[Fix] Spawn, don't fork, the benchmark server process (#34712)

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

python/sglang/benchmark/endpoint.py

@@ -50,14 +50,13 @@ def _launch_server_target(launch_server_func: Callable, server_args: ServerArgs)
 
 
 def launch_or_reuse_server(launch_server_func: Callable, server_args: ServerArgs):
-    # Resolve in the parent, before the fork. The pipeline probes the device
-    # (the default attention backend reads the CUDA capability), and a forked
-    # child cannot re-initialize CUDA once this process has.
+    # Resolving probes the device: the default attention backend reads the CUDA
+    # capability, and XPU reads mem_get_info. This process owns a live context after.
     server_args.resolve_once()
 
     base_url = resolve_base_url("", server_args.host, server_args.port)
 
-    # Reuse an already-running server instead of forking a second one onto the
+    # Reuse an already-running server instead of launching a second one onto the
     # occupied port, where it would orphan, compete for the GPU, and OOM.
     if server_is_up(base_url, timeout=5):
         print(
@@ -66,7 +65,9 @@ def launch_or_reuse_server(launch_server_func: Callable, server_args: ServerArgs
         )
         return None, base_url
 
-    proc = multiprocessing.Process(
+    # Spawn, not the platform default: a fork inherits the context resolve_once()
+    # initialized above, and CUDA/XPU cannot be re-initialized in a forked child.
+    proc = multiprocessing.get_context("spawn").Process(
         target=_launch_server_target,
         args=(
             launch_server_func,
@@ -99,7 +100,8 @@ class BenchEndpoint:
     """
 
     base_url: str
-    _proc: Optional[multiprocessing.Process] = None
+    # SpawnProcess is a sibling of multiprocessing.Process, not a subclass.
+    _proc: Optional[multiprocessing.process.BaseProcess] = None
 
     def close(self) -> None:
         if self._proc is not None:

test/registered/unit/bench/test_benchmark_endpoint.py

@@ -0,0 +1,53 @@
+"""Unit tests for sglang/benchmark/endpoint.py"""
+
+import multiprocessing
+import unittest
+from unittest import mock
+
+from sglang.benchmark.endpoint import launch_or_reuse_server
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_cpu_ci(est_time=10, suite="base-a-test-cpu")
+
+
+def _noop_launch_server(server_args):
+    """Module-level, not a lambda: spawn pickles ``target`` at Process.start()."""
+
+
+class _StubServerArgs:
+    """resolve_once() is where a real record probes the device; stubbing it is what
+    keeps this on CPU. Nothing but host/port is read before the child is created."""
+
+    host = "127.0.0.1"
+    port = 30000
+
+    def resolve_once(self):
+        pass
+
+
+class TestLaunchOrReuseServer(CustomTestCase):
+    def test_server_child_is_a_spawn_process(self):
+        """#34709: the child must be spawned -- a fork inherits the accelerator
+        context resolve_once() left in the parent and dies on its first device op.
+        """
+        with (
+            # Patch BaseProcess, not get_context: mocking the context away would
+            # erase the start method under test. Only the launch is suppressed.
+            mock.patch.object(multiprocessing.process.BaseProcess, "start"),
+            mock.patch.object(
+                multiprocessing.process.BaseProcess, "is_alive", return_value=True
+            ),
+            # Down for the reuse probe, up on the first startup poll; extend the
+            # list if the startup loop ever gains another probe.
+            mock.patch(
+                "sglang.benchmark.endpoint.server_is_up", side_effect=[False, True]
+            ),
+        ):
+            proc, _ = launch_or_reuse_server(_noop_launch_server, _StubServerArgs())
+
+        self.assertIsInstance(proc, multiprocessing.context.SpawnProcess)
+
+
+if __name__ == "__main__":
+    unittest.main()