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.

Enable Pytorch Attention for gfx1150 (#12973)

lostdisc3814bf4454ef
diff --git a/comfy/model_management.py b/comfy/model_management.py
index 442d5a40a..a4af5ddb2 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -400,7 +400,7 @@ try:
if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
if aotriton_supported(arch): # AMD efficient attention implementation depends on aotriton.
if torch_version_numeric >= (2, 7): # works on 2.6 but doesn't actually seem to improve much
- if any((a in arch) for a in ["gfx90a", "gfx942", "gfx950", "gfx1100", "gfx1101", "gfx1151"]): # TODO: more arches, TODO: gfx950
+ if any((a in arch) for a in ["gfx90a", "gfx942", "gfx950", "gfx1100", "gfx1101", "gfx1150", "gfx1151"]): # TODO: more arches, TODO: gfx950
ENABLE_PYTORCH_ATTENTION = True
if rocm_version >= (7, 0):
if any((a in arch) for a in ["gfx1200", "gfx1201"]):