SGLang issue 40723: Leaked sgl_shm_mm_* segments fill /dev/shm -> permanent crashloop
repo
Fields
id
sources/mgpu-sglang-40723-devshm-leakPublisher
sgl-project/sglang
Kind
repo
Published
2026-09-22
Accessed
2026-10-03
notes
Fleet report: ~700 H20 GPUs, TP4 pods, 4 -> 101 affected pods in ~30 min with no recovery without pod deletion. Contains the exact SIGBUS and NCCL ENOSPC strings.
Notes
Fleet report: ~700 H20 GPUs, TP4 pods, 4 -> 101 affected pods in ~30 min with no recovery without pod deletion. Contains the exact SIGBUS and NCCL ENOSPC strings.
Cross-references
This record references no other records.