min_dalle/models/dalle_bart_decoder.py

from typing import Tuple, List
import torch
from torch import nn, LongTensor, FloatTensor, BoolTensor
from .dalle_bart_encoder import GLU, AttentionBase
import gc
import tracemalloc

IMAGE_TOKEN_COUNT = 256


class DecoderCrossAttention(AttentionBase):
    def forward(
        self,
        decoder_state: FloatTensor,
        encoder_state: FloatTensor,
        attention_mask: BoolTensor
    ) -> FloatTensor:
        keys = self.k_proj.forward(encoder_state)
        values = self.v_proj.forward(encoder_state)
        queries = self.q_proj.forward(decoder_state)
        return super().forward(keys, values, queries, attention_mask)


class DecoderSelfAttention(AttentionBase):
    def __init__(self, head_count: int, embed_count: int):
        super().__init__(head_count, embed_count)


    def forward(
        self, 
        decoder_state: FloatTensor,
        attention_state: FloatTensor,
        attn_mask: BoolTensor,
        token_index: LongTensor
    ) -> Tuple[FloatTensor, FloatTensor]:
        keys = self.k_proj.forward(decoder_state)
        values = self.v_proj.forward(decoder_state)
        queries = self.q_proj.forward(decoder_state)
        attn_state_new = torch.cat([keys, values]).to(attention_state.dtype)
        attention_state[:, token_index] = attn_state_new
        batch_count = decoder_state.shape[0]
        keys = attention_state[:batch_count]
        values = attention_state[batch_count:]
        decoder_state = super().forward(keys, values, queries, attn_mask)
        return decoder_state, attention_state


class DecoderLayer(nn.Module):
    def __init__(
        self, 
        head_count: int, 
        embed_count: int,
        glu_embed_count: int,
        device: str
    ):
        super().__init__()
        self.pre_self_attn_layer_norm = nn.LayerNorm(embed_count)
        self.self_attn = DecoderSelfAttention(head_count, embed_count)
        self.self_attn_layer_norm = nn.LayerNorm(embed_count)
        self.pre_encoder_attn_layer_norm = nn.LayerNorm(embed_count)
        self.encoder_attn = DecoderCrossAttention(head_count, embed_count)
        self.encoder_attn_layer_norm = nn.LayerNorm(embed_count)
        self.glu = GLU(embed_count, glu_embed_count)
        self.token_indices = torch.arange(IMAGE_TOKEN_COUNT, device=device)


    def forward(
        self,
        decoder_state: FloatTensor,
        encoder_state: FloatTensor,
        attention_state: FloatTensor,
        attention_mask: BoolTensor,
        token_index: LongTensor
    ) -> Tuple[FloatTensor, FloatTensor]:
        # Self Attention
        self_attn_mask = self.token_indices < token_index + 1
        self_attn_mask = self_attn_mask[None][[0] * decoder_state.shape[0]]
        residual = decoder_state
        decoder_state = self.pre_self_attn_layer_norm.forward(decoder_state)
        decoder_state, attention_state = self.self_attn.forward(
            decoder_state=decoder_state,
            attention_state=attention_state,
            attn_mask=self_attn_mask,
            token_index=token_index
        )
        decoder_state = self.self_attn_layer_norm.forward(decoder_state)
        decoder_state = residual + decoder_state

        # Cross Attention
        residual = decoder_state
        decoder_state = self.pre_encoder_attn_layer_norm.forward(decoder_state)
        decoder_state = self.encoder_attn.forward(
            decoder_state=decoder_state,
            encoder_state=encoder_state,
            attention_mask=attention_mask
        )
        decoder_state = self.encoder_attn_layer_norm.forward(decoder_state)
        decoder_state = residual + decoder_state

        # Feed forward
        residual = decoder_state
        decoder_state = self.glu.forward(decoder_state)
        decoder_state = residual + decoder_state

        

        return decoder_state, attention_state


class DalleBartDecoder(nn.Module):
    def __init__(
        self,
        image_vocab_count: int,
        embed_count: int,
        attention_head_count: int,
        glu_embed_count: int,
        layer_count: int,
        device: str
    ):
        super().__init__()
        self.layer_count = layer_count
        self.embed_count = embed_count
        self.image_vocab_count = image_vocab_count
        self.embed_tokens = nn.Embedding(image_vocab_count + 1, embed_count)
        self.embed_positions = nn.Embedding(IMAGE_TOKEN_COUNT, embed_count)
        self.layers: List[DecoderLayer] = nn.ModuleList([
            DecoderLayer(
                head_count=attention_head_count,
                embed_count=embed_count,
                glu_embed_count=glu_embed_count,
                device=device
            ) 
            for _ in range(layer_count)
        ])
        self.layernorm_embedding = nn.LayerNorm(embed_count)
        self.final_ln = nn.LayerNorm(embed_count)
        self.lm_head = nn.Linear(embed_count, image_vocab_count + 1, bias=False)
        self.token_indices = torch.arange(IMAGE_TOKEN_COUNT, device=device)


    def forward(
        self,
        settings: FloatTensor,
        attention_mask: BoolTensor,
        encoder_state: FloatTensor,
        attention_state: FloatTensor,
        prev_tokens: LongTensor,
        token_index: LongTensor
    ) -> Tuple[LongTensor, FloatTensor]:
        image_count = encoder_state.shape[0] // 2
        token_index_batched = token_index[[0] * image_count * 2]
        prev_tokens = prev_tokens[list(range(image_count)) * 2]
        prev_tokens.clamp_(0, self.image_vocab_count)
        decoder_state = self.embed_tokens.forward(prev_tokens)
        decoder_state += self.embed_positions.forward(token_index_batched)
        decoder_state = self.layernorm_embedding.forward(decoder_state)
        decoder_state = decoder_state[:, None]
        
        tracemalloc.start()
        print("--")
        # displaying the memory
        print(tracemalloc.get_traced_memory())
        
        for i in range(self.layer_count):
            decoder_state, attention_state[i] = self.layers[i].forward(
                decoder_state,
                encoder_state,
                attention_state[i],
                attention_mask,
                token_index
            )
        print(tracemalloc.get_traced_memory())
        decoder_state = self.final_ln(decoder_state)
        logits = self.lm_head(decoder_state)
        print(tracemalloc.get_traced_memory())
        del decoder_state 
        temperature = settings[[0]]
        top_k = settings[[1]].to(torch.long)
        print(tracemalloc.get_traced_memory())
        supercondition_factor = settings[[2]]
        logits = logits[:, -1, : 2 ** 14]
        logits: FloatTensor = (
            logits[:image_count] * (1 - supercondition_factor) + 
            logits[image_count:] * supercondition_factor
        )
        print(tracemalloc.get_traced_memory())
        del supercondition_factor 
        logits_sorted, _ = logits.sort(descending=True)
        is_kept = logits >= logits_sorted[:, top_k - 1]
        del top_k
        logits -= logits_sorted[:, [0]]
        del logits_sorted
        logits /= temperature
        del temperature
        logits.exp_()
        logits *= is_kept.to(torch.float32)
        del is_kept
        image_tokens = torch.multinomial(logits, 1)[:, 0]
        del logits
        gc.collect()
        
        print(tracemalloc.get_traced_memory())
        
        return image_tokens, attention_state