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.

Apple most likely is never fixing the fp16 attention bug. (#8485)

comfyanonymous6e28a46454f2
diff --git a/comfy/model_management.py b/comfy/model_management.py
index 283815611..054291432 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -1052,7 +1052,7 @@ def pytorch_attention_flash_attention():
global ENABLE_PYTORCH_ATTENTION
if ENABLE_PYTORCH_ATTENTION:
#TODO: more reliable way of checking for flash attention?
- if is_nvidia(): #pytorch flash attention only works on Nvidia
+ if is_nvidia():
return True
if is_intel_xpu():
return True
@@ -1068,7 +1068,7 @@ def force_upcast_attention_dtype():
upcast = args.force_upcast_attention
macos_version = mac_version()
- if macos_version is not None and ((14, 5) <= macos_version < (16,)): # black image bug on recent versions of macOS
+ if macos_version is not None and ((14, 5) <= macos_version): # black image bug on recent versions of macOS, I don't think it's ever getting fixed
upcast = True
if upcast: