[Feature] NUMA binding support for GPU workers (#38635)

Signed-off-by: Shengqi Chen <harry-chen@outlook.com>
Co-authored-by: Jason Li <jasonlizhengjian@gmail.com>
Co-authored-by: Roger Wang <hey@rogerw.io>
This commit is contained in:
Shengqi Chen
2026-04-08 09:55:24 -07:00
committed by GitHub
co-authored by Jason Li Roger Wang
parent 512c5eb455
commit 75e01a39a1
13 changed files with 817 additions and 7 deletions
+13
View File
@@ -416,6 +416,9 @@ class EngineArgs:
nnodes: int = ParallelConfig.nnodes
node_rank: int = ParallelConfig.node_rank
distributed_timeout_seconds: int | None = ParallelConfig.distributed_timeout_seconds
numa_bind: bool = ParallelConfig.numa_bind
numa_bind_nodes: list[int] | None = ParallelConfig.numa_bind_nodes
numa_bind_cpus: list[str] | None = ParallelConfig.numa_bind_cpus
tensor_parallel_size: int = ParallelConfig.tensor_parallel_size
prefill_context_parallel_size: int = ParallelConfig.prefill_context_parallel_size
decode_context_parallel_size: int = ParallelConfig.decode_context_parallel_size
@@ -861,6 +864,13 @@ class EngineArgs:
"--distributed-timeout-seconds",
**parallel_kwargs["distributed_timeout_seconds"],
)
parallel_group.add_argument("--numa-bind", **parallel_kwargs["numa_bind"])
parallel_group.add_argument(
"--numa-bind-nodes", **parallel_kwargs["numa_bind_nodes"]
)
parallel_group.add_argument(
"--numa-bind-cpus", **parallel_kwargs["numa_bind_cpus"]
)
parallel_group.add_argument(
"--tensor-parallel-size", "-tp", **parallel_kwargs["tensor_parallel_size"]
)
@@ -1826,6 +1836,9 @@ class EngineArgs:
cp_kv_cache_interleave_size=self.cp_kv_cache_interleave_size,
_api_process_count=self._api_process_count,
_api_process_rank=self._api_process_rank,
numa_bind=self.numa_bind,
numa_bind_nodes=self.numa_bind_nodes,
numa_bind_cpus=self.numa_bind_cpus,
)
speculative_config = self.create_speculative_config(