # Set device for model parallelism if self.model_parallel: torch.cuda.set_device(self.decoder.first_device) hidden_states = hidden_states.to(self.decoder.first_device) if decoder_input_ids is not None: decoder_input_ids = decoder_input_ids.to(self.decoder.first_device) if attention_mask is not None: attention_mask = attention_mask.to(self.decoder.first_device) if decoder_attention_mask is not None: decoder_attention_mask = decoder_attention_mask.to(self.decoder.first_device) # Decode decoder_outputs = self.decoder( input_ids=decoder_input_ids, attention_mask=decoder_attention_mask, inputs_embeds=decoder_inputs_embeds, past_key_values=past_key_values, encoder_hidden_states=hidden_states, encoder_attention_mask=attention_mask, head_mask=decoder_head_mask, cross_attn_head_mask=cross_attn_head_mask, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) sequence_output = decoder_outputs[0] # Set device for model parallelism if self.model_parallel: torch.cuda.set_device(self.encoder.first_device) self.lm_head = self.lm_head.to(self.encoder.first_device) sequence_output = sequence_output.to(self.lm_head.weight.device) if self.config.tie_word_embeddings: # Rescale output before projecting on vocab # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586 sequence_output = sequence_output * (self.model_dim**-0.5) lm_logits = self.lm_head(sequence_output) loss = None if labels is not None: loss_fct = CrossEntropyLoss(ignore_index=-100) loss = loss_fct(lm_logits.view(-1, lm_logits.size(-1)), labels.view(-1)) # TODO(thom): Add z_loss https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L666 if not return_dict: output = (lm_logits,) + decoder_outputs[1:] + encoder_outputs return ((loss,) + output) if loss is not None else output return Seq2SeqLMOutput( loss=loss, logits=lm_logits, past_key_values=decoder_outputs.past_key_values, decoder_hidden_states=decoder_outputs.hidden_states, decoder_attentions=decoder_outputs.attentions, cross_attentions=decoder_outputs.cross_attentions, encoder_last_hidden_state=encoder_outputs.last_hidden_state, encoder_hidden_states=encoder_outputs.hidden_states, encoder_attentions=encoder_outputs.attentions, )