Patrick Devine
Patrick Devine
This now carries #3084
@amitkot the quantize stuff I demoed in Paris isn't quite ready to go, but this should have been able to convert stock gemma. What was the huggingface repo of the...
@amitkot when I ran it I got: ``` % ./ollama run pdevine/gemma-hebrew >>> שלום! מה שלומך היום? Kiedy אתה נולד? Kiedy הוא נולד? Kiedy הם נולדו? Kiedy אתה הולדת? Kiedy...
@amitkot ah, yeah this change is very different from that. You just create a modelfile like: ``` FROM /Users/pdevine/git/Hebrew-Gemma-11B-V2 TEMPLATE """user {{ if .System }}{{ .System }} {{ end }}{{...
I'm going to close this, but we can reopen it if it's still a problem.
Thanks for the issue, @chrisbward . I think you'll need to ask the Huggingface team to implement that. I think it'd be a really cool feature though.
@Manouchehri @YueChenkkk The `invalid wire-format data` issue should be fixed with the `0.3.8`. The problem is that the HF model included both the safetensors and pytorch weights and Ollama was...
I think we can safely close this.
Also, I would definitely recommend checking out some of the new reasoning models.
This was incorporated into a different PR/commit.