lingconv_t5.py

import warnings
import copy
from typing import Optional, Tuple, Union

import torch
from torch.nn import CrossEntropyLoss

from transformers.modeling_outputs import (
    BaseModelOutput,
    Seq2SeqLMOutput,
    BaseModelOutputWithPastAndCrossAttentions,
)
from transformers.models.t5.modeling_t5 import T5Stack, T5ForConditionalGeneration, __HEAD_MASK_WARNING_MSG
from transformers import T5Config

from transformers.cache_utils import Cache, DynamicCache, EncoderDecoderCache
from transformers.utils import (
    is_torchdynamo_compiling,
)
import logging

logger = logging.getLogger(__name__)

class LingConvT5Stack(T5Stack):
    def __init__(self, config: T5Config, embed_tokens=None):
        super().__init__(config, embed_tokens)

        # Add new attributes for ling injection
        self.ling_injection_layer = getattr(config, 'ling_injection_layer', -1)
        self.ling_injection_type = getattr(config, 'ling_injection_type', 'none') # 'none', 'first', 'all'

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        inputs_embeds=None,
        head_mask=None,
        cross_attn_head_mask=None,
        past_key_values=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
        cache_position=None,
        ling_embed=None,
    ):
        # Model parallel
        if self.model_parallel:
            torch.cuda.set_device(self.first_device)
            self.embed_tokens = self.embed_tokens.to(self.first_device)
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if input_ids is not None and inputs_embeds is not None:
            err_msg_prefix = "decoder_" if self.is_decoder else ""
            raise ValueError(
                f"You cannot specify both {err_msg_prefix}input_ids and {err_msg_prefix}inputs_embeds at the same time"
            )
        elif input_ids is not None:
            input_shape = input_ids.size()
            input_ids = input_ids.view(-1, input_shape[-1])
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            err_msg_prefix = "decoder_" if self.is_decoder else ""
            raise ValueError(f"You have to specify either {err_msg_prefix}input_ids or {err_msg_prefix}inputs_embeds")

        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                )
                use_cache = False

        if inputs_embeds is None:
            if self.embed_tokens is None:
                raise ValueError("You have to initialize the model with valid token embeddings")
            inputs_embeds = self.embed_tokens(input_ids)

        batch_size, seq_length = input_shape

        if use_cache is True:
            if not self.is_decoder:
                raise ValueError(f"`use_cache` can only be set to `True` if {self} is used as a decoder")

        # initialize past_key_values
        return_legacy_cache = False
        return_self_attention_cache = False
        if self.is_decoder and (use_cache or past_key_values is not None):
            if isinstance(past_key_values, Cache) and not isinstance(past_key_values, EncoderDecoderCache):
                return_self_attention_cache = True
                past_key_values = EncoderDecoderCache(past_key_values, DynamicCache())
            elif not isinstance(past_key_values, EncoderDecoderCache):
                return_legacy_cache = True
                logger.warning_once(
                    "Passing a tuple of `past_key_values` is deprecated and will be removed in Transformers v4.48.0. "
                    "You should pass an instance of `EncoderDecoderCache` instead, e.g. "
                    "`past_key_values=EncoderDecoderCache.from_legacy_cache(past_key_values)`."
                )
                past_key_values = EncoderDecoderCache.from_legacy_cache(past_key_values)
            elif past_key_values is None:
                past_key_values = EncoderDecoderCache(DynamicCache(), DynamicCache())
        elif not self.is_decoder:
            # do not pass cache object down the line for encoder stack
            # it messes indexing later in decoder-stack because cache object is modified in-place
            past_key_values = None

        past_key_values_length = past_key_values.get_seq_length() if past_key_values is not None else 0
        if cache_position is None:
            cache_position = torch.arange(
                past_key_values_length, past_key_values_length + seq_length, device=inputs_embeds.device
            )

        if attention_mask is None and not is_torchdynamo_compiling():
            # required mask seq length can be calculated via length of past cache
            mask_seq_length = past_key_values_length + seq_length
            attention_mask = torch.ones(batch_size, mask_seq_length, device=inputs_embeds.device)

        if self.config.is_decoder:
            causal_mask = self._update_causal_mask(
                attention_mask,
                inputs_embeds,
                cache_position,
                past_key_values.self_attention_cache if past_key_values is not None else None,
                output_attentions,
            )
        elif attention_mask is not None:
            causal_mask = attention_mask[:, None, None, :]
            causal_mask = causal_mask.to(dtype=inputs_embeds.dtype)
            causal_mask = (1.0 - causal_mask) * torch.finfo(inputs_embeds.dtype).min
        else:
            causal_mask = None

        # If a 2D or 3D attention mask is provided for the cross-attention
        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
        if self.is_decoder and encoder_hidden_states is not None:
            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
            if encoder_attention_mask is None:
                encoder_attention_mask = torch.ones(
                    encoder_hidden_shape, device=inputs_embeds.device, dtype=torch.long
                )
            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
        else:
            encoder_extended_attention_mask = None

        # Prepare head mask if needed
        head_mask = self.get_head_mask(head_mask, self.config.num_layers)
        cross_attn_head_mask = self.get_head_mask(cross_attn_head_mask, self.config.num_layers)
        all_hidden_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None
        all_cross_attentions = () if (output_attentions and self.is_decoder) else None
        position_bias = None
        encoder_decoder_position_bias = None

        hidden_states = self.dropout(inputs_embeds)

        for i, layer_module in enumerate(self.block):
            layer_head_mask = head_mask[i]
            cross_attn_layer_head_mask = cross_attn_head_mask[i]
            # Model parallel
            if self.model_parallel:
                torch.cuda.set_device(hidden_states.device)
                # Ensure that attention_mask is always on the same device as hidden_states
                if causal_mask is not None:
                    causal_mask = causal_mask.to(hidden_states.device)
                if position_bias is not None:
                    position_bias = position_bias.to(hidden_states.device)
                if encoder_hidden_states is not None:
                    encoder_hidden_states = encoder_hidden_states.to(hidden_states.device)
                if encoder_extended_attention_mask is not None:
                    encoder_extended_attention_mask = encoder_extended_attention_mask.to(hidden_states.device)
                if encoder_decoder_position_bias is not None:
                    encoder_decoder_position_bias = encoder_decoder_position_bias.to(hidden_states.device)
                if layer_head_mask is not None:
                    layer_head_mask = layer_head_mask.to(hidden_states.device)
                if cross_attn_layer_head_mask is not None:
                    cross_attn_layer_head_mask = cross_attn_layer_head_mask.to(hidden_states.device)
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    layer_module.forward,
                    hidden_states,
                    causal_mask,
                    position_bias,
                    encoder_hidden_states,
                    encoder_extended_attention_mask,
                    encoder_decoder_position_bias,
                    layer_head_mask,
                    cross_attn_layer_head_mask,
                    None,  # past_key_value is always None with gradient checkpointing
                    use_cache,
                    output_attentions,
                    return_dict,
                    cache_position,
                )
            else:
                layer_outputs = layer_module(
                    hidden_states,
                    attention_mask=causal_mask,
                    position_bias=position_bias,
                    encoder_hidden_states=encoder_hidden_states,
                    encoder_attention_mask=encoder_extended_attention_mask,
                    encoder_decoder_position_bias=encoder_decoder_position_bias,
                    layer_head_mask=layer_head_mask,
                    cross_attn_layer_head_mask=cross_attn_layer_head_mask,
                    past_key_value=past_key_values,
                    use_cache=use_cache,
                    output_attentions=output_attentions,
                    return_dict=return_dict,
                    cache_position=cache_position,
                )

            # layer_outputs is a tuple with:
            # hidden-states, key-value-states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights)
            if use_cache is False:
                layer_outputs = layer_outputs[:1] + (None,) + layer_outputs[1:]

            hidden_states, next_decoder_cache = layer_outputs[:2]

            # Add linguistic embedding injection after specified layer
            if (self.is_decoder and 
                self.ling_injection_layer == i and 
                ling_embed is not None and 
                self.ling_injection_type != 'none'):
                
                hidden_states = hidden_states + ling_embed

            # We share the position biases between the layers - the first layer store them
            # layer_outputs = hidden-states, key-value-states (self-attention position bias), (self-attention weights),
            # (cross-attention position bias), (cross-attention weights)
            position_bias = layer_outputs[2]
            if self.is_decoder and encoder_hidden_states is not None:
                encoder_decoder_position_bias = layer_outputs[4 if output_attentions else 3]

            if output_attentions:
                all_attentions = all_attentions + (layer_outputs[3],)
                if self.is_decoder:
                    all_cross_attentions = all_cross_attentions + (layer_outputs[5],)

            # Model Parallel: If it's the last layer for that device, put things on the next device
            if self.model_parallel:
                for k, v in self.device_map.items():
                    if i == v[-1] and "cuda:" + str(k) != self.last_device:
                        hidden_states = hidden_states.to("cuda:" + str(k + 1))

        hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.dropout(hidden_states)

        # Add last layer
        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        next_cache = next_decoder_cache if use_cache else None
        if return_self_attention_cache:
            next_cache = past_key_values.self_attention_cache
        if return_legacy_cache:
            next_cache = past_key_values.to_legacy_cache()

        if not return_dict:
            return tuple(
                v
                for v in [
                    hidden_states,
                    next_cache,
                    all_hidden_states,
                    all_attentions,
                    all_cross_attentions,
                ]
                if v is not None
            )
        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_attentions,
            cross_attentions=all_cross_attentions,
        )

class LingConvT5ForConditionalGeneration(T5ForConditionalGeneration):
    def __init__(self, config):
        super().__init__(config)
        # Replace default decoder with our custom decoder
        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 = LingConvT5Stack(decoder_config, embed_tokens=self.shared)

    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,
        cache_position: Optional[torch.LongTensor] = None,
        ling_embed: Optional[torch.FloatTensor] = None,
    ) -> Union[Tuple[torch.FloatTensor], Seq2SeqLMOutput]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[-100, 0, ...,
            config.vocab_size - 1]`. All labels set to `-100` are ignored (masked), the loss is only computed for
            labels in `[0, ..., config.vocab_size]`

        Returns:

        Examples:

        ```python
        >>> from transformers import AutoTokenizer, T5ForConditionalGeneration

        >>> tokenizer = AutoTokenizer.from_pretrained("google-t5/t5-small")
        >>> model = T5ForConditionalGeneration.from_pretrained("google-t5/t5-small")

        >>> # training
        >>> input_ids = tokenizer("The <extra_id_0> walks in <extra_id_1> park", return_tensors="pt").input_ids
        >>> labels = tokenizer("<extra_id_0> cute dog <extra_id_1> the <extra_id_2>", return_tensors="pt").input_ids
        >>> outputs = model(input_ids=input_ids, labels=labels)
        >>> loss = outputs.loss
        >>> logits = outputs.logits

        >>> # inference
        >>> input_ids = tokenizer(
        ...     "summarize: studies have shown that owning a dog is good for you", return_tensors="pt"
        ... ).input_ids  # Batch size 1
        >>> outputs = model.generate(input_ids)
        >>> print(tokenizer.decode(outputs[0], skip_special_tokens=True))
        >>> # studies have shown that owning a dog is good for you.
        ```"""
        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]

        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,
            cache_position=cache_position,
            ling_embed=ling_embed,
        )

        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)
            # move labels to correct device to enable PP
            labels = labels.to(lm_logits.device)
            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,
        )