What's Changed
- Fix MiniCPM vision dtype with quantized embeddings by @shiftedx in #2511
- Simplify MiniCPM-V 4.7 vision cache handling by @Blaizzy in #2514
- qwen3_5: read strided K/V in ragged decode attention instead of copying the KV cache every step by @griffinwork40 in #2510
- Fix Mistral Large 3 image feature merging by @Blaizzy in #2516
- Add Nemotron 3.5 Super VL with MTP and Bolt support by @Blaizzy in #2528
- Add live image resolution controls to the server by @Blaizzy in #2529
- Release prompt cache of finished SpeculativeGenerationBatch by @sady4850 in #2525
- Read PIL images as channels-last in the Gemma 4 image processors by @MohammadHijjawi97 in #2367
- sam3: use PyTorch channel grouping in the pixel decoder GroupNorm by @ju-ius in #2513
- Read PIL images as channels-last in the EmbeddingGemma 2 processors by @zviratko in #2489
- Add grpo for vlms by @Goekdeniz-Guelmez in #2504
- Fix remaining affine QMV precision mismatches on MLX 0.32.4 by @griffinwork40 in #2527
- Add support for Flux-3 image generation and editing by @lucasnewman in #2537
- Add MOSS-Image-SFT image generation model by @lucasnewman in #2541
- Honor checkpoint sampling defaults in CLI and generation APIs by @Blaizzy in #2545
- Fix rotating cache sinks and absolute position handling by @Blaizzy in #2548
- Bump version to 0.7.9 by @lucasnewman in #2555
New Contributors
- @shiftedx made their first contribution in #2511
- @griffinwork40 made their first contribution in #2510
- @sady4850 made their first contribution in #2525
- @MohammadHijjawi97 made their first contribution in #2367
- @ju-ius made their first contribution in #2513
- @zviratko made their first contribution in #2489
Full Changelog: v0.7.8...v0.7.9