●Change8

v0.17.8-rc2

📦 ollamaView on GitHub →
✨ 2 features🐛 7 fixes🔧 8 symbols

Summary

This release focuses on stability and performance improvements, including fixes for GLM tool parsing, localhost handling, and updates to MLX and ROCm support. It also refactors the MLX runner sampler interface.

Migration Steps

  1. If you rely on the previous sampler interface chain in x/mlxrunner, update your code to use the new single stateful Sampler.

✨ New Features

  • mlx: Implemented int4 groupsize 64 support.
  • mlxrunner: Replaced sampler interface chain with a single stateful Sampler.

🐛 Bug Fixes

  • Fixed unclosed arg_value tags in GLM tool calls during parsing.
  • Reapplied fix preventing the requirement to pull stubs for cloud models.
  • Fixed localhost handling during model creation.
  • Improved Docker build process with smarter parallelism.
  • Handled stream disconnects gracefully in the cloud proxy.
  • Fixed an issue where resetting to defaults disabled auto-update.
  • Fixed Windows build process.

Affected Symbols

GLM tool calls (parsers)cloud models (stubs)localhost handling (create)docker buildcloud_proxy (stream disconnects)x/mlxrunner (sampler interface)app (reset to defaults/auto-update)mlx (model creation parameters)