[Perf][Multimodal] Sort multimodal encoder inputs by modality (#60196)

Signed-off-by: Jaebok Lee <jaebok9541@naver.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
zzaebok
2026-10-09 10:55:36 +08:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 295c062db8
commit 702ce313f4
5 changed files with 112 additions and 27 deletions
+16 -1
View File
@@ -78,7 +78,7 @@ def test_encoder_graph_unsupported_modality_uses_eager_output():
"vllm.v1.worker.gpu.mm.encoder_runner.group_and_batch_mm_kwargs",
return_value=[("audio", 1, {"audio_values": torch.zeros(2)})],
):
result = runner.execute_mm_encoder([])
result = runner.execute_mm_encoder([("audio", MagicMock())])
assert len(result) == 1
assert result[0] is output
manager.execute.assert_not_called()
@@ -290,6 +290,21 @@ def test_execute_mm_encoder_caches_outputs_without_gathering():
state.encoder_runner.gather_mm_embeddings.assert_not_called()
def test_execute_mm_encoder_sorts_by_modality():
"""Items are encoded sorted by modality; outputs keep the input order."""
mm_kwargs = [("video", 0), ("image", 1), ("video", 2), ("image", 3)]
runner = _make_runner([], [])
runner.model = MagicMock()
runner.model.embed_multimodal.side_effect = lambda x: [torch.full((1, 1), x)]
with patch(
"vllm.v1.worker.gpu.mm.encoder_runner.group_and_batch_mm_kwargs",
side_effect=lambda items, **_: [(m, 1, {"x": x}) for m, x in items],
) as group:
outputs = runner.execute_mm_encoder(mm_kwargs)
assert group.call_args.args[0] == [mm_kwargs[i] for i in (1, 3, 0, 2)]
assert [int(o) for o in outputs] == [0, 1, 2, 3]
def test_execute_mm_encoder_is_a_noop_without_scheduled_items():
"""A step that schedules no encoder input must not touch the encoder."""
cache = EncoderCache()
+64
View File
@@ -624,6 +624,70 @@ def test_set_active_mm_loras_builds_tower_and_connector_mappings():
assert connector_mapping.index_mapping == ((7,) * 14 + (7,) * 13 + (0,) * 12)
def test_set_active_mm_loras_follows_modality_sorted_encoder_order():
model = Mock()
model.get_mm_lora_token_counts.side_effect = (
lambda *, modality, mm_kwargs, num_mm_embeds: (num_mm_embeds, None)
)
lora_manager = Mock()
lora_manager.supports_tower_connector_lora.return_value = True
encoder_cache = EncoderCache()
for req_id, modality, length in [("req-a", "video", 1), ("req-b", "image", 2)]:
encoder_cache.mm_features[req_id] = [
MultiModalFeatureSpec(
data=None,
modality=modality,
identifier=f"{req_id}-0",
mm_position=PlaceholderRange(offset=0, length=length),
)
]
lora_state = LoraState(max_num_reqs=2)
lora_request = LoRARequest("vision-lora", 7, "/tmp/vision-lora")
lora_state.add_request("req-a", 0, lora_request)
lora_state.add_request("req-b", 1, None)
set_active_mm_loras(
model=model,
lora_manager=lora_manager,
encoder_cache=encoder_cache,
req_id_to_index={"req-a": 0, "req-b": 1},
lora_state=lora_state,
scheduled_encoder_inputs={"req-a": [0], "req-b": [0]},
)
# The image is encoded before the video, so its tokens come first.
_, tower_mapping = lora_manager.set_active_adapters.call_args_list[0].args
assert tower_mapping.index_mapping == (0, 0, 7)
def test_batch_mm_inputs_from_scheduler_sorts_by_modality():
features = {
f"req{i}": MultiModalFeatureSpec(
data=Mock(),
modality=m,
identifier=f"hash{i}",
mm_position=PlaceholderRange(offset=i, length=1),
)
for i, m in enumerate(["video", "image", "video"])
}
runner = SimpleNamespace(
requests={r: SimpleNamespace(mm_features=[f]) for r, f in features.items()}
)
scheduler_output = SimpleNamespace(
scheduled_encoder_inputs={r: [0] for r in features}
)
mm_hashes, mm_kwargs, mm_lora_refs = GPUModelRunner._batch_mm_inputs_from_scheduler(
runner, scheduler_output
)
order = ["req1", "req0", "req2"]
assert mm_hashes == [features[r].identifier for r in order]
assert mm_kwargs == [(features[r].modality, features[r].data) for r in order]
assert mm_lora_refs == [(r, features[r].mm_position) for r in order]
def test_update_states_new_request(model_runner, dist_init):
req_id = "req_0"