What's Changed
- Support Laguna S by @Blaizzy in #1650
- Add Inkling from Thinky support by @Lazarus-931 in #1637
- Improve generation API type hints by @Blaizzy in #1599
- fix(qwen3_omni_moe): treat audio feature mask as sample-domain (mel-frame lengths) by @Lazarus-931 in #1657
- implement apc as a pluggable, capability driven cache adapter for #1629 by @Lazarus-931 in #1638
- Fix compile validation by @Lazarus-931 in #1655
- Add top-nσ sampler by @Lazarus-931 in #1652
- chore(deps): make training-only datasets an optional [train] extra by @Lazarus-931 in #1659
- add glm4_moe_lite_mtp drafter (GLM-4.7-Flash MTP split tool) by @samithaj in #1570
- Bump version to 0.6.7 by @Blaizzy in #1661
- fix(qwen3_omni_moe): accept batched (1, mel, T) audio features in encoder by @Lazarus-931 in #1656
- feat(sampler): add p-less sampling by @Lazarus-931 in #1653
- docs: document the ui extra required by the Gradio chat UI by @natejswenson in #1646
- fix(lfm2_vl): make the training forward pass runnable for LoRA/SFT by @chintan-projects in #1631
- Fix Qwen3 Omni deepstack: visual embeds were computed but never injected by @hellopahe in #1635
- fix(qwen3-omni): keep mRoPE positions correct across prefill and decode by @Lazarus-931 in #1621
- Add decoder_input_ids warm start for DiffusionGemma by @Blaizzy in #1589
- Fix Laguna NVFP4 sanitize by @Blaizzy in #1664
- Add typical-p (locally typical) sampler by @Lazarus-931 in #1663
New Contributors
- @samithaj made their first contribution in #1570
- @natejswenson made their first contribution in #1646
- @chintan-projects made their first contribution in #1631
Full Changelog: v0.6.6...v0.6.7