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 13, 17:00 UTC
[Test] Add non-MoE DP test coverage (#28235)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Matthew Bonanni committed
Nov 6, 2025 at 15:59 UTC
ca90f503041f17ace06a288f06ebfc455402e3bc
1 file changed
+19
-10
tests/v1/distributed/test_async_llm_dp.py
+19
-10
@@ -20,13 +20,6 @@ from vllm.v1.metrics.stats import IterationStats, MultiModalCacheStats, Schedule
20
21
DP_SIZE = int(os.getenv("DP_SIZE", 2))
22
23
-engine_args = AsyncEngineArgs(
24
- model="ibm-research/PowerMoE-3b",
25
- enforce_eager=True,
26
- tensor_parallel_size=int(os.getenv("TP_SIZE", 1)),
27
- data_parallel_size=DP_SIZE,
28
-)
29
-
23
24
async def generate(
25
engine: AsyncLLM,
@@ -65,6 +58,13 @@ async def generate(
58
return count, request_id
59
60
61
+@pytest.mark.parametrize(
62
+ "model",
63
+ [
64
+ "ibm-research/PowerMoE-3b",
65
+ "hmellor/tiny-random-LlamaForCausalLM",
66
+ ],
67
+)
68
@pytest.mark.parametrize(
69
"output_kind",
70
[
@@ -76,7 +76,10 @@ async def generate(
76
@pytest.mark.parametrize("async_scheduling", [True, False])
77
@pytest.mark.asyncio
78
async def test_load(
79
- output_kind: RequestOutputKind, data_parallel_backend: str, async_scheduling: bool
79
+ model: str,
80
+ output_kind: RequestOutputKind,
81
+ data_parallel_backend: str,
82
+ async_scheduling: bool,
83
):
84
if async_scheduling and data_parallel_backend == "ray":
85
# TODO(NickLucche) Re-enable when async scheduling is supported
@@ -107,8 +110,14 @@ async def test_load(
110
with ExitStack() as after:
111
prompt = "This is a test of data parallel"
112
110
- engine_args.data_parallel_backend = data_parallel_backend
111
- engine_args.async_scheduling = async_scheduling
113
+ engine_args = AsyncEngineArgs(
114
+ model=model,
115
+ enforce_eager=True,
116
+ tensor_parallel_size=int(os.getenv("TP_SIZE", 1)),
117
+ data_parallel_size=DP_SIZE,
118
+ data_parallel_backend=data_parallel_backend,
119
+ async_scheduling=async_scheduling,
120
+ )
121
engine = AsyncLLM.from_engine_args(
122
engine_args, stat_loggers=[SimpleStatsLogger]
123
)