refactor: simplify attention mask handling

2026-07-20 20:36:16 +08:00
parent a6e920fdb0
commit d7ac66fb73
6 changed files with 34 additions and 38 deletions
@@ -76,9 +76,8 @@ class GQA(nn.Module):
        rotary_emb: Tensor,
        attn_mask: Tensor = None,
        paged_cache: Optional[CacheView] = None,
+        is_causal: bool = False,
    ) -> Tensor:
-        is_causal = attn_mask is None
-
        q = self._split_heads(self.q_proj(x), self.n_heads)
        k = self._split_heads(self.k_proj(x), self.n_kv_heads)
        v = self._split_heads(self.v_proj(x), self.n_kv_heads)
@@ -163,9 +162,9 @@ class MLA(nn.Module):
        rotary_emb: Tensor,
        attn_mask: Tensor = None,
        paged_cache: Optional[CacheView] = None,
+        is_causal: bool = False,
    ) -> Tensor:
        bsz, seq_len, _ = x.size()
-        is_causal = attn_mask is None

        q = self.q_proj(x)
        q = q.view(bsz, seq_len, self.n_heads, self.head_dim)
@@ -26,12 +26,14 @@ class DecoderBlock(nn.Module):
        rotary_emb: Tensor,
        attention_mask: Optional[Tensor] = None,
        paged_cache: Optional[CacheView] = None,
+        is_causal: bool = False,
    ) -> Tensor:
        attn_output = self.attention(
            self.input_norm(x),
            rotary_emb,
            attention_mask,
            paged_cache,
+            is_causal,
        )
        x = attn_output + x
        x = self.mlp(self.post_attention_norm(x)) + x