Issue evidence · Serve · sgl-project/sglang

deepep_v2 crashes at startup on a cold DeepGEMM cache

GitHub issue: https://github.com/sgl-project/sglang/issues/44244

Not run

With --moe-a2a-backend deepep_v2 and an empty DeepGEMM JIT cache, non-first ranks time out in DeepEP kernels while rank 0 compiles.

What Badgr ran

Nothing was run: it needs a multi-GPU node with DeepEP v2 support, which did not fit the $3 cap on rentable GPUs.

What came back

  • The report attributes it to only the first rank compiling and DeepEP's default timeout.

A funded multi-GPU run would compare a cold and a warmed cache.

Checked 2026-10-14 in Badgr’s local development environment.

← All issue evidence