SGLang Runs 816B MoE on a Single TPU v7 Host

SGLang, together with RadixArk and Google, ran the 816B Muse Spark Open MoE model on a single TPU v7 host via SGLang-TorchTPU, treating all 62 layers as one Pallas megakernel to reach 25.9k tok/s per chip and 433 tok/s per host, with cost beating GB200.

2026-11-11 ~ 2026-11-12 · 2 related posts

• PyTorch megakernel runs 816B MoE on one TPU v7 host at 433 tok/s, beats GB200 on cost — PyTorch · 2026-11-11
• SGLang hits 25.9K tok/s per chip on TPU v7, running 816B MoE as one megakernel — ying11231 · 2026-11-12