Issue evidence · Serve · sgl-project/sglang

EP2 (tp=2 + ep=2) crashes with illegal memory access on Qwen3.8-Flash-Next-NVFP4

GitHub issue: https://github.com/sgl-project/sglang/issues/44291

Not run

A TP=2, EP=2 launch of Qwen3.8-Flash-Next-NVFP4 on RTX PRO 6000 crashes with an illegal memory access.

What Badgr ran

Nothing was run: it needs two RTX PRO 6000 Blackwell GPUs and the NVFP4 model, over the $5 cap.

What came back

  • The issue has no code blocks.

Not run.

Checked 2026-10-15 in Badgr’s local development environment.

← All issue evidence