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
[ci/build] fix broken tests in LLM.collective_rpc (#15350)
Signed-off-by: youkaichao <youkaichao@gmail.com>
youkaichao committed
Mar 23, 2025 at 14:49 UTC
f68cce8e640f9f70340c17a3600ab7c50b8e2296
2 files changed
+3
-12
.buildkite/test-pipeline.yaml
+1
-1
@@ -515,7 +515,7 @@ steps:
515
- vllm/worker/model_runner.py
516
- entrypoints/llm/test_collective_rpc.py
517
commands:
518
- - pytest -v -s entrypoints/llm/test_collective_rpc.py
518
+ - VLLM_ENABLE_V1_MULTIPROCESSING=0 pytest -v -s entrypoints/llm/test_collective_rpc.py
519
- pytest -v -s ./compile/test_basic_correctness.py
520
- pytest -v -s ./compile/test_wrapper.py
521
- VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
tests/entrypoints/llm/test_collective_rpc.py
+2
-11
@@ -21,18 +21,9 @@ def test_collective_rpc(tp_size, backend):
21
def echo_rank(self):
22
return self.rank
23
24
- from vllm.worker.worker import Worker
25
-
26
- class MyWorker(Worker):
27
-
28
- def echo_rank(self):
29
- return self.rank
30
-
24
llm = LLM(model="meta-llama/Llama-3.2-1B-Instruct",
25
enforce_eager=True,
26
load_format="dummy",
27
tensor_parallel_size=tp_size,
35
- distributed_executor_backend=backend,
36
- worker_cls=MyWorker)
37
- for method in ["echo_rank", echo_rank]:
38
- assert llm.collective_rpc(method) == list(range(tp_size))
28
+ distributed_executor_backend=backend)
29
+ assert llm.collective_rpc(echo_rank) == list(range(tp_size))