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 6, 20:30 UTC
[BUG] Fix async rlhf tests (#35811)
Signed-off-by: ahao-anyscale <ahao@anyscale.com>
Aaron Hao committed
Mar 11, 2026 at 15:06 UTC
d6b61e5166ac3eec7f828d0a102c30a76f6aecf3
2 files changed
+5
-1
.buildkite/test_areas/distributed.yaml
+1
-1
@@ -149,7 +149,7 @@ steps:
149
num_devices: 2
150
commands:
151
- pytest -v -s tests/distributed/test_context_parallel.py
152
- # - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/offline_inference/new_weight_syncing/rlhf_async_new_apis.py --- failing, need to re-enable
152
+ - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/offline_inference/new_weight_syncing/rlhf_async_new_apis.py
153
- VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
154
- pytest -v -s tests/v1/distributed/test_dbo.py
155
vllm/v1/worker/gpu_worker.py
+4
@@ -1006,6 +1006,10 @@ class Worker(WorkerBase):
1006
load_weights=load_weights_direct,
1007
)
1008
1009
+ # NCCL broadcast/packed path are asynchronous.
1010
+ # Sync here so the next step uses the new weights.
1011
+ torch.accelerator.synchronize()
1012
+
1013
def shutdown(self) -> None:
1014
# has_kv_transfer_group can be None during interpreter shutdown.
1015
if ensure_kv_transfer_shutdown is not None: