[Core] Drop the Ray driver-node GPU pin left over from the legacy executor (#61713)
Signed-off-by: Woosuk Kwon <woosuk@inferact.ai> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5.5
parent
3f1e57225a
commit
4a89316efa
@@ -1,79 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""Make sure ray assigns GPU workers to the correct node.
|
||||
|
||||
Run:
|
||||
```sh
|
||||
cd $VLLM_PATH/tests
|
||||
|
||||
pytest distributed/test_multi_node_assignment.py
|
||||
```
|
||||
"""
|
||||
|
||||
import os
|
||||
import time
|
||||
|
||||
import pytest
|
||||
import ray
|
||||
from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy
|
||||
|
||||
from vllm import initialize_ray_cluster
|
||||
from vllm.config import ParallelConfig
|
||||
from vllm.utils.network_utils import get_ip
|
||||
from vllm.v1.executor.ray_utils import available_resources_per_node
|
||||
|
||||
VLLM_MULTI_NODE = os.getenv("VLLM_MULTI_NODE", "0") == "1"
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
not VLLM_MULTI_NODE, reason="Need at least 2 nodes to run the test."
|
||||
)
|
||||
def test_multi_node_assignment() -> None:
|
||||
# NOTE: important to keep this class definition here
|
||||
# to let ray use cloudpickle to serialize it.
|
||||
class Actor:
|
||||
def get_ip(self):
|
||||
return get_ip()
|
||||
|
||||
for _ in range(10):
|
||||
config = ParallelConfig(1, 2)
|
||||
initialize_ray_cluster(config)
|
||||
|
||||
current_ip = get_ip()
|
||||
workers = []
|
||||
for bundle_id, bundle in enumerate(config.placement_group.bundle_specs):
|
||||
if not bundle.get("GPU", 0):
|
||||
continue
|
||||
scheduling_strategy = PlacementGroupSchedulingStrategy(
|
||||
placement_group=config.placement_group,
|
||||
placement_group_capture_child_tasks=True,
|
||||
placement_group_bundle_index=bundle_id,
|
||||
)
|
||||
|
||||
worker = ray.remote(
|
||||
num_cpus=0,
|
||||
num_gpus=1,
|
||||
scheduling_strategy=scheduling_strategy,
|
||||
)(Actor).remote()
|
||||
worker_ip = ray.get(worker.get_ip.remote())
|
||||
assert worker_ip == current_ip
|
||||
workers.append(worker)
|
||||
|
||||
for worker in workers:
|
||||
ray.kill(worker)
|
||||
|
||||
ray.util.remove_placement_group(config.placement_group)
|
||||
|
||||
gpus_needed = len(workers)
|
||||
deadline = time.monotonic() + 60
|
||||
node_id = ray.get_runtime_context().get_node_id()
|
||||
while True:
|
||||
free_gpus = available_resources_per_node().get(node_id, {}).get("GPU", 0)
|
||||
if free_gpus >= gpus_needed:
|
||||
break
|
||||
|
||||
remaining = deadline - time.monotonic()
|
||||
if remaining <= 0:
|
||||
raise TimeoutError("placement group resources were not released")
|
||||
|
||||
time.sleep(min(1, remaining))
|
||||
Reference in New Issue
Block a user