wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-03-16 22:58:19 +08:00
Code Issues Actions 11 Packages Projects Releases Wiki Activity

Merge cb213aee66 into 3814bf4454

Browse Source
This commit is contained in:
Alex Butler 2026-03-15 22:59:58 +01:00 committed by GitHub
commit 62833446d0
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 5 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
README.md
View File

@ -374,7 +374,11 @@ You can enable experimental memory efficient attention on recent pytorch in Comf
```TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 python main.py --use-pytorch-cross-attention```
You can also try setting this env variable `PYTORCH_TUNABLEOP_ENABLED=1` which might speed things up at the cost of a very slow initial run.
You can also try:
* Tunable ops: Setting `PYTORCH_TUNABLEOP_ENABLED=1` which might speed things up at the cost of a very slow initial runs. After running online tuning for a while consider disabling it with `PYTORCH_TUNABLEOP_TUNING=0` to only used the tuned settings and avoid slowdowns.
* MIOpen: Currently disabled by default. Enable with `COMFYUI_ENABLE_MIOPEN=1`. Be aware that miopen will autotune by default, consider disabling it with `MIOPEN_FIND_MODE=FAST` to avoid tuning slowdowns.
* Flash attention: Install from [flash-attention](https://github.com/Dao-AILab/flash-attention) & enable with `FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE` and arg `--use-flash-attention`. See also notes in the repo on triton autotuning.
* If you are encountering VRAM OOMs `PYTORCH_NO_HIP_MEMORY_CACHING=1` may help.
# Notes