[Renderer] Move InputPreprocessor into Renderer (2/2) (#34560)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-02-17 05:29:01 -08:00
committed by GitHub
parent c61a98f529
commit 574fe75245
32 changed files with 984 additions and 1054 deletions
+6 -9
View File
@@ -195,18 +195,15 @@ def test_chat_batch_failure_cleanup(llm_for_failure_test):
valid_msg = [{"role": "user", "content": "Hello"}]
long_text = "This is a very long text to test the error " * 50
invalid_msg = [{"role": "user", "content": long_text}]
batch_1 = [
valid_msg,
valid_msg,
invalid_msg,
]
batch_2 = [
valid_msg,
valid_msg,
]
batch_1 = [valid_msg, valid_msg, invalid_msg]
batch_2 = [valid_msg, valid_msg]
sampling_params = SamplingParams(temperature=0, max_tokens=10)
with pytest.raises(ValueError, match="context length is only"):
llm.chat(batch_1, sampling_params=sampling_params)
assert llm.llm_engine.get_num_unfinished_requests() == 0
outputs_2 = llm.chat(batch_2, sampling_params=sampling_params)
assert len(outputs_2) == len(batch_2)
assert llm.llm_engine.get_num_unfinished_requests() == 0
@@ -489,8 +489,9 @@ def _assert_inputs_equal(
if ignore_mm_keys is None:
ignore_mm_keys = set()
a_rest = {k: v for k, v in a.items() if k != "mm_kwargs"}
b_rest = {k: v for k, v in b.items() if k != "mm_kwargs"}
ignore_prompt_keys = ("prompt", "mm_kwargs")
a_rest = {k: v for k, v in a.items() if k not in ignore_prompt_keys}
b_rest = {k: v for k, v in b.items() if k not in ignore_prompt_keys}
assert a_rest == b_rest, msg
@@ -0,0 +1,165 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from vllm.assets.image import ImageAsset
from vllm.assets.video import VideoAsset
from vllm.config import CacheConfig, ModelConfig, VllmConfig
from vllm.renderers.hf import HfRenderer
from vllm.tokenizers.registry import tokenizer_args_from_config
cherry_pil_image = ImageAsset("cherry_blossom").pil_image
stop_pil_image = ImageAsset("stop_sign").pil_image
baby_reading_np_ndarrays = VideoAsset("baby_reading").np_ndarrays
def _build_renderer(
*, mm_cache_gb: float = 4.0, enable_prefix_caching: bool = True
) -> HfRenderer:
model_config = ModelConfig(
model="Qwen/Qwen2.5-VL-3B-Instruct",
max_model_len=128,
mm_processor_cache_gb=mm_cache_gb,
)
vllm_config = VllmConfig(
model_config=model_config,
cache_config=CacheConfig(enable_prefix_caching=enable_prefix_caching),
)
_, tokenizer_name, _, kwargs = tokenizer_args_from_config(model_config)
return HfRenderer.from_config(
vllm_config,
tokenizer_kwargs={**kwargs, "tokenizer_name": tokenizer_name},
)
def test_multi_modal_uuids_length_mismatch_raises():
renderer = _build_renderer()
mm_data = {"image": [cherry_pil_image, stop_pil_image]}
# Mismatch: 2 items but only 1 uuid provided
mm_uuids = {"image": ["hash_cherry"]}
mm_processor = renderer.get_mm_processor()
mm_items = mm_processor.info.parse_mm_data(mm_data)
with pytest.raises(ValueError, match="must have same length as"):
renderer._process_mm_uuids(mm_data, mm_items, mm_uuids, "req-1")
def test_multi_modal_uuids_missing_modality_raises():
renderer = _build_renderer()
mm_data = {
"image": [cherry_pil_image],
"video": None,
}
# Only image uuids provided; video missing should raise
mm_uuids = {"image": ["hash_cherry"]}
mm_processor = renderer.get_mm_processor()
mm_items = mm_processor.info.parse_mm_data(mm_data)
with pytest.raises(ValueError, match="is empty but .* is missing"):
renderer._process_mm_uuids(mm_data, mm_items, mm_uuids, "req-2")
@pytest.mark.parametrize(
"mm_cache_gb, enable_prefix_caching",
[
(4.0, True), # default behavior
(4.0, False), # prefix caching disabled
(0.0, True), # processor cache disabled
],
)
def test_multi_modal_uuids_accepts_none_and_passes_through(
monkeypatch, mm_cache_gb: float, enable_prefix_caching: bool
):
renderer = _build_renderer(
mm_cache_gb=mm_cache_gb,
enable_prefix_caching=enable_prefix_caching,
)
mm_data = {
"image": [cherry_pil_image, stop_pil_image],
"video": baby_reading_np_ndarrays,
}
# Use a consistent two-image scenario across all configurations
mm_uuids = {"image": [None, "hash_stop"], "video": None}
mm_processor = renderer.get_mm_processor()
mm_items = mm_processor.info.parse_mm_data(mm_data)
processed_mm_uuids = renderer._process_mm_uuids(
mm_data, mm_items, mm_uuids, "req-3"
)
assert processed_mm_uuids == mm_uuids
@pytest.mark.parametrize(
"mm_cache_gb, enable_prefix_caching",
[
(4.0, True), # default behavior
(4.0, False), # prefix caching disabled
(0.0, True), # processor cache disabled
],
)
def test_multi_modal_uuids_accepts_empty(
monkeypatch, mm_cache_gb: float, enable_prefix_caching: bool
):
renderer = _build_renderer(
mm_cache_gb=mm_cache_gb,
enable_prefix_caching=enable_prefix_caching,
)
# While None means cached multi-modal input requiring UUIDs
# an empty list means no multi-modal input
mm_data = {"image": [], "video": []} # type: ignore[var-annotated]
mm_uuids = {"image": [], "video": None} # type: ignore[var-annotated]
mm_processor = renderer.get_mm_processor()
mm_items = mm_processor.info.parse_mm_data(mm_data)
processed_mm_uuids = renderer._process_mm_uuids(
mm_data, mm_items, mm_uuids, "req-4"
)
assert processed_mm_uuids == mm_uuids
def test_multi_modal_uuids_ignored_when_caching_disabled(monkeypatch):
# When both processor cache is 0 and prefix caching disabled, the
# processor builds overrides from request id instead of using user UUIDs.
renderer = _build_renderer(mm_cache_gb=0.0, enable_prefix_caching=False)
request_id = "req-42"
mm_data = {
"image": [cherry_pil_image, stop_pil_image],
"video": baby_reading_np_ndarrays,
}
mm_uuids = {"image": ["hash_cherry", "hash_stop"], "video": ["hash_video"]}
mm_processor = renderer.get_mm_processor()
mm_items = mm_processor.info.parse_mm_data(mm_data)
processed_mm_uuids = renderer._process_mm_uuids(
mm_data, mm_items, mm_uuids, request_id
)
# Expect request-id-based overrides are passed through
assert set(mm_uuids.keys()) == {"image", "video"}
assert len(mm_uuids["image"]) == 2
assert len(mm_uuids["video"]) == 1
assert processed_mm_uuids["image"][0].startswith(
f"{request_id}-image-"
) and processed_mm_uuids["image"][0].endswith("-0")
assert processed_mm_uuids["image"][1].startswith(
f"{request_id}-image-"
) and processed_mm_uuids["image"][1].endswith("-1")
assert processed_mm_uuids["video"][0].startswith(
f"{request_id}-video-"
) and processed_mm_uuids["video"][0].endswith("-0")
-2
View File
@@ -20,7 +20,6 @@ MM_BEAM_WIDTHS = [2]
MODELS = ["TinyLlama/TinyLlama-1.1B-Chat-v1.0"]
@pytest.mark.skip_v1 # V1 engine does not yet support beam search
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", MAX_TOKENS)
@@ -62,7 +61,6 @@ def test_beam_search_single_input(
)
@pytest.mark.skip_v1 # V1 engine does not yet support beam search
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", MAX_TOKENS)
@@ -1,174 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from vllm.assets.image import ImageAsset
from vllm.assets.video import VideoAsset
from vllm.config import CacheConfig, ModelConfig, VllmConfig
from vllm.multimodal import MultiModalUUIDDict
from vllm.sampling_params import SamplingParams
from vllm.v1.engine.input_processor import InputProcessor
cherry_pil_image = ImageAsset("cherry_blossom").pil_image
stop_pil_image = ImageAsset("stop_sign").pil_image
baby_reading_np_ndarrays = VideoAsset("baby_reading").np_ndarrays
def _build_input_processor(
*, mm_cache_gb: float = 4.0, enable_prefix_caching: bool = True
) -> InputProcessor:
model_config = ModelConfig(
model="Qwen/Qwen2.5-VL-3B-Instruct",
max_model_len=128,
mm_processor_cache_gb=mm_cache_gb,
)
vllm_config = VllmConfig(
model_config=model_config,
cache_config=CacheConfig(enable_prefix_caching=enable_prefix_caching),
)
return InputProcessor(vllm_config)
def test_multi_modal_uuids_length_mismatch_raises():
input_processor = _build_input_processor()
prompt = {
"prompt": "USER: <image>\nDescribe\nASSISTANT:",
"multi_modal_data": {"image": [cherry_pil_image, stop_pil_image]},
# Mismatch: 2 items but only 1 uuid provided
"multi_modal_uuids": {"image": ["hash_cherry"]},
}
with pytest.raises(ValueError, match="must have same length as"):
input_processor.process_inputs(
request_id="req-1",
prompt=prompt, # type: ignore[arg-type]
params=SamplingParams(),
)
def test_multi_modal_uuids_missing_modality_raises():
input_processor = _build_input_processor()
prompt = {
"prompt": "USER: <image><video>\nDescribe\nASSISTANT:",
# Two modalities provided in data
"multi_modal_data": {
"image": [cherry_pil_image],
"video": None,
},
# Only image uuids provided; video missing should raise
"multi_modal_uuids": {"image": ["hash_cherry"]},
}
with pytest.raises(ValueError, match="is empty but .* is missing"):
input_processor.process_inputs(
request_id="req-2",
prompt=prompt, # type: ignore[arg-type]
params=SamplingParams(),
)
@pytest.mark.parametrize(
"mm_cache_gb, enable_prefix_caching",
[
(4.0, True), # default behavior
(4.0, False), # prefix caching disabled
(0.0, True), # processor cache disabled
],
)
def test_multi_modal_uuids_accepts_none_and_passes_through(
monkeypatch, mm_cache_gb: float, enable_prefix_caching: bool
):
input_processor = _build_input_processor(
mm_cache_gb=mm_cache_gb,
enable_prefix_caching=enable_prefix_caching,
)
# Capture the overrides passed to InputPreprocessor.preprocess
captured: dict[str, object] = {}
def fake_preprocess(
prompt, *, tokenization_kwargs=None, lora_request=None, mm_uuids=None
):
captured["mm_uuids"] = mm_uuids
# Minimal processed inputs for decoder-only flow
return {"type": "token", "prompt_token_ids": [1]}
# Monkeypatch only the bound preprocess method on this instance
monkeypatch.setattr(
input_processor.input_preprocessor, "preprocess", fake_preprocess, raising=True
)
# Use a consistent two-image scenario across all configurations
mm_uuids = {"image": [None, "hash_stop"], "video": None}
prompt = {
"prompt": "USER: <image><image>\nTwo images\nASSISTANT:",
"multi_modal_data": {
"image": [cherry_pil_image, stop_pil_image],
"video": baby_reading_np_ndarrays,
},
"multi_modal_uuids": mm_uuids,
}
input_processor.process_inputs(
request_id="req-3",
prompt=prompt, # type: ignore[arg-type]
params=SamplingParams(),
)
assert captured["mm_uuids"] == mm_uuids
def test_multi_modal_uuids_ignored_when_caching_disabled(monkeypatch):
# When both processor cache is 0 and prefix caching disabled, the
# processor builds overrides from request id instead of using user UUIDs.
input_processor = _build_input_processor(
mm_cache_gb=0.0, enable_prefix_caching=False
)
captured: dict[str, MultiModalUUIDDict] = {}
def fake_preprocess(
prompt, *, tokenization_kwargs=None, lora_request=None, mm_uuids=None
):
captured["mm_uuids"] = mm_uuids
return {"type": "token", "prompt_token_ids": [1]}
monkeypatch.setattr(
input_processor.input_preprocessor, "preprocess", fake_preprocess, raising=True
)
request_id = "req-42"
mm_uuids = {"image": ["hash_cherry", "hash_stop"], "video": ["hash_video"]}
prompt = {
"prompt": "USER: <image><image><video>\nDescribe\nASSISTANT:",
"multi_modal_data": {
"image": [cherry_pil_image, stop_pil_image],
"video": [baby_reading_np_ndarrays],
},
"multi_modal_uuids": mm_uuids,
}
input_processor.process_inputs(
request_id=request_id,
prompt=prompt, # type: ignore[arg-type]
params=SamplingParams(),
)
# Expect request-id-based overrides are passed through
assert set(mm_uuids.keys()) == {"image", "video"}
assert len(mm_uuids["image"]) == 2
assert len(mm_uuids["video"]) == 1
assert captured["mm_uuids"]["image"][0].startswith(
f"{request_id}-image-"
) and captured["mm_uuids"]["image"][0].endswith("-0")
assert captured["mm_uuids"]["image"][1].startswith(
f"{request_id}-image-"
) and captured["mm_uuids"]["image"][1].endswith("-1")
assert captured["mm_uuids"]["video"][0].startswith(
f"{request_id}-video-"
) and captured["mm_uuids"]["video"][0].endswith("-0")