[Bugfix] Restore multimodal support on the plain "vllm" throughput backend (#52168)

Signed-off-by: Michal Ganczarenko <michal.ganczarenko@intel.com>
This commit is contained in:
Michał Ganczarenko
2026-08-28 10:28:25 +00:00
committed by GitHub
parent 5f213ed159
commit d1922cb5a7
2 changed files with 50 additions and 4 deletions
+47
View File
@@ -230,3 +230,50 @@ def test_get_requests_resolves_mmvu(
requests = get_requests(args, hf_tokenizer)
assert len(requests) == 3
assert all(isinstance(r, SampleRequest) for r in requests)
@pytest.mark.benchmark
def test_get_requests_allows_multimodal_on_plain_vllm_backend(
monkeypatch: pytest.MonkeyPatch,
hf_tokenizer: PreTrainedTokenizerBase,
) -> None:
"""Regression: --backend vllm must accept multimodal datasets such as ASR.
_run_vllm_requests forwards multi_modal_data straight to LLM.generate, so
the vllm-chat-only gate in get_samples rejected a backend that always
worked.
"""
import vllm.benchmarks.datasets.datasets as dsmod
class _StubASRDataset:
IS_MULTIMODAL = True
SUPPORTED_DATASET_PATHS = {"openslr/librispeech_asr"}
def __init__(self, **kwargs):
self.__dict__.update(kwargs)
def sample(self, *, num_requests, tokenizer, **kwargs): # noqa: ARG002
return [
SampleRequest(prompt="q", prompt_len=1, expected_output_len=1)
for _ in range(num_requests)
]
monkeypatch.setattr(dsmod, "ASRDataset", _StubASRDataset)
parser = FlexibleArgumentParser()
add_cli_args(parser)
args = parser.parse_args(
[
"--dataset-name",
"hf",
"--dataset-path",
"openslr/librispeech_asr",
"--backend",
"vllm",
"--num-prompts",
"3",
]
)
requests = get_requests(args, hf_tokenizer)
assert len(requests) == 3
assert all(isinstance(r, SampleRequest) for r in requests)
+3 -4
View File
@@ -567,10 +567,9 @@ def _to_serve_args(args: argparse.Namespace) -> argparse.Namespace:
def get_requests(args, tokenizer):
serve_args = _to_serve_args(args)
# Throughput reuses bench serve's dataset dispatch. vllm-chat is the only
# multimodal-capable throughput backend, so it is the only one allowed past
# the multimodal gates inside get_samples; other backends raise there.
mm_backends = ("vllm-chat",) if args.backend == "vllm-chat" else ()
# Both offline vLLM backends forward multi_modal_data to the engine;
# "hf" and "mii" are text-only and get rejected inside get_samples.
mm_backends = ("vllm", "vllm-chat")
requests = get_samples(serve_args, tokenizer, multimodal_backends=mm_backends)
requests = assign_loras(requests, args)
requests = filter_requests_for_dp(requests, args.data_parallel_size)