Joshua Lochner

I've uploaded q8 (uint8/int8) weights for the model. Can you try it out? Setting `{dtype: 'q8'}` will use those weights.

Oh thanks, I know what the issue is and I'll fix it tomorrow!

Okay the corrected quantized weights are now up! Please let me know if it works now. 😇 (Just remember to clear the cache)

Thanks for all your testing! I don't think this is an issue from your side, I think it's a problem with the quantization settings. Will do some investigating tomorrow. In...

Wow that is amazing! 🔥 Great stuff! 🚀 The 8-bit quantized version still seems to have some issues (audio is not perfect), but I'll play around with a few more...

Hi again! I've done some additional testing and added per-model dtypes and devices, so you can do the following: ```js const model = await MusicgenForConditionalGeneration.from_pretrained(model_id, { dtype: { text_encoder: 'q8',...

Hey again 👋 I put out a (very simple) demo myself: https://huggingface.co/spaces/Xenova/musicgen-web https://github.com/xenova/transformers.js/assets/26504141/f20b683d-2fd5-4a66-81e1-775c859a0c51 The progress tracking is now possible thanks to the `Streamer` API, which I added today. I've added...

Are you referring to an example/demo application we've created? If so, please be more specific :)

@llmexperiment You may be interested in looking at the HF transformers implementation ([PR here](https://github.com/huggingface/transformers/pull/28094)), which supports a fallback if `causal-conv1d` is not found in the environment. I've also been trying...

update: https://github.com/huggingface/huggingface.js/pull/525 The above template is now supported (as well as all models on the HF Hub with >= 10 downloads), but I do still think it will be a...