github gpustack/runner v0.1.32.post5

8 hours ago

Preserve platform-specific dependency metadata and improve Ascend prefill/decode KV transfer reliability. (#402)

Fixes

  • Preserve each CPU platform's measured dependencies in grouped backend and service listings.
  • Resolve Ascend Mooncake peer endpoints by absolute handshake port across data parallel groups and different KV port ranges.
  • Keep HIXL channel names distinct across peer restarts in the CANN vLLM image.

Ascend validation covered two Linux arm64 hosts with eight Ascend 910B2 devices each: raw KV transfers, one Prefill-only recovery with Decode and Router processes preserved, gateway/SSE requests, and ordinary inference.

Full Changelog: v0.1.32.post4...v0.1.32.post5

Don't miss a new runner release

NewReleases is sending notifications on new releases.