src/lmr/models copy 2/transformer_bert/components/attention.py

import torch
import torch.nn as nn
from torchtune.modules import RotaryPositionalEmbeddings
from torch.nn.attention.flex_attention import flex_attention
from torch.nn.attention import sdpa_kernel, SDPBackend

class Attention(nn.Module):
    def __init__(self, config):
        super().__init__()
        
        self.config = config
        
        # Linear layers...
        self.W_q = nn.Linear(config.embed_dim, config.hidden_dim, bias=False)
        self.W_k = nn.Linear(config.embed_dim, config.hidden_dim, bias=False)
        self.W_v = nn.Linear(config.embed_dim, config.hidden_dim, bias=True)
        self.W_o = nn.Linear(config.hidden_dim, config.embed_dim, bias=True)
        
        # RoPE
        d_k = config.hidden_dim // config.n_heads
        self.rotary_embeddings = RotaryPositionalEmbeddings(d_k, max_seq_len=config.max_seq_len + 10)
        
        self.drop_resid = nn.Dropout(0.1)
        
    def forward(self, x):
        
        B, S, E = x.shape
        H = self.config.n_heads
        D_h = self.config.hidden_dim // H
        
        # 1. Project and Reshape (B, S, E) -> (B, S, H, D_h)
        q = self.W_q(x).view(B, S, H, D_h)
        k = self.W_k(x).view(B, S, H, D_h)
        v = self.W_v(x).view(B, S, H, D_h)
        
        # 2. Apply RoPE and Transpose to (B, H, S, D_h)
        q = self.rotary_embeddings(q).transpose(1, 2).contiguous()
        k = self.rotary_embeddings(k).transpose(1, 2).contiguous()
        v = v.transpose(1, 2).contiguous() # V is often transposed without RoPE
        
        # 3. Custom Bidirectional Masking Function (No-op)
        def _score_mod(scores, b, h, i, j):
            # This implements bidirectional attention by applying NO mask.
            # All tokens are visible to all other tokens.
            return scores 

        # 4. Attention with Explicit Backend Control
        # sdpa_kernel controls the backend used for the attention operation within the block
        with sdpa_kernel(SDPBackend.FLASH_ATTENTION):
            attn_output = flex_attention(
                q, k, v,
                score_mod=_score_mod, # <-- Apply the custom bidirectional (no-op) mask
                scale=None,
            )
        
        # 5. Combine Heads and Project Back
        # (B, H, S, D_h) -> (B, S, H*D_h)
        attn_output = attn_output.transpose(1, 2).contiguous().view(B, S, self.config.hidden_dim)
        attn_output = self.W_o(attn_output)
        attn_output = self.drop_resid(attn_output)
        
        return attn_output