Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Fix last commit not working on older pytorch. (#9346)

comfyanonymous4e5c230f6a95
diff --git a/comfy/ops.py b/comfy/ops.py
index 2be35f76a..18e7db705 100644
--- a/comfy/ops.py
+++ b/comfy/ops.py
@@ -32,20 +32,21 @@ def scaled_dot_product_attention(q, k, v, *args, **kwargs):
try:
if torch.cuda.is_available():
from torch.nn.attention import SDPBackend, sdpa_kernel
-
- SDPA_BACKEND_PRIORITY = [
- SDPBackend.FLASH_ATTENTION,
- SDPBackend.EFFICIENT_ATTENTION,
- SDPBackend.MATH,
- ]
-
- SDPA_BACKEND_PRIORITY.insert(0, SDPBackend.CUDNN_ATTENTION)
-
- def scaled_dot_product_attention(q, k, v, *args, **kwargs):
- # Use this (rather than the decorator syntax) to eliminate graph
- # break for pytorch < 2.9
- with sdpa_kernel(SDPA_BACKEND_PRIORITY, set_priority=True):
- return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
+ import inspect
+ if "set_priority" in inspect.signature(sdpa_kernel).parameters:
+ SDPA_BACKEND_PRIORITY = [
+ SDPBackend.FLASH_ATTENTION,
+ SDPBackend.EFFICIENT_ATTENTION,
+ SDPBackend.MATH,
+ ]
+
+ SDPA_BACKEND_PRIORITY.insert(0, SDPBackend.CUDNN_ATTENTION)
+
+ def scaled_dot_product_attention(q, k, v, *args, **kwargs):
+ with sdpa_kernel(SDPA_BACKEND_PRIORITY, set_priority=True):
+ return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
+ else:
+ logging.warning("Torch version too old to set sdpa backend priority.")
except (ModuleNotFoundError, TypeError):
logging.warning("Could not set sdpa backend priority.")