What's Changed
- [Test] Pin the online hub path once for the GGUF integration tests by @abhijithneilabraham in #1026
- [Bugfix] Validate the diffusion stability threshold and name the refused parameter by @abhijithneilabraham in #1028
- DSpark: skip dense corrected-logits build in compiled draft by @begcdn in #1020
- [Perf] Compute the TurboQuant admission thresholds once per forward by @abhijithneilabraham in #1015
- [Contributing] Add shared agent contribution instructions by @WindChimeRan in #1039
- [Feature] Default the KV offload host pool to two max_model_len requests by @abhijithneilabraham in #1032
- [Bugfix] Count the compressor in the KV commit probe and bound its vm_stat read by @abhijithneilabraham in #1034
Full Changelog: v0.31.0.dev20261007013927...v0.31.0.dev20261007085756