xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 07:57:45 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Misc] Clean up processor tests (#14771)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung 2025-03-14 02:25:37 +08:00 committed by GitHub
parent 01b3fd0af7
commit 8e9ffd37d6
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
8 changed files with 30 additions and 46 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
tests/models/multimodal/processing/test_h2ovl.py
View File

@ -96,14 +96,14 @@ def _run_check(
tokenizer = processor.info.get_tokenizer() tokenizer = processor.info.get_tokenizer()
config = processor.info.get_hf_config() config = processor.info.get_hf_config()
prompt = "<image>" * len(images)
mm_data = {"image": images} mm_data = {"image": images}
total_expected_num_patches = sum( total_expected_num_patches = sum(
_get_expected_num_patches(config, image, len(images), min_num, max_num) _get_expected_num_patches(config, image, len(images), min_num, max_num)
for image in images) for image in images)
processed_inputs = processor.apply("<image>" * len(images), mm_data, processed_inputs = processor.apply(prompt, mm_data, mm_processor_kwargs)
mm_processor_kwargs)
# Ensure we have the right number of placeholders per num_crops size # Ensure we have the right number of placeholders per num_crops size
image_token_id = tokenizer.convert_tokens_to_ids("<IMG_CONTEXT>") image_token_id = tokenizer.convert_tokens_to_ids("<IMG_CONTEXT>")
@ -152,9 +152,7 @@ def test_processor_override(
} }
ctx = build_model_context( ctx = build_model_context(
model_name=model_id, model_id,
tokenizer_name=model_id,
trust_remote_code=True,
mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None, mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None,
limit_mm_per_prompt={"image": len(size_factors)}, limit_mm_per_prompt={"image": len(size_factors)},
) )

10
tests/models/multimodal/processing/test_idefics3.py
View File

@ -9,10 +9,8 @@ from vllm.transformers_utils.tokenizer import cached_tokenizer_from_config
from ....conftest import _ImageAssets from ....conftest import _ImageAssets
from ...utils import build_model_context from ...utils import build_model_context
models = ["HuggingFaceM4/Idefics3-8B-Llama3"]
@pytest.mark.parametrize("model_id", ["HuggingFaceM4/Idefics3-8B-Llama3"])
@pytest.mark.parametrize("model", models)
# yapf: disable # yapf: disable
@pytest.mark.parametrize( @pytest.mark.parametrize(
("mm_processor_kwargs", "expected_toks_per_img"), ("mm_processor_kwargs", "expected_toks_per_img"),
@ -25,7 +23,7 @@ models = ["HuggingFaceM4/Idefics3-8B-Llama3"]
@pytest.mark.parametrize("kwargs_on_init", [True, False]) @pytest.mark.parametrize("kwargs_on_init", [True, False])
def test_processor_override( def test_processor_override(
image_assets: _ImageAssets, image_assets: _ImageAssets,
model: str, model_id: str,
mm_processor_kwargs: dict[str, object], mm_processor_kwargs: dict[str, object],
expected_toks_per_img: int, expected_toks_per_img: int,
num_imgs: int, num_imgs: int,
@ -36,9 +34,7 @@ def test_processor_override(
# in this test and assume that the kwargs will be correctly expanded by # in this test and assume that the kwargs will be correctly expanded by
# the partial when calling the custom input processor. # the partial when calling the custom input processor.
ctx = build_model_context( ctx = build_model_context(
model_name=model, model_id,
tokenizer_name=model,
trust_remote_code=True,
mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None, mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None,
limit_mm_per_prompt={"image": num_imgs}, limit_mm_per_prompt={"image": num_imgs},
) )

8
tests/models/multimodal/processing/test_internvl.py
View File

@ -56,14 +56,14 @@ def _run_check(
tokenizer = processor.info.get_tokenizer() tokenizer = processor.info.get_tokenizer()
config = processor.info.get_hf_config() config = processor.info.get_hf_config()
prompt = "<image>" * len(images)
mm_data = {"image": images} mm_data = {"image": images}
total_expected_num_patches = sum( total_expected_num_patches = sum(
_get_expected_num_patches(config, image, len(images), min_num, max_num) _get_expected_num_patches(config, image, len(images), min_num, max_num)
for image in images) for image in images)
processed_inputs = processor.apply("<image>" * len(images), mm_data, processed_inputs = processor.apply(prompt, mm_data, mm_processor_kwargs)
mm_processor_kwargs)
# Ensure we have the right number of placeholders per num_crops size # Ensure we have the right number of placeholders per num_crops size
image_token_id = tokenizer.convert_tokens_to_ids("<IMG_CONTEXT>") image_token_id = tokenizer.convert_tokens_to_ids("<IMG_CONTEXT>")
@ -109,9 +109,7 @@ def test_processor_override(
} }
ctx = build_model_context( ctx = build_model_context(
model_name=model_id, model_id,
tokenizer_name=model_id,
trust_remote_code=True,
mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None, mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None,
limit_mm_per_prompt={"image": len(size_factors)}, limit_mm_per_prompt={"image": len(size_factors)},
) )

9
tests/models/multimodal/processing/test_llava_next.py
View File

@ -36,8 +36,7 @@ def _validate_image_max_tokens_one(
@pytest.mark.parametrize("model_id", ["llava-hf/llava-v1.6-mistral-7b-hf"]) @pytest.mark.parametrize("model_id", ["llava-hf/llava-v1.6-mistral-7b-hf"])
def test_processor_max_tokens(model_id): def test_processor_max_tokens(model_id):
ctx = build_model_context( ctx = build_model_context(
model_name=model_id, model_id,
tokenizer_name=model_id,
mm_processor_kwargs=None, mm_processor_kwargs=None,
limit_mm_per_prompt={"image": 1}, limit_mm_per_prompt={"image": 1},
) )
@ -136,8 +135,7 @@ def _test_image_prompt_replacements(
@pytest.mark.parametrize("num_imgs", [1, 2]) @pytest.mark.parametrize("num_imgs", [1, 2])
def test_processor_prompt_replacements_regression(model_id, num_imgs): def test_processor_prompt_replacements_regression(model_id, num_imgs):
ctx = build_model_context( ctx = build_model_context(
model_name=model_id, model_id,
tokenizer_name=model_id,
mm_processor_kwargs=None, mm_processor_kwargs=None,
limit_mm_per_prompt={"image": num_imgs}, limit_mm_per_prompt={"image": num_imgs},
) )
@ -166,8 +164,7 @@ def test_processor_prompt_replacements_regression(model_id, num_imgs):
@pytest.mark.parametrize("num_imgs", [1]) @pytest.mark.parametrize("num_imgs", [1])
def test_processor_prompt_replacements_all(model_id, num_imgs): def test_processor_prompt_replacements_all(model_id, num_imgs):
ctx = build_model_context( ctx = build_model_context(
model_name=model_id, model_id,
tokenizer_name=model_id,
mm_processor_kwargs=None, mm_processor_kwargs=None,
limit_mm_per_prompt={"image": num_imgs}, limit_mm_per_prompt={"image": num_imgs},
) )

9
tests/models/multimodal/processing/test_llava_onevision.py
View File

@ -37,8 +37,7 @@ def _validate_image_max_tokens_one(
["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"]) ["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"])
def test_processor_max_tokens(model_id): def test_processor_max_tokens(model_id):
ctx = build_model_context( ctx = build_model_context(
model_name=model_id, model_id,
tokenizer_name=model_id,
mm_processor_kwargs=None, mm_processor_kwargs=None,
limit_mm_per_prompt={"image": 1}, limit_mm_per_prompt={"image": 1},
) )
@ -136,8 +135,7 @@ def _test_image_prompt_replacements(
@pytest.mark.parametrize("num_imgs", [1, 2]) @pytest.mark.parametrize("num_imgs", [1, 2])
def test_processor_prompt_replacements_regression(model_id, num_imgs): def test_processor_prompt_replacements_regression(model_id, num_imgs):
ctx = build_model_context( ctx = build_model_context(
model_name=model_id, model_id,
tokenizer_name=model_id,
mm_processor_kwargs=None, mm_processor_kwargs=None,
limit_mm_per_prompt={"image": num_imgs}, limit_mm_per_prompt={"image": num_imgs},
) )
@ -167,8 +165,7 @@ def test_processor_prompt_replacements_regression(model_id, num_imgs):
@pytest.mark.parametrize("num_imgs", [1]) @pytest.mark.parametrize("num_imgs", [1])
def test_processor_prompt_replacements_all(model_id, num_imgs): def test_processor_prompt_replacements_all(model_id, num_imgs):
ctx = build_model_context( ctx = build_model_context(
model_name=model_id, model_id,
tokenizer_name=model_id,
mm_processor_kwargs=None, mm_processor_kwargs=None,
limit_mm_per_prompt={"image": num_imgs}, limit_mm_per_prompt={"image": num_imgs},
) )

4
tests/models/multimodal/processing/test_phi3v.py
View File

@ -35,9 +35,7 @@ def test_processor_override(
from vllm.model_executor.models.phi3v import _IMAGE_TOKEN_ID from vllm.model_executor.models.phi3v import _IMAGE_TOKEN_ID
ctx = build_model_context( ctx = build_model_context(
model_name=model_id, model_id,
tokenizer_name=model_id,
trust_remote_code=True,
mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None, mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None,
limit_mm_per_prompt={"image": num_imgs}, limit_mm_per_prompt={"image": num_imgs},
) )

3
tests/models/multimodal/processing/test_qwen2_vl.py
View File

@ -30,8 +30,7 @@ def test_processor_override(
): ):
"""Ensure Qwen2VLMultiModalProcessor handles min/max pixels properly.""" """Ensure Qwen2VLMultiModalProcessor handles min/max pixels properly."""
ctx = build_model_context( ctx = build_model_context(
model_name=model_id, model_id,
tokenizer_name=model_id,
mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None, mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None,
limit_mm_per_prompt={"image": num_imgs}, limit_mm_per_prompt={"image": num_imgs},
) )

25
tests/models/utils.py
View File

@ -10,6 +10,8 @@ from vllm.config import ModelConfig, TaskOption
from vllm.inputs import InputContext from vllm.inputs import InputContext
from vllm.sequence import Logprob, PromptLogprobs, SampleLogprobs from vllm.sequence import Logprob, PromptLogprobs, SampleLogprobs
from .registry import HF_EXAMPLE_MODELS
TokensText = tuple[list[int], str] TokensText = tuple[list[int], str]
@ -250,10 +252,8 @@ def check_logprobs_close(
def build_model_context( def build_model_context(
model_name: str, model_id: str,
task: TaskOption = "auto", task: TaskOption = "auto",
tokenizer_name: Optional[str] = None,
trust_remote_code: bool = False,
dtype: Optional[Union[str, torch.dtype]] = None, dtype: Optional[Union[str, torch.dtype]] = None,
mm_processor_kwargs: Optional[dict] = None, mm_processor_kwargs: Optional[dict] = None,
limit_mm_per_prompt: Optional[dict] = None, limit_mm_per_prompt: Optional[dict] = None,
@ -262,9 +262,7 @@ def build_model_context(
"""Creates an InputContext for a given model. """Creates an InputContext for a given model.
Args: Args:
model_name: Name of the model being considered. model_id: ID of the model being considered.
tokenizer_name: Name of the tokenizer being considered.
trust_remote_code: Whether or not to allow loading remote code.
mm_processor_kwargs: optional processor kwargs for to be leveraged mm_processor_kwargs: optional processor kwargs for to be leveraged
in the input processor, mapper, dummy data creation, etc. in the input processor, mapper, dummy data creation, etc.
limit_mm_per_prompt: Multimodal limits. limit_mm_per_prompt: Multimodal limits.
@ -272,21 +270,24 @@ def build_model_context(
Returns: Returns:
InputContext for the model being considered. InputContext for the model being considered.
""" """
if tokenizer_name is None: model_info = HF_EXAMPLE_MODELS.find_hf_info(model_id)
tokenizer_name = model_name model_info.check_available_online(on_fail="skip")
model_info.check_transformers_version(on_fail="skip")
if dtype is None: if dtype is None:
dtype = "half" dtype = "half"
model_config = ModelConfig( model_config = ModelConfig(
model_name, model_id,
task=task, task=task,
tokenizer=tokenizer_name, tokenizer=model_info.tokenizer or model_id,
tokenizer_mode="auto", tokenizer_mode=model_info.tokenizer_mode,
trust_remote_code=trust_remote_code, trust_remote_code=model_info.trust_remote_code,
dtype=dtype, dtype=dtype,
seed=0, seed=0,
mm_processor_kwargs=mm_processor_kwargs, mm_processor_kwargs=mm_processor_kwargs,
limit_mm_per_prompt=limit_mm_per_prompt, limit_mm_per_prompt=limit_mm_per_prompt,
disable_mm_preprocessor_cache=disable_mm_preprocessor_cache, disable_mm_preprocessor_cache=disable_mm_preprocessor_cache,
hf_overrides=model_info.hf_overrides,
) )
return InputContext(model_config) return InputContext(model_config)