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 15, 19:45 UTC
[Misc] Clean up processor tests (#14771)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Mar 14, 2025 at 02:25 UTC
8e9ffd37d6499c6855b04b2a4d8f5326b5e2f78f
8 files changed
+30
-46
tests/models/multimodal/processing/test_h2ovl.py
+3
-5
@@ -96,14 +96,14 @@ def _run_check(
96
tokenizer = processor.info.get_tokenizer()
97
config = processor.info.get_hf_config()
98
99
+ prompt = "<image>" * len(images)
100
mm_data = {"image": images}
101
102
total_expected_num_patches = sum(
103
_get_expected_num_patches(config, image, len(images), min_num, max_num)
104
for image in images)
105
105
- processed_inputs = processor.apply("<image>" * len(images), mm_data,
106
- mm_processor_kwargs)
106
+ processed_inputs = processor.apply(prompt, mm_data, mm_processor_kwargs)
107
108
# Ensure we have the right number of placeholders per num_crops size
109
image_token_id = tokenizer.convert_tokens_to_ids("<IMG_CONTEXT>")
@@ -152,9 +152,7 @@ def test_processor_override(
152
}
153
154
ctx = build_model_context(
155
- model_name=model_id,
156
- tokenizer_name=model_id,
157
- trust_remote_code=True,
155
+ model_id,
156
mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None,
157
limit_mm_per_prompt={"image": len(size_factors)},
158
)
tests/models/multimodal/processing/test_idefics3.py
+3
-7
@@ -9,10 +9,8 @@ from vllm.transformers_utils.tokenizer import cached_tokenizer_from_config
9
from ....conftest import _ImageAssets
10
from ...utils import build_model_context
11
12
-models = ["HuggingFaceM4/Idefics3-8B-Llama3"]
12
14
-
15
-@pytest.mark.parametrize("model", models)
13
+@pytest.mark.parametrize("model_id", ["HuggingFaceM4/Idefics3-8B-Llama3"])
14
# yapf: disable
15
@pytest.mark.parametrize(
16
("mm_processor_kwargs", "expected_toks_per_img"),
@@ -25,7 +23,7 @@ models = ["HuggingFaceM4/Idefics3-8B-Llama3"]
23
@pytest.mark.parametrize("kwargs_on_init", [True, False])
24
def test_processor_override(
25
image_assets: _ImageAssets,
28
- model: str,
26
+ model_id: str,
27
mm_processor_kwargs: dict[str, object],
28
expected_toks_per_img: int,
29
num_imgs: int,
@@ -36,9 +34,7 @@ def test_processor_override(
34
# in this test and assume that the kwargs will be correctly expanded by
35
# the partial when calling the custom input processor.
36
ctx = build_model_context(
39
- model_name=model,
40
- tokenizer_name=model,
41
- trust_remote_code=True,
37
+ model_id,
38
mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None,
39
limit_mm_per_prompt={"image": num_imgs},
40
)
tests/models/multimodal/processing/test_internvl.py
+3
-5
@@ -56,14 +56,14 @@ def _run_check(
56
tokenizer = processor.info.get_tokenizer()
57
config = processor.info.get_hf_config()
58
59
+ prompt = "<image>" * len(images)
60
mm_data = {"image": images}
61
62
total_expected_num_patches = sum(
63
_get_expected_num_patches(config, image, len(images), min_num, max_num)
64
for image in images)
65
65
- processed_inputs = processor.apply("<image>" * len(images), mm_data,
66
- mm_processor_kwargs)
66
+ processed_inputs = processor.apply(prompt, mm_data, mm_processor_kwargs)
67
68
# Ensure we have the right number of placeholders per num_crops size
69
image_token_id = tokenizer.convert_tokens_to_ids("<IMG_CONTEXT>")
@@ -109,9 +109,7 @@ def test_processor_override(
109
}
110
111
ctx = build_model_context(
112
- model_name=model_id,
113
- tokenizer_name=model_id,
114
- trust_remote_code=True,
112
+ model_id,
113
mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None,
114
limit_mm_per_prompt={"image": len(size_factors)},
115
)
tests/models/multimodal/processing/test_llava_next.py
+3
-6
@@ -36,8 +36,7 @@ def _validate_image_max_tokens_one(
36
@pytest.mark.parametrize("model_id", ["llava-hf/llava-v1.6-mistral-7b-hf"])
37
def test_processor_max_tokens(model_id):
38
ctx = build_model_context(
39
- model_name=model_id,
40
- tokenizer_name=model_id,
39
+ model_id,
40
mm_processor_kwargs=None,
41
limit_mm_per_prompt={"image": 1},
42
)
@@ -136,8 +135,7 @@ def _test_image_prompt_replacements(
135
@pytest.mark.parametrize("num_imgs", [1, 2])
136
def test_processor_prompt_replacements_regression(model_id, num_imgs):
137
ctx = build_model_context(
139
- model_name=model_id,
140
- tokenizer_name=model_id,
138
+ model_id,
139
mm_processor_kwargs=None,
140
limit_mm_per_prompt={"image": num_imgs},
141
)
@@ -166,8 +164,7 @@ def test_processor_prompt_replacements_regression(model_id, num_imgs):
164
@pytest.mark.parametrize("num_imgs", [1])
165
def test_processor_prompt_replacements_all(model_id, num_imgs):
166
ctx = build_model_context(
169
- model_name=model_id,
170
- tokenizer_name=model_id,
167
+ model_id,
168
mm_processor_kwargs=None,
169
limit_mm_per_prompt={"image": num_imgs},
170
)
tests/models/multimodal/processing/test_llava_onevision.py
+3
-6
@@ -37,8 +37,7 @@ def _validate_image_max_tokens_one(
37
["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"])
38
def test_processor_max_tokens(model_id):
39
ctx = build_model_context(
40
- model_name=model_id,
41
- tokenizer_name=model_id,
40
+ model_id,
41
mm_processor_kwargs=None,
42
limit_mm_per_prompt={"image": 1},
43
)
@@ -136,8 +135,7 @@ def _test_image_prompt_replacements(
135
@pytest.mark.parametrize("num_imgs", [1, 2])
136
def test_processor_prompt_replacements_regression(model_id, num_imgs):
137
ctx = build_model_context(
139
- model_name=model_id,
140
- tokenizer_name=model_id,
138
+ model_id,
139
mm_processor_kwargs=None,
140
limit_mm_per_prompt={"image": num_imgs},
141
)
@@ -167,8 +165,7 @@ def test_processor_prompt_replacements_regression(model_id, num_imgs):
165
@pytest.mark.parametrize("num_imgs", [1])
166
def test_processor_prompt_replacements_all(model_id, num_imgs):
167
ctx = build_model_context(
170
- model_name=model_id,
171
- tokenizer_name=model_id,
168
+ model_id,
169
mm_processor_kwargs=None,
170
limit_mm_per_prompt={"image": num_imgs},
171
)
tests/models/multimodal/processing/test_phi3v.py
+1
-3
@@ -35,9 +35,7 @@ def test_processor_override(
35
from vllm.model_executor.models.phi3v import _IMAGE_TOKEN_ID
36
37
ctx = build_model_context(
38
- model_name=model_id,
39
- tokenizer_name=model_id,
40
- trust_remote_code=True,
38
+ model_id,
39
mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None,
40
limit_mm_per_prompt={"image": num_imgs},
41
)
tests/models/multimodal/processing/test_qwen2_vl.py
+1
-2
@@ -30,8 +30,7 @@ def test_processor_override(
30
):
31
"""Ensure Qwen2VLMultiModalProcessor handles min/max pixels properly."""
32
ctx = build_model_context(
33
- model_name=model_id,
34
- tokenizer_name=model_id,
33
+ model_id,
34
mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None,
35
limit_mm_per_prompt={"image": num_imgs},
36
)
tests/models/utils.py
+13
-12
@@ -10,6 +10,8 @@ from vllm.config import ModelConfig, TaskOption
10
from vllm.inputs import InputContext
11
from vllm.sequence import Logprob, PromptLogprobs, SampleLogprobs
12
13
+from .registry import HF_EXAMPLE_MODELS
14
+
15
TokensText = tuple[list[int], str]
16
17
@@ -250,10 +252,8 @@ def check_logprobs_close(
252
253
254
def build_model_context(
253
- model_name: str,
255
+ model_id: str,
256
task: TaskOption = "auto",
255
- tokenizer_name: Optional[str] = None,
256
- trust_remote_code: bool = False,
257
dtype: Optional[Union[str, torch.dtype]] = None,
258
mm_processor_kwargs: Optional[dict] = None,
259
limit_mm_per_prompt: Optional[dict] = None,
@@ -262,9 +262,7 @@ def build_model_context(
262
"""Creates an InputContext for a given model.
263
264
Args:
265
- model_name: Name of the model being considered.
266
- tokenizer_name: Name of the tokenizer being considered.
267
- trust_remote_code: Whether or not to allow loading remote code.
265
+ model_id: ID of the model being considered.
266
mm_processor_kwargs: optional processor kwargs for to be leveraged
267
in the input processor, mapper, dummy data creation, etc.
268
limit_mm_per_prompt: Multimodal limits.
@@ -272,21 +270,24 @@ def build_model_context(
270
Returns:
271
InputContext for the model being considered.
272
"""
275
- if tokenizer_name is None:
276
- tokenizer_name = model_name
273
+ model_info = HF_EXAMPLE_MODELS.find_hf_info(model_id)
274
+ model_info.check_available_online(on_fail="skip")
275
+ model_info.check_transformers_version(on_fail="skip")
276
+
277
if dtype is None:
278
dtype = "half"
279
280
model_config = ModelConfig(
281
- model_name,
281
+ model_id,
282
task=task,
283
- tokenizer=tokenizer_name,
284
- tokenizer_mode="auto",
285
- trust_remote_code=trust_remote_code,
283
+ tokenizer=model_info.tokenizer or model_id,
284
+ tokenizer_mode=model_info.tokenizer_mode,
285
+ trust_remote_code=model_info.trust_remote_code,
286
dtype=dtype,
287
seed=0,
288
mm_processor_kwargs=mm_processor_kwargs,
289
limit_mm_per_prompt=limit_mm_per_prompt,
290
disable_mm_preprocessor_cache=disable_mm_preprocessor_cache,
291
+ hf_overrides=model_info.hf_overrides,
292
)
293
return InputContext(model_config)