Spaces:
Running
on
Zero
extremely slow and lower quality in my own duplicated space
hi, i duplicated this space without modifying any code and configurations, but it could not reproduce the same quality and time consumption. i tested with a 19min audio, it costed 10s to transcribe here but in my own space it's nearly 1 min, and the quality of the output text was significantly lower. any ideas? thanks in advance.
Hi,
Are you running on ZeroGPU or CPU?
Does it say "Running on ZERO"?
Hi,
Thanks for the reply.Yes I did run it on a ZeroGPU space, and I managed to figure out why.
It seems to be the version compatibility issue of flash-attn and torch.
Maybe the default version of these packages has been changed.
I got it running by assigning fixed version of torch and moving the pip install of flash-attn back into the requirements.txt in the way described in last part of this discussion:
https://huggingface.co/spaces/zero-gpu-explorers/README/discussions/63
Thanks again for your reply.
Glad you found the fix! I will probably have to implement this solution too when the space next restarts.