github vllm-project/vllm-metal v0.30.0.dev20260928205037

pre-release4 hours ago

What's Changed

  • [Multimodal] Require multimodal_backbone_mode and document bidirectional_layer_kinds by @Lavmee in #862
  • [Attention] Benchmark image-block attention: tiled kernel vs recompute by @Lavmee in #863
  • [Attention] Log the image-block attention path at startup by @Lavmee in #868
  • [Spec Decode] Check VLLM_METAL_SPEC_INGEST_CHUNK when the drafter is built by @Lavmee in #869
  • Load K-quant GGUF checkpoints by @ricky-chaoju in #870

Full Changelog: v0.30.0.dev20260928125641...v0.30.0.dev20260928205037

Don't miss a new vllm-metal release

NewReleases is sending notifications on new releases.