SDPA backend priority (Comfy-Org#9299)

Author: contentis

comfy/ldm/hunyuan3d/vae.py

@@ -178,7 +178,7 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
 
 class CrossAttentionProcessor:
     def __call__(self, attn, q, k, v):
-        out = F.scaled_dot_product_attention(q, k, v)
+        out = ops.scaled_dot_product_attention(q, k, v)
         return out
 
 

comfy/ldm/modules/attention.py

@@ -448,7 +448,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
             mask = mask.unsqueeze(1)
 
     if SDP_BATCH_LIMIT >= b:
-        out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
+        out = ops.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
         if not skip_output_reshape:
             out = (
                 out.transpose(1, 2).reshape(b, -1, heads * dim_head)
@@ -461,7 +461,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
                 if mask.shape[0] > 1:
                     m = mask[i : i + SDP_BATCH_LIMIT]
 
-            out[i : i + SDP_BATCH_LIMIT] = torch.nn.functional.scaled_dot_product_attention(
+            out[i : i + SDP_BATCH_LIMIT] = ops.scaled_dot_product_attention(
                 q[i : i + SDP_BATCH_LIMIT],
                 k[i : i + SDP_BATCH_LIMIT],
                 v[i : i + SDP_BATCH_LIMIT],

comfy/ldm/modules/diffusionmodules/model.py

@@ -285,7 +285,7 @@ def pytorch_attention(q, k, v):
     )
 
     try:
-        out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False)
+        out = ops.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False)
         out = out.transpose(2, 3).reshape(orig_shape)
     except model_management.OOM_EXCEPTION:
         logging.warning("scaled_dot_product_attention OOMed: switched to slice attention")

comfy/ops.py

@@ -23,9 +23,18 @@
 import comfy.float
 import comfy.rmsnorm
 import contextlib
+from torch.nn.attention import SDPBackend, sdpa_kernel
 
 cast_to = comfy.model_management.cast_to #TODO: remove once no more references
 
+SDPA_BACKEND_PRIORITY = [
+    SDPBackend.FLASH_ATTENTION,
+    SDPBackend.EFFICIENT_ATTENTION,
+    SDPBackend.MATH,
+]
+if torch.cuda.is_available():
+    SDPA_BACKEND_PRIORITY.insert(0, SDPBackend.CUDNN_ATTENTION)
+
 def cast_to_input(weight, input, non_blocking=False, copy=True):
     return comfy.model_management.cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy)
 
@@ -249,6 +258,10 @@ def conv_nd(s, dims, *args, **kwargs):
         else:
             raise ValueError(f"unsupported dimensions: {dims}")
 
+    @staticmethod
+    @sdpa_kernel(backends=SDPA_BACKEND_PRIORITY, set_priority=True)
+    def scaled_dot_product_attention(q, k, v, *args, **kwargs):
+        return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
 
 class manual_cast(disable_weight_init):
     class Linear(disable_weight_init.Linear):