Eric Buehler

Add the OpenVLA model

2

https://huggingface.co/openvla/openvla-7b

Currently, AnyMoE only support homogenous expert types. This restricts the user to using only fine-tuned or only LoRA adapter experts. Implementing heterogeneous expert support will enable, for example, mixing fine-tuned...

Fixes #2480.

[Tracking] FLUX T5 XXL model produces NaN when on CUDA and using F16

1

Perhaps we can use clamping, as per: > https://github.com/huggingface/transformers/blob/main/src/transformers/models/t5/modeling_t5.py#L748-L755 Using BF16 works on CUDA.

Currently, our Error type is already useful but could benefit from some of the QoL features which the anyhow crate has. This PR adds the following: - Make `Error::msg` more...

The Phi 3.5 MoE model is a ~42B parameter model with 16 experts, using 2 active. This PR implements the model and provides a simple inference example. Additionally, this PR...

Add Parler TTS

1

Model: parler-tts/parler-tts-mini-v1 TODO list: - [ ] ParlerForCausalGen - [x] T5 encoder-decoder - [ ] `ParlerTTSForCausalLM` decoder - [ ] `ParlerTTSDecoder` - [ ] `ParlerTTSSinusoidalPositionalEmbedding` - [ ] Other RoPE...