Vatsal Aggarwal

One backprop step for Natgrad & Adam

1

**PR type:** new feature **Related issue(s)/PRs:** ## Summary **Proposed changes** Stochastic Variational Inference based GP models typically combine NatGrad to train the variational parameters along with Adam to train the...

https://github.com/metavoiceio/metavoice-src/issues/1 ``` PYTORCH_ENABLE_MPS_FALLBACK=1 python fam/llm/sample.py \ --huggingface_repo_id="metavoiceio/metavoice-1B-v0.1" \ --spk_cond_path="assets/bria.mp3" \ --text "Hi." ```

Hey great work! Just wondering any chance you would have the loss evolution for your pretrained models?

Loss evolution for pre-trained models?

1

Hey great work! Just wondering any chance you would have the loss evolution for your pretrained models?

MetaVoice-1B: fix degradation compared to Python version

7

The MetaVoice-1B model has significant degradation compared to the Python version. I believe one of the main causes is using a 64x smaller decoder model (instead of multiband diffusion and...

torch.compile(fullgraph=True) support for flash-decoding

8

I was optimizing inference for a CPU overhead bound model and ran into problems compiling the model to a CUDA graph with flash decoding. I've tried to wrap the flash...