What's Changed
- fix(server): clear MLX buffer cache after /v1/embeddings requests by @pujitha24 in #2665
- Keep the model config out of the Muse Glimmer processor config by @nicoloangileri in #2535
- Honor max tokens defaults from the model when not specified by @lucasnewman in #2685
- Bump version to 0.7.13 by @lucasnewman in #2698
- (feat) Add QAT to training by @Goekdeniz-Guelmez in #2590
- Add MiniMax-H3-Image model by @lucasnewman in #2702
- Add Ling 3.1 Flash support by @Blaizzy in #2696
- Add native Ripple-65B-Preview and DFlash2 support by @Blaizzy in #2697
- Skip the deepstack eval in qwen3_vl and qwen3_vl_moe while training by @ahuzmeza in #2705
- Avoid shared-GPU reference stalls in distributed tests by @Blaizzy in #2706
- Bump version to 0.8.0 by @Blaizzy in #2707
- Add quantized model support for MiniMax-H3-Image by @lucasnewman in #2703
- docs: ask contributors to sign commits in CONTRIBUTING.md by @e4c5 in #2693
- Use eps=1e-6 in the RF-DETR projector LayerNorms by @ishaanzee in #2709
- Expose DeepSeek sharding and fix Mistral processor loading by @Blaizzy in #2717
- Gemma 4: read per_layer_config for the full-attention shapes by @jjbuck in #2719
- Support official Red Hat and OpenBMB DSpark checkpoints by @Blaizzy in #2676
- Add native Mellum MTP and Qwen text-only MTP aliases by @Blaizzy in #2677
- Add native DeepSeek V4.1 DSpark support by @Blaizzy in #2678
- [Server] Add distributed streaming, resumable serving, and cluster management by @Blaizzy in #2712
- [Server] Preserve RDMA links when importing configured workers by @Blaizzy in #2730
- feat(server): expose native model loading progress by @Blaizzy in #2732
- fix(mistral): normalize native Large 4 checkpoint metadata by @Blaizzy in #2736
- Add ERNIE 5.5 Flash Base with native ECHO decoding by @Blaizzy in #2708
- fix(deepseek): preserve mixed FP8/FP4 loading and dense vision by @Blaizzy in #2737
- Add experimental fast synchronization to distributed launches by @Blaizzy in #2741
- Add Inkling 2 by @Blaizzy in #2738
- Add speculative drafters to the cluster API by @Blaizzy in #2739
- fix(cluster): reuse verified checkpoints after reboot by @Blaizzy in #2744
- feat(server): serve startup loading progress over HTTP by @Blaizzy in #2735
- Fix the DPOP penalty placement by @abhijithneilabraham in #2726
- Keep XTC cutoffs separate for each batch row by @abhijithneilabraham in #2727
- Preserve DoRA input widths for 3-bit and 6-bit weights by @abhijithneilabraham in #2728
- Reject quantized embeddings when creating DoRA adapters by @abhijithneilabraham in #2729
- Resize RF-DETR masks with bilinear interpolation, as the reference does by @ishaanzee in #2713
- fix(tool-streaming): catch SyntaxError in ToolCallDeltaState on malformed arguments (#2635 follow-up) by @deokgoo in #2711
- Skip Gemma's embed_audio projector when quantizing by @jjbuck in #2715
New Contributors
- @nicoloangileri made their first contribution in #2535
- @e4c5 made their first contribution in #2693
- @jjbuck made their first contribution in #2719
- @deokgoo made their first contribution in #2711
Full Changelog: v0.7.12...v0.8.0