Skip to content

Improve AMD performance. #10302

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Oct 12, 2025
+2 −5
Merged

Improve AMD performance. #10302

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 2 additions & 5 deletions comfy/model_management.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -332,6 +332,7 @@ def amd_min_version(device=None, min_rdna_version=0):
SUPPORT_FP8_OPS = args.supports_fp8_compute
try:
if is_amd():
torch.backends.cudnn.enabled = False # Seems to improve things a lot on AMD
try:
rocm_version = tuple(map(int, str(torch.version.hip).split(".")[:2]))
except:
Expand Down Expand Up @@ -925,11 +926,7 @@ def vae_dtype(device=None, allowed_dtypes=[]):
if d == torch.float16 and should_use_fp16(device):
return d

# NOTE: bfloat16 seems to work on AMD for the VAE but is extremely slow in some cases compared to fp32
# slowness still a problem on pytorch nightly 2.9.0.dev20250720+rocm6.4 tested on RDNA3
# also a problem on RDNA4 except fp32 is also slow there.
# This is due to large bf16 convolutions being extremely slow.
if d == torch.bfloat16 and ((not is_amd()) or amd_min_version(device, min_rdna_version=4)) and should_use_bf16(device):
if d == torch.bfloat16 and should_use_bf16(device):
return d

return torch.float32
Expand Down