Meta optimizes Muse Spark Open on TPU v7 with SGLang-TorchTPU

A megakernel unifies 62 decoder layers into one Pallas program, enabling pipelined weight loading, KV access, and cross-device transfers. Key optimizations include hybrid attention, NVFP4 expert weights, and enhanced KV caching per chip. On SemiAnalysis's AgentX benchmark with 1M context, throughput reaches 25,906 tokens/s per chip at p90 interactivity under 47 seconds.

Summarised from the original article by an open-source model, then checked automatically: every number and English name must appear in the source. The original article is authoritative.