[DO NOT MERGE YET - STILL TESTING] Add more ONNX weights (dtypes)
Thanks Joshua!
If you have the magic prompt you used to generate all of those flavors, we can update https://github.com/mozilla/distilvit/blob/main/distilvit/train.py#L331 for the next time I train again
Sure thing - it's using the v3 conversion script: https://github.com/xenova/transformers.js/blob/v3/scripts/convert.py
Command:
python -m scripts.convert --quantize --model_id Mozilla/distilvit
I'm still doing some testing with v3 (mainly testing WebGPU support accuracy of different quantizations) and will keep you updated!
oh ok, neat thanks.
This is our currently vendored version in FF - https://hg.mozilla.org/mozilla-central/file/tip/toolkit/components/ml/vendor/build.sh
Looking at the latest onnx releases, it should be simpler to vendor for us once we can move to transformers.js v3
Thanks Joshua, I added them with the most recent v3