Story

In brief

Tenstorrent accelerators join vLLM as an out-of-tree platform plugin, driven by mesh-architecture choices: phase-based scheduling, single-process data parallelism on Galaxy, on-device sampling with host fallback, and...