What's Changed
- adding kolibri1 support by @Lazarus-931 in #2423
- Add EmbeddingGemma 2 multimodal embedding support by @Blaizzy in #2446
- qwen3_5: use mlx-lm's packed gated delta kernel for prefill by @wyanzhao in #2442
- Use fused SDPA in the RF-DETR backbone attention by @ishaanzee in #2443
- Compute grid_sample in float32 for half-precision inputs by @ishaanzee in #2454
- Fix RF-DETR small variant config by @ishaanzee in #2453
- Make RF-DETR sanitize idempotent by @ishaanzee in #2452
- Fix off-by-one RF-DETR backbone feature layers by @ishaanzee in #2450
- Fix tracker RoPE axis order to match Meta checkpoints by @atesahmet0 in #2435
- Add support for Clef decision model by @Lazarus-931 in #2459
- Support Xing4.0 model by @Lazarus-931 in #2465
- feat: add LiquidAI d1-3B and d1-omni-600M decision models by @ykhrustalev in #2467
- inkling: use the Metal kernels only when Metal is available by @joshuaswarren in #2462
- Use the separable interpolation kernel for the RF-DETR segmentation u… by @ishaanzee in #2468
- Fix/2426 expert activation by @Lazarus-931 in #2449
- moe_offload: write a weight index describing the repacked output by @Lazarus-931 in #2427
- Migrate to httpx2, fix CI by @lucasnewman in #2487
- fix: recognize LFM2's down projection in mixed quant recipes by @ykhrustalev in #2497
- Preserve NVFP4 global scales by @Blaizzy in #2437
- Load bundled drafters directly in CLI and server by @Blaizzy in #2457
- Stabilize tool-schema key order for prefix caching by @Blaizzy in #2498
- Add MiniCPM-V 4.7 1B and 35B-A3B with a custom processor by @Blaizzy in #2499
- Reformat trainer for all modalities by @Goekdeniz-Guelmez in #1939
- Fix Muse Glimmer ignoring explicit RoPE scaling settings by @spector-in-london in #2494
- Fix cancellation during multi-row prefill by @Blaizzy in #2273
- Fix compaction when max output fills the context window by @Blaizzy in #2458
- Stream tool arguments in Chat Completions and Responses by @Blaizzy in #2502
- Allow compaction between completed tool groups by @Blaizzy in #2501
New Contributors
- @wyanzhao made their first contribution in #2442
- @ishaanzee made their first contribution in #2443
- @atesahmet0 made their first contribution in #2435
- @joshuaswarren made their first contribution in #2462
- @spector-in-london made their first contribution in #2494
Full Changelog: v0.7.6...v0.7.8