Open SinanAkkoyun opened 2 days ago
This is a request, so I'm removing the bug tag. Nested directories are currently not supported due to complexity and I'm on the fence of supporting this in the first place. A model name is a name, not a path.
Currently, if you want to have different bpw models, append -6bpw
to the name of the folder rather than sub-nesting.
[ Linux | CUDA 12.2 | py 3.11 ]
Describe the bug
Hi, thanks for your awesome work!
When infering
Llama3.1/70B-Instruct/6.0bpw
over the OAI API, it throws:However, when infering the model string
6.0bpw
it works.Reproduction steps
Restructure your models directory to have nested structure:
Llama3.1
Expected behavior
It should only listen to the full model string
Acknowledgements