[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
+6 -2
View File
@@ -30,6 +30,7 @@ from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.tasks import POOLING_TASKS, SupportedTask
from vllm.tracing import instrument, maybe_init_worker_tracer
from vllm.transformers_utils.config import maybe_register_config_serialize_by_value
from vllm.utils import numa_utils
from vllm.utils.gc_utils import (
freeze_gc_heap,
maybe_attach_gc_debug_callback,
@@ -1055,6 +1056,8 @@ class EngineCoreProc(EngineCore):
set_process_title(process_title)
maybe_init_worker_tracer("vllm.engine_core", "engine_core", process_title)
decorate_logs()
if parallel_config.numa_bind:
numa_utils.log_current_affinity_state(process_title)
if data_parallel and vllm_config.kv_transfer_config is not None:
# modify the engine_id and append the local_dp_rank to it to ensure
@@ -1255,8 +1258,9 @@ class EngineCoreProc(EngineCore):
return
output = UtilityOutput(call_id)
# Lazily look-up utility method so that failure will be handled/returned.
get_result = lambda: (method := getattr(self, method_name)) and method(
*self._convert_msgspec_args(method, args)
get_result = lambda: (
(method := getattr(self, method_name))
and method(*self._convert_msgspec_args(method, args))
)
enqueue_output = lambda out: self.output_queue.put_nowait(
(client_idx, EngineCoreOutputs(utility_output=out))