SGLang issue 36943: --enable-symm-mem all TP schedulers deadlock in ncclCommWindowRegister
repo
Fields
id
sources/mgpu-sglang-36943-symm-mem-deadlockPublisher
sgl-project/sglang
Kind
repo
Published
2026-08-29
Accessed
2026-10-03
notes
15 watchdog kills across 10 engines in two days. Full py-spy + gdb stacks show a blocking timeout-less recvmsg; /health stays 200 and GPU util is 0% throughout.
Notes
15 watchdog kills across 10 engines in two days. Full py-spy + gdb stacks show a blocking timeout-less recvmsg; /health stays 200 and GPU util is 0% throughout.
Cross-references
This record references no other records.