shenglin.net/ComfyUI-MochiWrapper-2
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update readme.md

Browse Source
This commit is contained in:
Jukka Seppänen 2024-10-24 02:55:09 +03:00 committed by GitHub
parent 00a550e81c
commit cafd0741f5
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 1 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
readme.md
View File

@ -6,18 +6,7 @@
https://github.com/user-attachments/assets/a714b70f-dcdb-4f91-8a3d-8da679a28d6e
## Requires flash_attn !
Not sure if this can be worked around, I compiled a wheel for my Windows setup (Python 3.12, torch 2.5.0+cu124) that worked for me:
https://huggingface.co/Kijai/Mochi_preview_comfy/blob/main/flash_attn-2.6.3-cp312-cp312-win_amd64.whl
Python 3.10 / CUDA 12.4 / Torch 2.4.1:
https://huggingface.co/Kijai/Mochi_preview_comfy/blob/main/flash_attn-2.6.3-cp310-cp310-win_amd64.whl
Other sources for pre-compiled wheels:
https://github.com/oobabooga/flash-attention/releases
Can use flash_attn, pytorch attention (sdpa) or [sage attention](https://github.com/thu-ml/SageAttention), sage being fastest.
Depending on frame count can fit under 20GB, VAE decoding is heavy and there is experimental tiled decoder (taken from CogVideoX -diffusers code) which allows higher frame counts, so far highest I've done is 97 with the default tile size 2x2 grid.