Headline
- Web UI: Improved support for reasoning models, streamlined user interface, and tool-calling label
- Measure power during experiments in the dev CLI
- Fixes: resolve the NPU driver check, avoid attention sinks, fix uv support
What's Changed
- This PR adds platform markers for Windows-only dependencies (
wmi,human-eval-windows) so thatuv addresolves correctly on Linux. by @ianbmacdonald in #349 - Power profiling by @amd-pworfolk in #303
- Added tool label by @keetrap in #352
- block default Enter behavior fix #358 by @RobertAgee in #360
- Update LLAMA_VERSION_VULKAN to b6392 by @jeremyfowers in #359
- Feat: Unified reasoning stream handling, animated thought UI, llama.cpp bump, and reasoning format auto-switch by @kpoineal in #356
- Sync ROCm and Vulkan versions by @danielholanda in #364
- Add two smaller models to "hot" and add Qwen3-4B-Instruct-2507-GGUF by @jeremyfowers in #363
- Move model picker next to status by @RobertAgee in #326
- Enable context shift and avoid issues with attention sinks by @danielholanda in #365
- Update gpt-oss models to mxfp4 GGUFs by @jeremyfowers in #362
- Update driver check by @ramkrishna2910 in #369
New Contributors
- @ianbmacdonald made their first contribution in #349
- @kpoineal made their first contribution in #356
Full Changelog: v8.1.8...v8.1.9