model.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-

from typing import List, Optional, Tuple, Union

import torch
import torch.nn as nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss

from transformers.modeling_outputs import (
    CausalLMOutputWithCrossAttentions,
    BaseModelOutputWithPastAndCrossAttentions,
)
from transformers import GPT2PreTrainedModel
from transformers.pytorch_utils import (
    find_pruneable_heads_and_indices,
    prune_conv1d_layer,
)
from transformers.activations import ACT2FN


class GPT2Attention(nn.Module):
    def __init__(self, config, block_size, is_cross_attention=False, 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.block_size = block_size
        if self.head_dim * self.num_heads != self.embed_dim:
            raise ValueError(
                f"`embed_dim` must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
                f" {self.num_heads})."
            )

        self.scale_attn_weights = config.scale_attn_weights
        self.is_cross_attention = is_cross_attention

        # 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.reorder_and_upcast_attn = config.reorder_and_upcast_attn

        if self.is_cross_attention:
            self.c_attn = nn.Linear(self.embed_dim, 2 * self.embed_dim)
            self.q_attn = nn.Linear(self.embed_dim, self.embed_dim)
        else:
            self.c_attn_q = nn.Linear(self.embed_dim, self.embed_dim)
            self.c_attn_k = nn.Linear(self.embed_dim, self.embed_dim)
            self.c_attn_v = nn.Linear(self.embed_dim, self.embed_dim)
        self.c_proj = nn.Linear(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 prune_heads(self, heads):
        if len(heads) == 0:
            return
        heads, index = find_pruneable_heads_and_indices(
            heads, self.num_heads, self.head_dim, self.pruned_heads
        )
        index_attn = torch.cat(
            [index, index + self.split_size, index + (2 * self.split_size)]
        )

        # Prune conv1d layers
        self.c_attn = prune_conv1d_layer(self.c_attn, index_attn, dim=1)
        self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)

        # Update hyper params
        self.split_size = (self.split_size // self.num_heads) * (
            self.num_heads - len(heads)
        )
        self.num_heads = self.num_heads - len(heads)
        self.pruned_heads = self.pruned_heads.union(heads)

    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 / (self.head_dim**0.5)

        # Layer-wise attention scaling
        if self.scale_attn_by_inverse_layer_idx:
            attn_weights = attn_weights / float(self.layer_idx + 1)

        if not self.is_cross_attention:
            # if only "normal" attention layer implements causal mask
            query_length, key_length = self.block_size, self.block_size
            causal_mask = self.bias[
                :, :, key_length - query_length : key_length, :key_length
            ].bool()

            attn_weights = torch.where(
                causal_mask,
                attn_weights,
                torch.tensor(-torch.inf, device=torch.device("cuda")),
            )

        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)

        # Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op otherwise
        attn_weights = attn_weights
        attn_weights = self.attn_dropout(attn_weights)

        # 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 _upcast_and_reordered_attn(
        self, query, key, value, attention_mask=None, head_mask=None
    ):
        # Use `torch.baddbmm` (a bit more efficient w/ alpha param for scaling -- from Megatron-LM)
        bsz, num_heads, q_seq_len, dk = query.size()
        _, _, k_seq_len, _ = key.size()

        # Preallocate attn_weights for `baddbmm`
        attn_weights = torch.empty(
            bsz * num_heads,
            q_seq_len,
            k_seq_len,
            dtype=torch.float32,
            device=torch.device("cuda"),
        )

        # Compute Scale Factor
        scale_factor = 1.0
        if self.scale_attn_weights:
            scale_factor /= float(value.size(-1)) ** 0.5

        if self.scale_attn_by_inverse_layer_idx:
            scale_factor /= float(self.layer_idx + 1)

        # Upcast (turn off autocast) and reorder (Scale K by 1 / root(dk))
        with autocast(enabled=False):
            q, k = query.reshape(-1, q_seq_len, dk), key.transpose(-1, -2).reshape(
                -1, dk, k_seq_len
            )
            attn_weights = torch.baddbmm(
                attn_weights, q.float(), k.float(), beta=0, alpha=scale_factor
            )
            attn_weights = attn_weights.reshape(bsz, num_heads, q_seq_len, k_seq_len)

        if not self.is_cross_attention:
            # 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
            ].bool()
            mask_value = torch.finfo(attn_weights.dtype).min
            # Need to be a tensor, otherwise we get error: `RuntimeError: expected scalar type float but found double`.
            # Need to be on the same device, otherwise `RuntimeError: ..., x and y to be on the same device`
            mask_value = torch.tensor(
                mask_value, dtype=attn_weights.dtype, device=torch.device("cuda")
            )
            attn_weights = torch.where(causal_mask, attn_weights, mask_value)

        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)

        # Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op if otherwise
        if attn_weights.dtype != torch.float32:
            raise RuntimeError(
                "Error with upcasting, attn_weights does not have dtype torch.float32"
            )
        attn_weights = attn_weights.type(value.dtype)
        attn_weights = self.attn_dropout(attn_weights)

        # 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):
        """
        Splits hidden_size dim into attn_head_size and num_heads
        """
        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):
        """
        Merges attn_head_size dim and num_attn_heads dim into hidden_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]],
        layer_past: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
    ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]:
        query = self.c_attn_q(hidden_states)
        key = self.c_attn_k(hidden_states)
        value = self.c_attn_v(hidden_states)

        query = self._split_heads(query, self.num_heads, self.head_dim)
        key = self._split_heads(key, self.num_heads, self.head_dim)
        value = self._split_heads(value, self.num_heads, self.head_dim)

        if layer_past is not None:
            past_key, past_value = layer_past
            key = torch.cat((past_key, key), dim=-2)
            value = torch.cat((past_value, value), dim=-2)

        present = (key, value)

        if self.reorder_and_upcast_attn:
            attn_output, attn_weights = self._upcast_and_reordered_attn(
                query, key, value, attention_mask, head_mask
            )
        else:
            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)
        attn_output = self.resid_dropout(attn_output)

        outputs = (attn_output, present)

        return outputs  # a, present, (attentions)


class GPT2MLP(nn.Module):
    def __init__(self, intermediate_size, config):
        super().__init__()
        embed_dim = config.hidden_size
        self.c_fc = nn.Linear(embed_dim, intermediate_size)
        self.c_proj = nn.Linear(intermediate_size, embed_dim)
        self.act = nn.GELU()
        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)
        hidden_states = self.dropout(hidden_states)
        return hidden_states


class GPT2Block(nn.Module):
    def __init__(self, config, block_size, 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, block_size, layer_idx=layer_idx)
        self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)

        if config.add_cross_attention:
            self.crossattention = GPT2Attention(
                config, block_size, is_cross_attention=True, layer_idx=layer_idx
            )
            self.ln_cross_attn = nn.LayerNorm(
                hidden_size, eps=config.layer_norm_epsilon
            )

        self.mlp = GPT2MLP(inner_dim, config)

    def forward(
        self,
        hidden_states: Optional[Tuple[torch.FloatTensor]],
        layer_past: Optional[Tuple[torch.Tensor]] = None,
        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,
            layer_past=layer_past,
            attention_mask=attention_mask,
            head_mask=head_mask,
        )
        attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)
        outputs = attn_outputs[1:]
        # residual connection
        hidden_states = attn_output + 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

        outputs = (hidden_states,) + outputs

        return outputs  # hidden_states, present, (attentions, cross_attentions)


class ModifiedGPT2Model(GPT2PreTrainedModel):
    _keys_to_ignore_on_load_missing = ["attn.masked_bias"]

    def __init__(self, config, block_size):
        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, block_size, layer_idx=i)
                for i in range(config.num_hidden_layers)
            ]
        )
        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
        self.block_size = block_size

        # Model parallel
        self.model_parallel = False
        self.device_map = None
        self.gradient_checkpointing = False

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
    ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
        input_ids = input_ids.view(-1, self.block_size)

        position_ids = torch.arange(
            0, self.block_size, dtype=torch.long, device=self.device
        )
        position_ids = position_ids.unsqueeze(0).view(-1, self.block_size)

        # GPT2Attention mask.
        if attention_mask is not None:
            attention_mask = attention_mask[:, None, None, :]
            attention_mask = (
                1.0 - attention_mask
            ) * -1000000.0  # -inf is replaced by -1000000.0

        # 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)
        head_mask = [None] * self.config.n_layer

        inputs_embeds = self.wte(input_ids)
        position_embeds = self.wpe(position_ids)
        hidden_states = inputs_embeds + position_embeds

        hidden_states = self.drop(hidden_states)

        presents = None
        all_self_attentions = None
        all_cross_attentions = None
        all_hidden_states = None
        for i, block in enumerate(self.h):
            outputs = block(
                hidden_states,
                layer_past=None,
                attention_mask=attention_mask,
                head_mask=head_mask[i],
            )
            hidden_states = outputs[0]

        hidden_states = self.ln_f(hidden_states)

        hidden_states = hidden_states.view(-1, self.block_size, self.embed_dim)
        # Add last hidden state

        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=presents,
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
            cross_attentions=all_cross_attentions,
        )


class ModifiedGPT2LMHeadModel(nn.Module):
    def __init__(self, config, block_size):
        super(ModifiedGPT2LMHeadModel, self).__init__()
        self.config = config
        self.transformer = ModifiedGPT2Model(config, block_size)
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
        self.loss_fct = CrossEntropyLoss()

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
        token_type_ids = kwargs.get("token_type_ids", None)
        # only last token for inputs_ids if past is defined in kwargs
        if past:
            input_ids = input_ids[:, -1].unsqueeze(-1)
            if token_type_ids is not None:
                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)

        attention_mask = kwargs.get("attention_mask", None)
        position_ids = kwargs.get("position_ids", None)

        if attention_mask is not None and position_ids is None:
            # create position_ids on the fly for batch generation
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            if past:
                position_ids = position_ids[:, -1].unsqueeze(-1)
        else:
            position_ids = None
        return {
            "input_ids": input_ids,
            "past_key_values": past,
            "position_ids": position_ids,
            "attention_mask": attention_mask,
            "token_type_ids": token_type_ids,
        }

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
    ):
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
        """
        transformer_outputs = self.transformer(
            input_ids,
            attention_mask=attention_mask,
        )
        hidden_states = transformer_outputs[0]

        lm_logits = self.lm_head(hidden_states)

        return lm_logits

        # loss = None
        # if labels is not None:
        #     # Shift so that tokens < n predict n
        #     shift_logits = lm_logits[..., :-1, :].contiguous()
        #     shift_labels = labels[..., 1:].contiguous()
        #     # Flatten the tokens
        #     loss = self.loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))

        # return CausalLMOutputWithCrossAttentions(
        #     loss=loss,
        #     logits=lm_logits,
        #     past_key_values=transformer_outputs.past_key_values,
        #     hidden_states=transformer_outputs.hidden_states,
        #     attentions=transformer_outputs.attentions,
        #     cross_attentions=transformer_outputs.cross_attentions,
        # )

    @staticmethod
    def _reorder_cache(
        past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
    ) -> Tuple[Tuple[torch.Tensor]]:
        """
        This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
        [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
        beam_idx at every generation step.
        """
        return tuple(
            tuple(
                past_state.index_select(0, beam_idx.to(past_state.device))
                for past_state in layer_past
            )
            for layer_past in past
        )