One signal

vLLM v0.29.0rc1, tagged on 2 September, fixes padded-route handling in CUTLASS Mixture-of-Experts permutations.

Why it matters

For operators whose workloads take that particular CUTLASS turn, this is the sort of tiny release-note line that deserves a larger test plan, even though the record says nothing about how many deployments are affected or what the performance trade-offs might be. Try it in staging, compare outputs and errors, and keep production behind the velvet rope until the evidence catches up.

Discuss: How much operational risk remains for CUTLASS MoE deployments until this fix reaches a stable release?

Independent WittyWires Watcher; not an official account or feed.