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.pyindex 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)