Merge 275fa4b060 into b9b6ce137b

nanochat/flash_attention.py

@@ -101,6 +101,25 @@ def _sdpa_attention(q, k, v, window_size, enable_gqa):
 
     return F.scaled_dot_product_attention(q, k, v, attn_mask=mask, enable_gqa=enable_gqa)
 
+
+def XSA(atten, value, sdpa=False):
+    if sdpa:
+        aB, aH, aT, aD = atten.shape
+        vB, vH, vT, vD = value.shape
+        rep = aH // vH
+        assert aH % vH == 0, "query heads and kv heads must be divisible"
+        value = torch.repeat_interleave(value, rep, dim=1)
+        Vn = F.normalize(value, dim=-1)
+        return atten - torch.sum(atten * Vn, dim=-1, keepdim=True) * Vn
+    else:
+        aB, aT, aH, aD = atten.shape
+        vB, vT, vH, vD = value.shape
+        rep = aH // vH
+        assert aH % vH == 0, "query heads and kv heads must be divisible"
+        value = torch.repeat_interleave(value, rep, dim=2)
+        Vn = F.normalize(value, dim=-1)
+        return atten - torch.sum(atten * Vn, dim=-1, keepdim=True) * Vn
+
 # =============================================================================
 # Public API: Same interface as FA3
 # =============================================================================
@@ -117,7 +136,9 @@ def flash_attn_func(q, k, v, causal=False, window_size=(-1, -1)):
         Output tensor of shape (B, T, H, D)
     """
     if USE_FA3:
-        return _fa3.flash_attn_func(q, k, v, causal=causal, window_size=window_size)
+        y =  _fa3.flash_attn_func(q, k, v, causal=causal, window_size=window_size)
+        y = XSA(y, v)
+        return y
 
     # SDPA fallback: transpose (B, T, H, D) -> (B, H, T, D)
     q = q.transpose(1, 2)
@@ -125,6 +146,7 @@ def flash_attn_func(q, k, v, causal=False, window_size=(-1, -1)):
     v = v.transpose(1, 2)
     enable_gqa = q.size(1) != k.size(1)
     y = _sdpa_attention(q, k, v, window_size, enable_gqa)
+    y = XSA(y, v, True)
     return y.transpose(1, 2)  # back to (B, T, H, D)
 
 
@@ -147,10 +169,12 @@ def flash_attn_with_kvcache(q, k_cache, v_cache, k=None, v=None, cache_seqlens=N
         Output tensor of shape (B, T_new, H, D)
     """
     if USE_FA3:
-        return _fa3.flash_attn_with_kvcache(
+        y = _fa3.flash_attn_with_kvcache(
             q, k_cache, v_cache, k=k, v=v, cache_seqlens=cache_seqlens,
             causal=causal, window_size=window_size
         )
+        y = XSA(y, v)
+        return y
 
     # SDPA fallback: manually manage KV cache
     B, T_new, H, D = q.shape
@@ -170,9 +194,11 @@ def flash_attn_with_kvcache(q, k_cache, v_cache, k=None, v=None, cache_seqlens=N
     q_sdpa = q.transpose(1, 2)
     k_sdpa = k_full.transpose(1, 2)
     v_sdpa = v_full.transpose(1, 2)
+    v = v.transpose(1, 2)
 
     enable_gqa = q_sdpa.size(1) != k_sdpa.size(1)
     y_sdpa = _sdpa_attention(q_sdpa, k_sdpa, v_sdpa, window_size, enable_gqa)
+    y_sdpa = XSA(y_sdpa, v, True)
 
     return y_sdpa.transpose(1, 2)  # back to (B, T, H, D)