Starstruk Logo MarkStarstruk

v1.1.0

snail007/goproxyv1.1.0Dec 10, 2025by p-e-w

AI Summary

This release significantly expands hardware compatibility and model support. Key additions include Apple Silicon (MPS) and multi-GPU support, alongside new model types like IBM Granite MoE and 'thinking' models. The update also focuses on robustness by improving quantization (MXFP4), dataset loading, and fixing data type loading issues.

Key Highlights

  • Added Apple Silicon (MPS) and multi-GPU support for enhanced performance.
  • Introduced support for IBM Granite MoE and 'thinking' models.
  • Implemented MXFP4 quantization support using Triton tensors.
  • Added notebook compatibility (Colab/Kaggle) and fixed float32 loading issues.
  • Improved refusal detection and local dataset loading capabilities.

New Features

  • Apple Silicon (MPS) support
  • Multi-GPU support
  • IBM Granite MoE support
  • MXFP4 quantized models with Triton tensors support
  • Support for models requiring `trust_remote_code`
  • Notebook (Colab/Kaggle) compatibility
  • Early stopping optimization support
  • Support for thinking models

Full Release Notes

## Changes

* @mbarnson added basic MPS (Apple Silicon) support in https://github.com/p-e-w/heretic/pull/5
* @red40maxxer reduced memory usage in https://github.com/p-e-w/heretic/pull/15
* @Ooooze added IBM Granite MoE support in https://github.com/p-e-w/heretic/pull/14
* @kldzj added multi-GPU support in https://github.com/p-e-w/heretic/pull/17 and https://github.com/p-e-w/heretic/pull/32
* @ricyoung fixed an error when Hugging Face user profile fields are missing in https://github.com/p-e-w/heretic/pull/20
* @tymat added support for MXFP4 quantized models with Triton tensors in https://github.com/p-e-w/heretic/pull/28
* @spikymoth improved support for loading local datasets in https://github.com/p-e-w/heretic/pull/33
* @kldzj added support for models that require `trust_remote_code` in https://github.com/p-e-w/heretic/pull/31
* @Vinayyyy7 added notebook (Colab/Kaggle) compatibility in https://github.com/p-e-w/heretic/pull/42
* @Vinayyyy7 fixed loading for certain models that default to the float32 dtype in https://github.com/p-e-w/heretic/pull/44
* @spikymoth improved refusal detection in https://github.com/p-e-w/heretic/pull/45
* @red40maxxer added a PR title lint to CI in https://github.com/p-e-w/heretic/pull/66
* @p-e-w added research features, support for stopping the optimization process early, and support for thinking models, and implemented an important padding fix suggested by @accemlcc

## New Contributors

* @mbarnson made their first contribution in https://github.com/p-e-w/heretic/pull/5
* @red40maxxer made their first contribution in https://github.com/p-e-w/heretic/pull/15
* @Ooooze made their first contribution in https://github.com/p-e-w/heretic/pull/14
* @kldzj made their first contribution in https://github.com/p-e-w/heretic/pull/17
* @ricyoung made their first contribution in https://github.com/p-e-w/heretic/pull/20
* @tymat made their first contribution in https://github.com/p-e-w/heretic/pull/28
* @spikymoth made their first contribution in https://github.com/p-e-w/heretic/pull/33
* @Vinayyyy7 made their first contribution in https://github.com/p-e-w/heretic/pull/42

**Full Changelog**: https://github.com/p-e-w/heretic/compare/v1.0.1...v1.1.0