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.pyindex 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: gfx950ENABLE_PYTORCH_ATTENTION = Trueif rocm_version >= (7, 0):if any((a in arch) for a in ["gfx1200", "gfx1201"]):