candle
candle copied to clipboard
Add quantized CLIP
I was wanting to test quantization for CLIP for use in WASM and wrote this for my own but wanted to upstream it.
For making the quantized version, I have a fork of clip.cpp which uses the hugging face keys to match the original model.
https://github.com/rockerBOO/clip.cpp/tree/hf-format
I am proposing this in a limited scope currently to gauge interest, but I can provide quantized models for testing if this PR is desired.
If anything needs updating or fixed, let me know.
Thank you.