v0.6.7

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