What's Changed
- Honour the left-padding mask in MiniCPM3 and Phixtral batches by @bhaskargurram-ai in #1925
- Fix incorrect assumption when parsing tool_calls[].function.arguments by @LxYuan0420 in #1904
- Tolerate a null rope_scaling config when constructing PhiMoE models by @gyanu2507 in #1889
- Update acknowledgments to include DeepSeek's DeepSeek v4.1 support by @Goekdeniz-Guelmez in #1936
- Fix incorrect masking axis in phi3small by @michalk8 in #1938
- Fix phimoe's sparsemixer by @michalk8 in #1939
- Fix: max_tokens: 0 -> Server accepts max_tokens: 0, generator then rejects it by @abhijithneilabraham in #1935
- Fix Qwen3 Coder argument parsing when the top-level type is missing by @zmylol in #1910
- Add global scale for nvfp4 quantized MoEs by @nastya236 in #1844
- Add tool call parser for MiniCPM5 by @dasashreeya in #1899
- Add prism_hadamard_qwen35 (Hadamard-rotated 2-bit Qwen3.5) by @bri-prism in #1941
- Fix GraniteMoeHybrid padded batch logprobs (#1908) by @aakashkardam in #1921
- Parallel head for qwen 3.5 by @nastya236 in #1960
- Fix long generation crash with MiniMax M3 by @pcuenca in #1958
- Support unconverted models by @nastya236 in #1962
- Add
meta-models/Muse-Spark-Open-816B-A42B-NVFP4by @nastya236 @michalk8 in #1976
New Contributors
- @bhaskargurram-ai made their first contribution in #1925
- @abhijithneilabraham made their first contribution in #1935
- @zmylol made their first contribution in #1910
- @bri-prism made their first contribution in #1941
- @aakashkardam made their first contribution in #1921
Full Changelog: v0.32.0...v0.31.4