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.

Auto enable mem efficient attention on gfx1100 on pytorch nightly 2.7

comfyanonymousd7b4bf21a2de
diff --git a/comfy/model_management.py b/comfy/model_management.py
index 212ce9af2..dd8a2a28f 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -236,6 +236,19 @@ try:
except:
pass
+
+try:
+ if is_amd():
+ arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName
+ logging.info("AMD arch: {}".format(arch))
+ if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
+ if int(torch_version[0]) >= 2 and int(torch_version[2]) >= 7: # works on 2.6 but doesn't actually seem to improve much
+ if arch in ["gfx1100"]: #TODO: more arches
+ ENABLE_PYTORCH_ATTENTION = True
+except:
+ pass
+
+
if ENABLE_PYTORCH_ATTENTION:
torch.backends.cuda.enable_math_sdp(True)
torch.backends.cuda.enable_flash_sdp(True)