Spaces:
Running
Space works well on A10, but failed on `ZeroGPU`
the space is https://huggingface.co/spaces/AnchorFake/MuseVSpace.
Current situation is that, MuseVSpace works well on A10 but failed on ZeroGPU.
The failed info
I found base docker image on ZeroGPU is docker.io/library/python:3.10.13@sha256:d5b1fbbc00fd3b55620a9314222498bebf09c4bf606425bf464709ed6a79f202, while on A10 is FROM docker.io/nvidia/cuda:12.3.2-cudnn9-devel-ubuntu22.04@sha256:fb1ad20f2552f5b3aafb2c9c478ed57da95e2bb027d15218d7a55b3a0e4b4413.
Is this the reason?
No it’s because you need to modify the code to work on ZERO
No it’s because you need to modify the code to work on ZERO
@mrfakename Thanks for your response. Where can i find the features or docs about ZEROGPU? i have no idea how to modify it.
@AnchorFake it is here - https://huggingface.co/zero-gpu-explorers
You need to use the decorator on the method that will require GPU usuage.
[email protected]
def generate(prompt):
@codelion
hi, thanks for your advice.
I have decorated the hg_predict method like this.
https://huggingface.co/spaces/AnchorFake/MuseVSpace/blob/main/app_gradio_space.py#L76
https://huggingface.co/spaces/AnchorFake/MuseVSpace/blob/main/app_gradio_space.py#L97
@AnchorFake the acutal code in your app that requires GPU is buried deep somewhere here - https://huggingface.co/spaces/AnchorFake/MuseVSpace/blob/main/MuseV/musev/pipelines/pipeline_controlnet_predictor.py#L244
I am not sure if it will work as it is. Ideally you should be able to decorate only the method that requires GPU with spaces.GPU. But as is in your code there are many places where this happens.