v0.22.0

What's Changed

• feat: add native Metal support by @AlpinDale in #1668
• chore: optimize metal backend performance by @AlpinDale in #1669
• perf: optimize GDN performance on Metal by @AlpinDale in #1670
• fix: logging messages in metal by @AlpinDale in #1671
• fix: gemma-4 with exllamav3 by @AlpinDale in #1672
• feat: add dflash support for gemma-4 by @AlpinDale in #1673
• sync: update to upstream 13c49f9845 by @AlpinDale in #1678
• [sync] vLLM 2026-07-04 — 13 commit(s) (13c49f984..d0a258477) by @AlpinDale in #1679
• docs: rebrand aphrodite to sonar by @AlpinDale in #1681
• [sync] vLLM 2026-07-04 — 2 commit(s) (d0a258477..d63c8e944) (stopped early at d63c8e944) by @AlpinDale in #1680
• [sync] vLLM 2026-07-04 — 63 commit(s) (d63c8e944..4a6bf3c77) by @AlpinDale in #1682
• fix: restore APHRODITE_USE_PRECOMPILED by @AlpinDale in #1683
• feat: add DeepSeek Sparse Attention kernels for sm89 (ada) by @AlpinDale in #1684
• chore: pre-commit baseline by @AlpinDale in #1685
• [sync] vLLM 2026-07-06 — 91b564730..9fde043f5 by @AlpinDale in #1688
• [sync] vLLM 2026-07-07 — 40cc2e832..d6875196a by @AlpinDale in #1689
• [sync] vLLM 2026-07-08 — d891b9bd5..f7fc0ca99 by @AlpinDale in #1690
• [sync] vLLM 2026-07-08 — 567a78432..f7fc0ca99 by @AlpinDale in #1691
• [sync] vLLM 2026-07-08 — 5e975eae1..5d5fab006 by @AlpinDale in #1692
• [sync] vLLM 2026-07-08 — c0e8e1f12..a5d19cbb9 by @AlpinDale in #1693
• fix(sampler): avoid IndexError on empty output_token_ids subset by @AlpinDale in #1695
• [sync] vLLM 2026-07-09 — 089e41287..2c17d33f4 by @AlpinDale in #1694
• test(moe): scale modelopt_fp4 tolerance by k at all sizes by @AlpinDale in #1696
• fix(quantization): break flashinfer_utils circular import via lazy MoEActivation by @AlpinDale in #1697
• style: apply pre-commit fixes (ruff format/lint + spdx headers) by @AlpinDale in #1698
• [sync] vLLM 2026-07-09 — b8c7c8653..429f40574 by @AlpinDale in #1699
• [sync] vLLM 2026-07-09 — a07765c6b..b0dec2a11 by @AlpinDale in #1701
• [sync] vLLM 2026-07-09 — 2ded1b24e..ea0fa34f4 by @AlpinDale in #1702
• feat: build qutlass fp4 kernels for SM110 (Thor) via the SM100 path by @AlpinDale in #1703
• fix: force 32 MiB cuBLAS workspace on Thor (SM 11.0) by @AlpinDale in #1704
• fix: point test assets back to the vllm-public-assets S3 bucket by @AlpinDale in #1705
• feat: bump torch to 2.13.0 (torchvision 0.28.0, triton 3.7.1) by @AlpinDale in #1706
• feat: swordfish kernels for Blackwell by @AlpinDale in #1707
• [sync] vLLM 2026-07-12 — 766469a4c..83762b77b by @AlpinDale in #1708
• [sync] vLLM 2026-07-12 — 5f8e73cb8..4c81772e8 by @AlpinDale in #1709
• [sync] vLLM 2026-07-15 — e26264f3e..647213129 by @AlpinDale in #1710
• fix: correct shard_ids for GDN in_proj_qkv under LoRA (Qwen3.5/3.6) by @Nakanokensetsu in #1711
• fix: recognize EXL3 base layers in LoRA device detection by @Nakanokensetsu in #1712
• [sync] vLLM 2026-07-18 — 3b39fd284..109b736b8 by @AlpinDale in #1713

New Contributors

• @Nakanokensetsu made their first contribution in #1711

Full Changelog: v0.21.0...v0.22.0