comfyanonymous
Extreme CPU use with Aysnc disabled / Torch compile disabled 13700k/RTX 5090
Custom Node Testing
- [x] I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)
Your question
I'm unsure when this started to happen but it was most likely within the last month, that when generating my CPU usage will go to 80-90% with my GPU at 90-100% (Which is normal) It seems to be bottlenecking my IT/s at around 8 when i normally get 10.5-11It/s (SDXL 1k res) I did hear their was a bug with Nodes 2.0 which I have disabled but this issue is still happening, I tried with all custom nodes disabled and the same problem happens. Even running with -HighVram and -SageAttention does not stop the high CPU use so I'm suspecting it must be some sort of bug or Nodes 2.0 is somehow still turned on.
Logs
Other
No response
You should try with
--disable-async-offload
Oh, sorry I forgot to mention I have run it with the launch augments,
--highvram --disable-smart-memory --disable-pinned-memory --disable-async-offload --use-pytorch-cross-attention And this did also not help.
Hello, I have similar issue with XPU, but for my 4090 I am using more older version of Pytorch, can you try to downgrade it to torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128 and check if this behaviour has something with Pytorch?
Hello, I have similar issue with XPU, but for my 4090 I am using more older version of Pytorch, can you try to downgrade it to torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128 and check if this behaviour has something with Pytorch?
It seems your right on the money, I downgraded a test comfy to 2.8.0 and it now uses 0 CPU and is much faster. I'm unsure why 2.9 is cuasing high CPU load with reduce speeds.
It seems that there is some issue related to our CPU or something else. This behavior occurs on my B60 GPU with the same Pytroch version. I will prepare a new issue for the PyTorch team and link this error in it, it would be helpful if you could join the issue as soon as it’s ready.