Issue evidence · Serve · sgl-project/sglang

--enable-symm-mem: the 4 GiB symmetric-memory pool and NCCL buffers are not taken out of --mem-fraction-static

GitHub issue: https://github.com/sgl-project/sglang/issues/44106

Not run

With --enable-symm-mem, the 4 GiB pool and NCCL buffers are not accounted for in --mem-fraction-static, so a configuration that runs without the flag runs out of memory after start-up.

What Badgr ran

Nothing was run, as planned: the reproduction is GLM 5.3 Flash on 8x B200 (178 GiB each) with MTP EAGLE, and no large GPU allocation was made. Read the issue's configuration only.

What came back

The reporter's TP8 run at --mem-fraction-static 0.90 ran out of CUDA memory in the prefill_shapes warmup, while 0.88 passed, and TP4 crashed on the first verify broadcast.

The reporter says they did not run main with the failing settings, so whether main still has this is unconfirmed.

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

← All issue evidence