Open manishiitg opened 1 year ago
This issue should be solved upstream in ctranslate2. Once ctranslate2 supports mistral, tgi can bump the version.
Allowing a pre-converted model is not really on the roadmap, tgi really wants to download .safetensors in pytorch format. in that way, ct2 quantization, just quick-fixes this issue.
Feature request
ctranslate2 doesn't support mistral model conversion out of the box https://github.com/OpenNMT/CTranslate2/issues/1501
currently when using ct2 it automatically converts the model on the fly.
can we allow support to convert model to ct2 offline? that way i can convert mistral model offline and then use it with TGI.
Motivation
mistral support
Your contribution
-