[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
+3 -35
View File
@@ -223,29 +223,6 @@ def _pct_sku_config() -> _PctSku | None:
return sku
def _get_gpu_index(
parallel_config, local_rank: int, dp_local_rank: int | None = None
) -> int:
"""Compute the physical GPU index used for NUMA lookup."""
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
):
if dp_local_rank is None:
dp_local_rank = parallel_config.data_parallel_rank_local
if dp_local_rank is None:
dp_local_rank = parallel_config.data_parallel_index
tp_pp_world_size = (
parallel_config.pipeline_parallel_size
* parallel_config.tensor_parallel_size
)
return local_rank + dp_local_rank * tp_pp_world_size
return local_rank
def _get_numa_node(parallel_config, gpu_index: int) -> int:
numa_nodes = parallel_config.numa_bind_nodes
if numa_nodes is None:
@@ -346,7 +323,7 @@ def _get_numactl_worker_args(
parallel_config, local_rank: int, dp_local_rank: int | None = None
) -> str:
"""Compute the numactl args for a single TP/PP worker subprocess."""
gpu_index = _get_gpu_index(parallel_config, local_rank, dp_local_rank)
gpu_index = parallel_config.get_local_gpu_index(local_rank, dp_local_rank)
numa_node = _get_numa_node(parallel_config, gpu_index)
cpu_binding = _get_cpu_binding(parallel_config, gpu_index, [numa_node])
@@ -384,17 +361,8 @@ def _get_enginecore_numa_nodes(
and parallel_config.data_parallel_backend != "ray"
and parallel_config.nnodes_within_dp == 1
):
if dp_local_rank is None:
dp_local_rank = parallel_config.data_parallel_rank_local
if dp_local_rank is None:
dp_local_rank = parallel_config.data_parallel_index
tp_pp_world_size = (
parallel_config.pipeline_parallel_size
* parallel_config.tensor_parallel_size
)
shard_start = dp_local_rank * tp_pp_world_size
shard_end = min(shard_start + tp_pp_world_size, len(numa_nodes))
shard_start = parallel_config.get_local_gpu_index(0, dp_local_rank)
shard_end = min(shard_start + parallel_config.world_size, len(numa_nodes))
shard_indices: range | tuple[int, ...] = range(shard_start, shard_end)
else:
shard_indices = range(len(numa_nodes))