A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 18:30 UTC
Fix v1 supported oracle for worker-cls and worker-extension-cls (#15324)
Signed-off-by: youkaichao <youkaichao@gmail.com> Co-authored-by: youkaichao <youkaichao@gmail.com>
hijkzzz committed
Mar 23, 2025 at 10:23 UTC
0661cfef7a3bcb2c4ec6b03d6e070ce750150b3d
3 files changed
+7
-12
.buildkite/test-pipeline.yaml
+2
-2
@@ -148,8 +148,8 @@ steps:
148
# TODO: create a dedicated test section for multi-GPU example tests
149
# when we have multiple distributed example tests
150
- pushd ../examples/offline_inference
151
- - python3 rlhf.py
152
- - RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
151
+ - VLLM_ENABLE_V1_MULTIPROCESSING=0 python3 rlhf.py
152
+ - VLLM_ENABLE_V1_MULTIPROCESSING=0 RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
153
- popd
154
155
- label: Metrics, Tracing Test # 10min
vllm/engine/arg_utils.py
-10
@@ -1459,16 +1459,6 @@ class EngineArgs:
1459
recommend_to_remove=False)
1460
return False
1461
1462
- if self.worker_cls != EngineArgs.worker_cls:
1463
- _raise_or_fallback(feature_name="--worker-cls",
1464
- recommend_to_remove=False)
1465
- return False
1466
-
1467
- if self.worker_extension_cls != EngineArgs.worker_extension_cls:
1468
- _raise_or_fallback(feature_name="--worker-extension-cls",
1469
- recommend_to_remove=False)
1470
- return False
1471
-
1462
if self.num_scheduler_steps != EngineArgs.num_scheduler_steps:
1463
_raise_or_fallback(feature_name="--num-scheduler-steps",
1464
recommend_to_remove=True)
vllm/utils.py
+5
@@ -2170,6 +2170,11 @@ def _maybe_force_spawn():
2170
if cuda_is_initialized():
2171
reason = "CUDA is initialized"
2172
elif is_in_ray_actor():
2173
+ # even if we choose to spawn, we need to pass the ray address
2174
+ # to the subprocess so that it knows how to connect to the ray cluster.
2175
+ # env vars are inherited by subprocesses, even if we use spawn.
2176
+ import ray
2177
+ os.environ["RAY_ADDRESS"] = ray.get_runtime_context().gcs_address
2178
reason = "In a Ray actor and can only be spawned"
2179
2180
if reason is not None: