A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 4, 13:45 UTC
[Bugfix] Exclude numa_bind fields from ParallelConfig DP hash (#41098)
Signed-off-by: yasong <yasong.wang@inferact.ai>
Roy Wang committed
Apr 28, 2026 at 15:52 UTC
9e92de51c61a47e5abb32d99b1930862473741d5
1 file changed
+8
vllm/config/parallel.py
+8
@@ -713,6 +713,14 @@ class ParallelConfig:
713
"worker_extension_cls",
714
"_api_process_count",
715
"_api_process_rank",
716
+ # NUMA binding is per-rank host-side memory locality; it does
717
+ # not affect collective-communication semantics. When numa_bind
718
+ # is enabled with auto-detection, each DP rank stores its own
719
+ # NUMA node in numa_bind_nodes (see vllm/utils/numa_utils.py
720
+ # `_get_numa_node`), which would otherwise diverge the DP hash.
721
+ "numa_bind",
722
+ "numa_bind_nodes",
723
+ "numa_bind_cpus",
724
}
725
726
from vllm.config.utils import get_hash_factors, hash_factors