[Bugfix] Share PCP-aware local GPU indexing for NUMA binding (#61237)

This commit is contained in:
Sherif Waly
2026-10-14 10:54:03 +02:00
committed by GitHub
parent 937d6c19df
commit be4ce2e505
6 changed files with 134 additions and 85 deletions
+1 -19
View File
@@ -444,25 +444,7 @@ class Worker(WorkerBase):
# This env var set by Ray causes exceptions with graph building.
os.environ.pop("NCCL_ASYNC_ERROR_HANDLING", None)
parallel_config = self.parallel_config
if (
parallel_config.distributed_executor_backend
not in ("ray", "external_launcher")
and parallel_config.data_parallel_backend != "ray"
and parallel_config.nnodes_within_dp == 1
):
# Use local DP rank if available, otherwise use global DP rank.
dp_local_rank = self.parallel_config.data_parallel_rank_local
if dp_local_rank is None:
dp_local_rank = self.parallel_config.data_parallel_index
tp_pcp_pp_world_size = (
self.parallel_config.pipeline_parallel_size
* self.parallel_config.prefill_context_parallel_size
* self.parallel_config.tensor_parallel_size
)
# DP_LOCAL_RANK * TP_PCP_PP_WORLD_SIZE + TP_LOCAL_RANK
self.local_rank += dp_local_rank * tp_pcp_pp_world_size
self.local_rank = parallel_config.get_local_gpu_index(self.local_rank)
# Publish the logical-to-physical mapping for topology queries
# such as NIC affinity and P2P checks.