Support numactl bind for CPU and memory before process starts (#14156)

python/sglang/srt/managers/data_parallel_controller.py

@@ -49,6 +49,7 @@ from sglang.srt.tracing.trace import (
     trace_slice_end,
     trace_slice_start,
 )
+from sglang.srt.utils import numa_utils
 from sglang.srt.utils.common import (
     bind_port,
     configure_ipv6,
@@ -440,7 +441,9 @@ class DataParallelController:
                             writer,
                         ),
                     )
-                    with memory_saver_adapter.configure_subprocess():
+                    with memory_saver_adapter.configure_subprocess(), numa_utils.configure_subprocess(
+                        server_args, gpu_id
+                    ):
                         proc.start()
                 self.scheduler_procs.append(proc)
                 scheduler_pipe_readers.append(reader)

python/sglang/srt/managers/scheduler.py

@@ -2631,7 +2631,9 @@ def run_scheduler_process(
         set_gpu_proc_affinity(
             server_args.pp_size, server_args.tp_size, server_args.nnodes, gpu_id
         )
-    if (numa_node := server_args.numa_node) is not None:
+    if (
+        numa_node := server_args.numa_node
+    ) is not None and not envs.SGLANG_NUMA_BIND_V2.get():
         numa_bind_to_node(numa_node[gpu_id])
 
     # Set up tracing