@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 6559d96796
Fork 0

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 11, 00:00 UTC

[ROCm][CI] Set TORCH_NCCL_BLOCKING_WAIT Distributed Tests On ROCm (#31259)

Signed-off-by: Micah Williamson <micah.williamson@amd.com>

Micah Williamson committed Dec 24, 2025 at 01:19 UTC 6559d96796da12709389488a36ff706ee1db43c6
1 file changed +9 -2
.buildkite/test-amd.yaml
+9 -2
@@ -222,6 +222,9 @@ steps:
222 - tests/v1/engine/test_engine_core_client.py
223 - tests/distributed/test_symm_mem_allreduce.py
224 commands:
225 + # Work around HIP bug tracked here: https://github.com/ROCm/hip/issues/3876
226 + # TODO: Remove when the bug is fixed in a future ROCm release
227 + - export TORCH_NCCL_BLOCKING_WAIT=1
228 # test with torchrun tp=2 and external_dp=2
229 - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
230 # test with torchrun tp=2 and pp=2
@@ -270,9 +273,10 @@ steps:
273 - vllm/v1/executor/uniproc_executor.py
274 - vllm/v1/worker/gpu_worker.py
275 commands:
273 - # https://github.com/NVIDIA/nccl/issues/1838
274 - #- export NCCL_CUMEM_HOST_ENABLE=0
276 # test with torchrun tp=2 and dp=4 with ep
277 + # Work around HIP bug tracked here: https://github.com/ROCm/hip/issues/3876
278 + # TODO: Remove when the bug is fixed in a future ROCm release
279 + - export TORCH_NCCL_BLOCKING_WAIT=1
280 - torchrun --nproc-per-node=8 ../examples/offline_inference/torchrun_dp_example.py --tp-size=2 --pp-size=1 --dp-size=4 --enable-ep
281
282 - label: EPLB Algorithm Test # 5min
@@ -1294,6 +1298,9 @@ steps:
1298 - tests/v1/shutdown
1299 - tests/v1/worker/test_worker_memory_snapshot.py
1300 commands:
1301 + # Work around HIP bug tracked here: https://github.com/ROCm/hip/issues/3876
1302 + # TODO: Remove when the bug is fixed in a future ROCm release
1303 + - export TORCH_NCCL_BLOCKING_WAIT=1
1304 - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py
1305 - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
1306 - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py