Add FlashAttention support to msa.py

openfold/model/msa.py

@@ -89,12 +89,14 @@ class MSAAttention(nn.Module):
     @torch.jit.ignore
     def _chunk(self, 
         m: torch.Tensor,
-        biases: List[torch.Tensor],
+        biases: Optional[List[torch.Tensor]],
         chunk_size: int,
         use_memory_efficient_kernel: bool, 
         use_lma: bool,
+        use_flash: bool,
+        flash_mask: Optional[torch.Tensor],
     ) -> torch.Tensor:
-        def fn(m, biases):
+        def fn(m, biases, flash_mask):
             m = self.layer_norm_m(m)
             return self.mha(
                 q_x=m, 
@@ -102,14 +104,23 @@ class MSAAttention(nn.Module):
                 biases=biases,
                 use_memory_efficient_kernel=use_memory_efficient_kernel,
                 use_lma=use_lma,
+                use_flash=use_flash,
+                flash_mask=flash_mask,
             )
 
+        inputs = {"m": m}
+        if(biases is not None):
+            inputs["biases"] = biases
+        else:
+            fn = partial(fn, biases=None)
+        if(use_flash and flash_mask is not None):
+            inputs["flash_mask"] = flash_mask
+        else:
+            fn = partial(fn, flash_mask=None)
+
         return chunk_layer(
             fn,
-            {
-                "m": m, 
-                "biases": biases, 
-            },
+            inputs,
             chunk_size=chunk_size,
             no_batch_dims=len(m.shape[:-2])
         )
@@ -175,6 +186,7 @@ class MSAAttention(nn.Module):
             m, mask_bias, z = self._prep_inputs(
                 m, z, mask, inplace_safe=inplace_safe
             )
+            m = self.layer_norm_m(m)
             q, k, v = self.mha._prep_qkv(m, m)
             return m, q, k, v, mask_bias, z
 
@@ -210,6 +222,7 @@ class MSAAttention(nn.Module):
         chunk_size: Optional[int] = None,
         use_memory_efficient_kernel: bool = False,
         use_lma: bool = False,
+        use_flash: bool = False,
         inplace_safe: bool = False,
         _chunk_logits: Optional[int] = None,
         _checkpoint_chunks: Optional[bool] = None,
@@ -237,6 +250,10 @@ class MSAAttention(nn.Module):
                 inplace_safe=inplace_safe,
             )
        
+        if(use_flash):
+            assert z is None
+            biases = None
+        else:    
             m, mask_bias, z = self._prep_inputs(
                 m, z, mask, inplace_safe=inplace_safe
             )
@@ -252,6 +269,8 @@ class MSAAttention(nn.Module):
                 chunk_size,
                 use_memory_efficient_kernel=use_memory_efficient_kernel, 
                 use_lma=use_lma,
+                use_flash=use_flash,
+                flash_mask=mask,
             )
         else:
             m = self.layer_norm_m(m)
@@ -261,6 +280,8 @@ class MSAAttention(nn.Module):
                 biases=biases,
                 use_memory_efficient_kernel=use_memory_efficient_kernel,
                 use_lma=use_lma,
+                use_flash=use_flash,
+                flash_mask=mask,
             )
 
         return m
@@ -336,6 +357,7 @@ class MSAColumnAttention(nn.Module):
         mask: Optional[torch.Tensor] = None, 
         chunk_size: Optional[int] = None,
         use_lma: bool = False,
+        use_flash: bool = False,
     ) -> torch.Tensor:
         """
         Args:
@@ -353,7 +375,13 @@ class MSAColumnAttention(nn.Module):
         if mask is not None:
             mask = mask.transpose(-1, -2)
 
-        m = self._msa_att(m, mask=mask, chunk_size=chunk_size, use_lma=use_lma)
+        m = self._msa_att(
+            m, 
+            mask=mask, 
+            chunk_size=chunk_size, 
+            use_lma=use_lma,
+            use_flash=use_flash,
+        )
 
         # [*, N_seq, N_res, C_in]
         m = m.transpose(-2, -3)