SGLang PR 6651: Add PCIe device cross-NUMA node detection to prevent model output corruption
repo
Fields
id
sources/mgpu-sglang-6651-cross-numa-allreducePublisher
sgl-project/sglang
Kind
repo
Published
2025-05-27
Accessed
2026-10-03
notes
2x L20 (no NVLink) with nvidia-smi topo showing SYS between GPU0 and GPU1: TP=2 garbles model output via custom all-reduce; --disable-custom-all-reduce restores correct output.
Notes
2x L20 (no NVLink) with nvidia-smi topo showing SYS between GPU0 and GPU1: TP=2 garbles model output via custom all-reduce; --disable-custom-all-reduce restores correct output.
Cross-references
This record references no other records.