src/model/transformer.py

"""State-of-the-art Transformer model implementation."""

import math
from typing import Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn import CrossEntropyLoss
from dataclasses import dataclass

import warnings
warnings.filterwarnings("ignore")

from .layers import RMSNorm, RotaryEmbedding, SwiGLU


@dataclass
class ModelOutput:
    """Model output container."""
    loss: Optional[torch.Tensor] = None
    logits: Optional[torch.Tensor] = None
    hidden_states: Optional[Tuple[torch.Tensor]] = None
    attentions: Optional[Tuple[torch.Tensor]] = None


class CausalSelfAttention(nn.Module):
    """Multi-head self-attention with causal mask and RoPE."""
    
    def __init__(self, config):
        super().__init__()
        assert config.hidden_size % config.num_attention_heads == 0
        
        self.num_attention_heads = config.num_attention_heads
        self.head_dim = config.hidden_size // config.num_attention_heads
        self.hidden_size = config.hidden_size
        
        self.q_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
        self.k_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
        self.v_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
        self.o_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
        
        self.attention_dropout = nn.Dropout(config.attention_dropout)
        self.rotary_emb = RotaryEmbedding(
            self.head_dim,
            max_position_embeddings=config.max_position_embeddings,
            base=config.rope_theta,
        )
        
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        use_cache: bool = False,
    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor]]]:
        bsz, q_len, _ = hidden_states.size()
        
        q = self.q_proj(hidden_states)
        k = self.k_proj(hidden_states)
        v = self.v_proj(hidden_states)
        
        q = q.view(bsz, q_len, self.num_attention_heads, self.head_dim).transpose(1, 2)
        k = k.view(bsz, q_len, self.num_attention_heads, self.head_dim).transpose(1, 2)
        v = v.view(bsz, q_len, self.num_attention_heads, self.head_dim).transpose(1, 2)
        
        # Apply rotary embeddings
        cos, sin = self.rotary_emb(v, seq_len=q_len)
        q, k = self.rotary_emb.apply_rotary_pos_emb(q, k, cos, sin, position_ids)
        
        # Flash attention or standard attention
        attn_weights = torch.matmul(q, k.transpose(2, 3)) / math.sqrt(self.head_dim)
        
        if attention_mask is not None:
            attn_weights = attn_weights + attention_mask
            
        attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(q.dtype)
        attn_weights = self.attention_dropout(attn_weights)
        
        attn_output = torch.matmul(attn_weights, v)
        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
        attn_output = self.o_proj(attn_output)
        
        return attn_output, None


class TransformerBlock(nn.Module):
    """Transformer block with RMSNorm and SwiGLU."""
    
    def __init__(self, config):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = CausalSelfAttention(config)
        self.mlp = SwiGLU(
            hidden_size=config.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
        )
        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.hidden_dropout = nn.Dropout(config.hidden_dropout)
        
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        use_cache: bool = False,
    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor]]]:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        
        # Self Attention
        hidden_states, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            use_cache=use_cache,
        )
        hidden_states = self.hidden_dropout(hidden_states)
        hidden_states = residual + hidden_states
        
        # MLP
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = self.hidden_dropout(hidden_states)
        hidden_states = residual + hidden_states
        
        return hidden_states, present_key_value


class TransformerModel(nn.Module):
    """Main transformer model."""
    
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.vocab_size = config.vocab_size
        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
        self.layers = nn.ModuleList(
            [TransformerBlock(config) for _ in range(config.num_hidden_layers)]
        )
        self.norm = RMSNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.gradient_checkpointing = False
        
        # Initialize weights
        self.apply(self._init_weights)
        
    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            torch.nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
            if module.bias is not None:
                torch.nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            torch.nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
            
    def get_input_embeddings(self):
        return self.embed_tokens
        
    def set_input_embeddings(self, value):
        self.embed_tokens = value
        
    def forward(
        self,
        input_ids: torch.LongTensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
        use_cache: Optional[bool] = None,
        output_attentions: bool = False,
        output_hidden_states: bool = False,
        return_dict: bool = True,
    ) -> torch.Tensor:
        batch_size, seq_length = input_ids.shape
        
        # Embed tokens
        hidden_states = self.embed_tokens(input_ids)
        
        # Create position IDs
        if position_ids is None:
            position_ids = torch.arange(
                seq_length, dtype=torch.long, device=input_ids.device
            ).unsqueeze(0)

        # Create causal mask
        causal_mask = torch.triu(
            torch.full((seq_length, seq_length), float('-inf'), device=input_ids.device),
            diagonal=1
        ).unsqueeze(0).unsqueeze(0)  # [1, 1, seq_len, seq_len]

        if attention_mask is not None:
            # Convert padding mask [batch, seq_len] to 4D [batch, 1, 1, seq_len] 
            # and combine with causal mask
            expanded_mask = attention_mask[:, None, None, :]  # [batch, 1, 1, seq_len]
            expanded_mask = (1.0 - expanded_mask) * -10000.0  # Convert 0s to -inf
            attention_mask = expanded_mask + causal_mask.expand(input_ids.shape[0], -1, -1, -1)
        else:
            attention_mask = causal_mask.expand(input_ids.shape[0], -1, -1, -1)
            
        # Forward through layers
        for layer in self.layers:
            if self.gradient_checkpointing and self.training:
                hidden_states, _ = torch.utils.checkpoint.checkpoint(
                    layer,
                    hidden_states,
                    attention_mask,
                    position_ids,
                    None,
                    False,
                    use_reentrant=False,
                )
            else:
                hidden_states, _ = layer(
                    hidden_states,
                    attention_mask=attention_mask,
                    position_ids=position_ids,
                    past_key_value=None,
                    use_cache=False,
                )
                
        hidden_states = self.norm(hidden_states)
        return hidden_states


class TransformerForCausalLM(nn.Module):
    """Transformer model with language modeling head."""
    
    def __init__(self, config):
        super().__init__()
        self.model = TransformerModel(config)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        
        # Tie weights
        self.lm_head.weight = self.model.embed_tokens.weight
        
    def forward(
        self,
        input_ids: torch.LongTensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: bool = False,
        output_hidden_states: bool = False,
        return_dict: bool = True,
    ) -> ModelOutput:
        hidden_states = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        
        logits = self.lm_head(hidden_states)
        
        loss = None
        if labels is not None:
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(
                shift_logits.view(-1, shift_logits.size(-1)),
                shift_labels.view(-1)
            )
            
        return ModelOutput(
            loss=loss,
            logits=logits,
            hidden_states=hidden_states,
            attentions=None,
        )
        
    def gradient_checkpointing_enable(self):
        self.model.gradient_checkpointing = True