the-crypt-keeper / can-ai-code

Self-evaluating interview for AI coders
https://huggingface.co/spaces/mike-ravkine/can-ai-code-results
MIT License
537 stars 30 forks source link

Evaluate TheBloke/Phind-CodeLlama-34B-v2-AWQ #94

Closed the-crypt-keeper closed 1 year ago

the-crypt-keeper commented 1 year ago

Latest vLLM adds AWQ support, it would be interesting to compare performance vs the native awq executor we already support:

the-crypt-keeper commented 1 year ago

its working through vLLM but it turns out this is a new kind of AWQ:

https://github.com/casper-hansen/AutoAWQ

these models are not working with even the latest version of the old awq repo:

RuntimeError: shape '[1, 29, 1024]' is invalid for input of size 237568

So looks like will have to implement autoawq to replace the legacy awq.

the-crypt-keeper commented 1 year ago

Completed via vllm-awq.