github vllm-project/vllm-metal v0.31.0.dev20261007210003

pre-release3 hours ago

What's Changed

  • [Test] DSpark: pin compiled-path token-ID parity with corrected_logits=False by @begcdn in #1045
  • [Bugfix] Accept integral DSpark candidate limits by @mhdimo in #1049
  • [Test] Harden DFlash and DSpark configuration coverage by @mhdimo in #1048
  • [Bugfix] Validate DSpark Q4 dimensions before loading weights by @StevenWang-CY in #1050
  • [KV Offload] Cap the default host pool to fit the Metal budget by @RobbieJ in #1046
  • [Perf] Enable GQA decode for multi-request batches by @suntp in #991

Full Changelog: v0.31.0.dev20261007095912...v0.31.0.dev20261007210003

Don't miss a new vllm-metal release

NewReleases is sending notifications on new releases.