Synchronize CPU ring workers before group teardown

Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
(cherry picked from commit 4ad71d8270)
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>

tests/diffusion/distributed/test_comm.py

@@ -14,6 +14,7 @@ from typing import Literal
 import pytest
 import torch
 
+from tests.helpers.distributed import configure_cpu_collective_worker, start_collective_workers
 from tests.helpers.mark import hardware_marks
 from vllm_omni.diffusion.distributed.comm import RingComm, SeqAllToAll4D, SeqAllToAll5D
 from vllm_omni.diffusion.distributed.parallel_state import (
@@ -56,6 +57,8 @@ def _init_worker(
             "MASTER_PORT": str(master_port),
         }
     )
+    if device_kind == "cpu":
+        configure_cpu_collective_worker()
     backend = "gloo" if device_kind == "cpu" else None
     init_distributed_environment(backend=backend)
 
@@ -227,6 +230,10 @@ def _run_ring_p2p(
             atol=atol,
             msg=f"[Rank {local_rank}] Ring P2P data mismatch",
         )
+        if device_kind == "cpu":
+            # P2P completion only synchronizes neighbors. Keep fast CPU ranks
+            # alive until every rank has finished constructing and using groups.
+            torch.distributed.barrier()
     finally:
         destroy_distributed_env()
 
@@ -243,7 +250,7 @@ def _spawn_4d_identity(
     device_kind: DeviceKind,
     master_port: int,
 ) -> None:
-    torch.multiprocessing.spawn(
+    start_collective_workers(
         _run_4d_identity,
         args=(
             world_size,
@@ -257,6 +264,7 @@ def _spawn_4d_identity(
             master_port,
         ),
         nprocs=world_size,
+        device_kind=device_kind,
     )
 
 
@@ -272,7 +280,7 @@ def _spawn_5d_identity(
     device_kind: DeviceKind,
     master_port: int,
 ) -> None:
-    torch.multiprocessing.spawn(
+    start_collective_workers(
         _run_5d_identity,
         args=(
             world_size,
@@ -286,6 +294,7 @@ def _spawn_5d_identity(
             master_port,
         ),
         nprocs=world_size,
+        device_kind=device_kind,
     )
 
 
@@ -299,10 +308,11 @@ def _spawn_ring_p2p(
     device_kind: DeviceKind,
     master_port: int,
 ) -> None:
-    torch.multiprocessing.spawn(
+    start_collective_workers(
         _run_ring_p2p,
         args=(world_size, dtype, batch_size, num_heads, head_size, device_kind, master_port),
         nprocs=world_size,
+        device_kind=device_kind,
     )
 
 
@@ -499,8 +509,9 @@ def _run_fused_qkv_matches_three_4d(
 @pytest.mark.cpu
 @pytest.mark.parametrize("world_size", [2, 4])
 def test_fused_qkv_all_to_all_matches_three_4d_exchanges_cpu(world_size: int) -> None:
-    torch.multiprocessing.spawn(
+    start_collective_workers(
         _run_fused_qkv_matches_three_4d,
         args=(world_size, "cpu", 29613),
         nprocs=world_size,
+        device_kind="cpu",
     )

tests/diffusion/distributed/test_pipeline_parallel.py

@@ -19,6 +19,7 @@ from vllm.utils.network_utils import get_file_store_init_method
 from vllm.v1.worker.gpu_worker import AsyncIntermediateTensors
 
 import vllm_omni.diffusion.distributed.pipeline_parallel as pp_module
+from tests.helpers.distributed import collective_context, start_collective_workers
 from tests.helpers.mark import hardware_marks
 from vllm_omni.diffusion.distributed.cfg_parallel import CFGParallelMixin
 from vllm_omni.diffusion.distributed.parallel_state import (
@@ -39,6 +40,25 @@ DeviceKind = Literal["cpu", "cuda"]
 _UNIT_MARKS = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
 
 
+@pytest.fixture(autouse=True)
+def _restore_collective_platforms():
+    """CPU baselines run in the parent process; retain its original platforms."""
+    import vllm.platforms
+
+    from vllm_omni.diffusion.distributed import comm, group_coordinator, parallel_state
+
+    original_platform = vllm.platforms.current_platform
+    original_omni_platforms = {
+        module: module.current_omni_platform for module in (comm, group_coordinator, parallel_state)
+    }
+    try:
+        yield
+    finally:
+        vllm.platforms.current_platform = original_platform
+        for module, platform in original_omni_platforms.items():
+            module.current_omni_platform = platform
+
+
 def _cleanup_distributed() -> None:
     """Destroy omni distributed state left dangling after a test.
 
@@ -409,6 +429,10 @@ def init_dist(
     device = _worker_device(local_rank, device_kind)
     if device_kind == "cuda":
         current_omni_platform.set_device(device)
+    else:
+        from tests.helpers.distributed import configure_cpu_collective_worker
+
+        configure_cpu_collective_worker()
     backend = "gloo" if device_kind == "cpu" else None
     init_distributed_environment(
         world_size=world_size,
@@ -500,13 +524,14 @@ def isend_irecv_worker(
 
 
 def _run_isend_irecv(pp_size: int, device_kind: DeviceKind, init_method: str) -> None:
-    mp_context = torch.multiprocessing.get_context("spawn")
+    mp_context = collective_context(device_kind)
     manager = mp_context.Manager()
     q = manager.Queue()
-    torch.multiprocessing.spawn(
+    start_collective_workers(
         isend_irecv_worker,
         args=(pp_size, init_method, device_kind, q),
         nprocs=pp_size,
+        device_kind=device_kind,
     )
     results = {label: tensor for label, tensor in [q.get(), q.get()]}
     torch.testing.assert_close(
@@ -656,15 +681,16 @@ def _run_predict_noise(
 
     baseline_out = compute_single_gpu_baseline(test_config, dtype, do_true_cfg, device_kind)
 
-    mp_context = torch.multiprocessing.get_context("spawn")
+    mp_context = collective_context(device_kind)
     manager = mp_context.Manager()
     pp_q = manager.Queue()
 
     world_size = pp_size * cfg_size
-    torch.multiprocessing.spawn(
+    start_collective_workers(
         predict_noise_worker,
         args=(world_size, init_method, pp_size, cfg_size, do_true_cfg, dtype, test_config, device_kind, pp_q),
         nprocs=world_size,
+        device_kind=device_kind,
     )
 
     pp_out = pp_q.get()
@@ -900,15 +926,16 @@ def _run_scheduler_step(
 
     baseline = compute_scheduler_step_baseline(test_config, do_true_cfg, device_kind)
 
-    mp_context = torch.multiprocessing.get_context("spawn")
+    mp_context = collective_context(device_kind)
     manager = mp_context.Manager()
     q = manager.Queue()
 
     world_size = pp_size * cfg_size
-    torch.multiprocessing.spawn(
+    start_collective_workers(
         scheduler_step_worker,
         args=(world_size, init_method, pp_size, cfg_size, do_true_cfg, test_config, device_kind, q),
         nprocs=world_size,
+        device_kind=device_kind,
     )
 
     result = q.get()

tests/helpers/cpu_collective_server.py

@@ -0,0 +1,19 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Import preload for the CPU-only collective forkserver.
+
+This module is imported in the forkserver process, before it starts workers.
+Its CUDA visibility does not alter the pytest process or CUDA parity workers.
+"""
+
+import os
+
+os.environ["CUDA_VISIBLE_DEVICES"] = ""
+
+import torch  # noqa: E402
+
+# Retain the expensive worker-module imports across CPU process groups.
+from tests.diffusion.distributed import test_comm, test_pipeline_parallel  # noqa: E402, F401
+
+if torch.cuda.is_initialized():
+    raise RuntimeError("CPU collective forkserver must not initialize CUDA")

tests/helpers/distributed.py

@@ -0,0 +1,62 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Process and platform setup for isolated collective test workers."""
+
+from collections.abc import Callable
+from multiprocessing.context import BaseContext
+
+import torch
+
+
+def collective_context(device_kind: str) -> BaseContext:
+    """Reuse imports in a clean CPU forkserver; CUDA workers use spawn."""
+    if device_kind == "cpu" and "forkserver" in torch.multiprocessing.get_all_start_methods():
+        torch.multiprocessing.set_forkserver_preload(["tests.helpers.cpu_collective_server"])
+        return torch.multiprocessing.get_context("forkserver")
+    return torch.multiprocessing.get_context("spawn")
+
+
+def start_collective_workers(
+    worker: Callable[..., None], *, args: tuple[object, ...], nprocs: int, device_kind: str
+) -> None:
+    context = collective_context(device_kind)
+    if context.get_start_method() == "spawn":
+        torch.multiprocessing.spawn(worker, args=args, nprocs=nprocs)
+    else:
+        torch.multiprocessing.start_processes(worker, args=args, nprocs=nprocs, start_method=context.get_start_method())
+
+
+def configure_cpu_collective_worker() -> None:
+    """Select CPU platforms inside a spawned worker that only uses Gloo.
+
+    Upstream coordinators resolve local ranks through the selected
+    platform's visible devices. CPU ranks can outnumber visible GPUs, so they
+    must use the CPU platform even when the test host has CUDA installed.
+    The worker still creates real process groups and runs real collectives.
+    """
+    import vllm.platforms
+    from vllm.platforms.cpu import CpuPlatform
+
+    from vllm_omni.diffusion.distributed import comm, group_coordinator, parallel_state
+    from vllm_omni.platforms.interface import UnspecifiedOmniPlatform
+
+    class CpuCollectivePlatform(UnspecifiedOmniPlatform):
+        dist_backend = "gloo"
+
+        @classmethod
+        def get_device_count(cls) -> int:
+            # CPU tensors share one logical device regardless of process rank.
+            return 1
+
+        @classmethod
+        def set_device(cls, device: torch.device) -> None:
+            assert device.type == "cpu"
+
+        @classmethod
+        def synchronize(cls) -> None:
+            pass
+
+    vllm.platforms.current_platform = CpuPlatform()
+    platform = CpuCollectivePlatform()
+    for module in (comm, group_coordinator, parallel_state):
+        module.current_omni_platform = platform

tests/helpers/runtime.py

@@ -659,14 +659,6 @@ class OmniRunner:
         self.seed = seed
         self._prompt_len_estimate_cache: dict[str, Any] = {}
         self.omni: Any = None
-        self._runner_process = psutil.Process()
-        self._initial_child_ids: set[tuple[int, float]] = set()
-        self._owned_engine_processes: dict[int, float] = {}
-        for child in self._runner_process.children(recursive=True):
-            try:
-                self._initial_child_ids.add((child.pid, child.create_time()))
-            except psutil.Error:
-                pass
         try:
             from vllm_omni.entrypoints.omni import Omni
 
@@ -678,14 +670,12 @@ class OmniRunner:
                 deploy_config=deploy_config,
                 **kwargs,
             )
-            self._remember_engine_processes(startup=True)
             startup_s = time.perf_counter() - startup_t0
             if log_stats:
                 print(f"OmniRunner startup took {startup_s:.3f}s (model={model_name})", flush=True)
         except BaseException:
             # ``with OmniRunner(...)`` never reaches ``__enter__``/``__exit__``
             # when construction fails after worker processes have started.
-            self._remember_engine_processes(startup=True)
             self.__exit__(None, None, None)
             raise
 
@@ -909,51 +899,18 @@ class OmniRunner:
     def stop_profile(self, stages: list[int] | None = None) -> list[Any]:
         return self.omni.stop_profile(stages=stages)
 
-    def _remember_engine_processes(self, *, startup: bool = False) -> None:
-        """Track startup workers, then only descendants of those workers.
-
-        Existing children may belong to another runner or a shared transcriber.
-        Record identities before close() can reparent surviving workers.
-        """
-        try:
-            if startup:
-                candidates = self._runner_process.children(recursive=True)
-            else:
-                candidates = []
-                for pid, created in list(self._owned_engine_processes.items()):
-                    try:
-                        proc = psutil.Process(pid)
-                        if proc.create_time() == created:
-                            candidates.extend(proc.children(recursive=True))
-                    except psutil.Error:
-                        pass
-            for proc in candidates:
-                try:
-                    identity = (proc.pid, proc.create_time())
-                    if identity in self._initial_child_ids:
-                        continue
-                    if startup and any(
-                        (parent.pid, parent.create_time()) in self._initial_child_ids for parent in proc.parents()
-                    ):
-                        continue
-                    cmdline = " ".join(proc.cmdline()).lower()
-                    if "enginecore" in cmdline or "enginecore" in proc.name().lower():
-                        self._owned_engine_processes[identity[0]] = identity[1]
-                except psutil.Error:
-                    pass
-        except psutil.Error as error:
-            print(f"Warning: could not inspect runner workers: {error}")
-
     def _cleanup_process(self):
         try:
+            keywords = ["enginecore"]
             matched = []
-            for pid, created in self._owned_engine_processes.items():
+            for proc in psutil.process_iter(["pid", "name", "cmdline", "username"]):
                 try:
-                    proc = psutil.Process(pid)
-                    if proc.create_time() == created:
-                        print(f"Found owned vllm process: PID={proc.pid}")
+                    cmdline = " ".join(proc.cmdline()).lower() if proc.cmdline() else ""
+                    name = proc.name().lower()
+                    if any(k in cmdline for k in keywords) or any(k in name for k in keywords):
+                        print(f"Found vllm process: PID={proc.pid}, cmd={cmdline[:100]}")
                         matched.append(proc)
-                except psutil.Error:
+                except (psutil.NoSuchProcess, psutil.AccessDenied):
                     pass
             for proc in matched:
                 try:
@@ -981,14 +938,11 @@ class OmniRunner:
         return self
 
     def __exit__(self, exc_type, exc_val, exc_tb):
-        self._remember_engine_processes()
-        try:
-            omni = getattr(self, "omni", None)
-            if omni is not None and hasattr(omni, "close"):
-                omni.close()
-        finally:
-            self._cleanup_process()
-            cleanup_test_environment()
+        omni = getattr(self, "omni", None)
+        if omni is not None and hasattr(omni, "close"):
+            omni.close()
+        self._cleanup_process()
+        cleanup_test_environment()
 
 
 # ---------------------------------------------------------------------------

tests/helpers/tests/test_runtime_omni_runner.py

@@ -1,15 +1,10 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
-"""Tests for OmniRunner startup rollback and worker ownership."""
+"""Tests for OmniRunner startup rollback."""
 
 from __future__ import annotations
 
-import subprocess
-import sys
-from collections.abc import Generator
-
-import psutil
 import pytest
 
 from tests.helpers.runtime import OmniRunner
@@ -43,114 +38,3 @@ def test_runner_init_rolls_back_on_omni_startup_failure(monkeypatch: pytest.Monk
     assert cleaned.count("process") == 1
     # Once at the start of ``__init__``, once from the constructor rollback.
     assert cleaned.count("env") == 2
-
-
-@pytest.fixture
-def live_workers() -> Generator[list[subprocess.Popen[bytes]], None, None]:
-    workers: list[subprocess.Popen[bytes]] = []
-    try:
-        yield workers
-    finally:
-        for worker in workers:
-            if worker.poll() is None:
-                worker.kill()
-            worker.wait(timeout=10)
-            if worker.stdout is not None:
-                worker.stdout.close()
-
-
-def _start_engine_worker(
-    workers: list[subprocess.Popen[bytes]], *, ignore_terminate: bool = False
-) -> subprocess.Popen[bytes]:
-    script = "import signal, time; "
-    if ignore_terminate:
-        script += "signal.signal(signal.SIGTERM, signal.SIG_IGN); "
-    script += "print('ready', flush=True); time.sleep(90)"
-    worker = subprocess.Popen([sys.executable, "-c", script, "enginecore"], stdout=subprocess.PIPE)
-    workers.append(worker)
-    assert worker.stdout is not None
-    assert worker.stdout.readline() == b"ready\n"
-    return worker
-
-
-@pytest.mark.parametrize("close_raises", [False, True])
-def test_runner_cleanup_preserves_unrelated_engine_workers(
-    monkeypatch: pytest.MonkeyPatch,
-    live_workers: list[subprocess.Popen[bytes]],
-    close_raises: bool,
-) -> None:
-    existing = _start_engine_worker(live_workers)
-    owned: list[subprocess.Popen[bytes]] = []
-    late_unrelated: list[subprocess.Popen[bytes]] = []
-    monkeypatch.setattr("tests.helpers.runtime.cleanup_test_environment", lambda: None)
-
-    class _OwnedOmni:
-        def __init__(self, *args: object, **kwargs: object) -> None:
-            owned.append(_start_engine_worker(live_workers))
-
-        def close(self) -> None:
-            late_unrelated.append(_start_engine_worker(live_workers))
-            if close_raises:
-                raise RuntimeError("close failed")
-
-    monkeypatch.setattr("vllm_omni.entrypoints.omni.Omni", _OwnedOmni)
-    if close_raises:
-        with pytest.raises(RuntimeError, match="close failed"), OmniRunner("fake-model"):
-            pass
-    else:
-        with OmniRunner("fake-model"):
-            pass
-
-    assert owned[0].poll() is not None
-    assert existing.poll() is None
-    assert late_unrelated[0].poll() is None
-
-
-def test_runner_startup_failure_cleans_only_new_workers(
-    monkeypatch: pytest.MonkeyPatch, live_workers: list[subprocess.Popen[bytes]]
-) -> None:
-    unrelated = _start_engine_worker(live_workers)
-    owned: list[subprocess.Popen[bytes]] = []
-    monkeypatch.setattr("tests.helpers.runtime.cleanup_test_environment", lambda: None)
-
-    class _FailingOmni:
-        def __init__(self, *args: object, **kwargs: object) -> None:
-            owned.append(_start_engine_worker(live_workers))
-            raise RuntimeError("startup failed")
-
-    monkeypatch.setattr("vllm_omni.entrypoints.omni.Omni", _FailingOmni)
-    with pytest.raises(RuntimeError, match="startup failed"):
-        OmniRunner("fake-model")
-
-    assert owned[0].poll() is not None
-    assert unrelated.poll() is None
-
-
-def test_runner_cleanup_rejects_stale_process_identity(live_workers: list[subprocess.Popen[bytes]]) -> None:
-    unrelated = _start_engine_worker(live_workers)
-    runner = object.__new__(OmniRunner)
-    runner._owned_engine_processes = {unrelated.pid: psutil.Process(unrelated.pid).create_time() + 1}
-
-    runner._cleanup_process()
-
-    assert unrelated.poll() is None
-
-
-def test_runner_cleanup_kills_owned_worker_ignoring_terminate(
-    monkeypatch: pytest.MonkeyPatch, live_workers: list[subprocess.Popen[bytes]]
-) -> None:
-    owned: list[subprocess.Popen[bytes]] = []
-    monkeypatch.setattr("tests.helpers.runtime.cleanup_test_environment", lambda: None)
-
-    class _StubbornOmni:
-        def __init__(self, *args: object, **kwargs: object) -> None:
-            owned.append(_start_engine_worker(live_workers, ignore_terminate=True))
-
-        def close(self) -> None:
-            pass
-
-    monkeypatch.setattr("vllm_omni.entrypoints.omni.Omni", _StubbornOmni)
-    with OmniRunner("fake-model"):
-        assert owned[0].poll() is None
-
-    assert owned[0].poll() is not None