d7543862bd · tduncan/Nvidia-Optimization-Stack-VLLM

examples/offline_inference/qwen2_5_omni/only_thinker.py

+2 -2

@@ -47,7 +47,7 @@ def get_mixed_modalities_query() -> QueryResult:

47 "image":

48 ImageAsset("cherry_blossom").pil_image.convert("RGB"),

49 "video":

50 - VideoAsset(name="sample_demo_1", num_frames=16).np_ndarrays,

50 + VideoAsset(name="baby_reading", num_frames=16).np_ndarrays,

51 },

52 },

53 limit_mm_per_prompt={

@@ -65,7 +65,7 @@ def get_use_audio_in_video_query() -> QueryResult:

65 "<|im_start|>user\n<|vision_bos|><|VIDEO|><|vision_eos|>"

66 f"{question}<|im_end|>\n"

67 f"<|im_start|>assistant\n")

68 - asset = VideoAsset(name="sample_demo_1", num_frames=16)

68 + asset = VideoAsset(name="baby_reading", num_frames=16)

69 audio = asset.get_audio(sampling_rate=16000)

70 assert not envs.VLLM_USE_V1, ("V1 does not support use_audio_in_video. "

71 "Please launch this example with "

examples/offline_inference/vision_language.py

+1 -1

@@ -1109,7 +1109,7 @@ def get_multi_modal_input(args):

1109

1110 if args.modality == "video":

1111 # Input video and question

1112 - video = VideoAsset(name="sample_demo_1",

1112 + video = VideoAsset(name="baby_reading",

1113 num_frames=args.num_frames).np_ndarrays

1114 vid_questions = ["Why is this video funny?"]

1115

tests/conftest.py

+21 -35

@@ -1,9 +1,7 @@

1 # SPDX-License-Identifier: Apache-2.0

2 -

2 import json

3 import os

4 import tempfile

6 -from collections import UserList

5 from enum import Enum

6 from typing import Any, Callable, Optional, TypedDict, TypeVar, Union

7

@@ -58,16 +56,12 @@ def _read_prompts(filename: str) -> list[str]:

56 return prompts

57

58

61 -class _ImageAssetPrompts(TypedDict):

59 +class ImageAssetPrompts(TypedDict):

60 stop_sign: str

61 cherry_blossom: str

62

63

66 -class _ImageAssetsBase(UserList[ImageAsset]):

67 - pass

68 -

69 -

70 -class _ImageAssets(_ImageAssetsBase):

64 +class ImageTestAssets(list[ImageAsset]):

65

66 def __init__(self) -> None:

67 super().__init__([

@@ -75,7 +69,7 @@ class _ImageAssets(_ImageAssetsBase):

69 ImageAsset("cherry_blossom"),

70 ])

71

78 - def prompts(self, prompts: _ImageAssetPrompts) -> list[str]:

72 + def prompts(self, prompts: ImageAssetPrompts) -> list[str]:

73 """

74 Convenience method to define the prompt for each test image.

75

@@ -85,35 +79,27 @@ class _ImageAssets(_ImageAssetsBase):

79 return [prompts["stop_sign"], prompts["cherry_blossom"]]

80

81

88 -class _VideoAssetPrompts(TypedDict):

89 - sample_demo_1: str

90 -

82 +class VideoAssetPrompts(TypedDict):

83 + baby_reading: str

84

92 -class _VideoAssetsBase(UserList[VideoAsset]):

93 - pass

85

95 -

96 -class _VideoAssets(_VideoAssetsBase):

86 +class VideoTestAssets(list[VideoAsset]):

87

88 def __init__(self) -> None:

89 super().__init__([

100 - VideoAsset("sample_demo_1"),

90 + VideoAsset("baby_reading"),

91 ])

92

103 - def prompts(self, prompts: _VideoAssetPrompts) -> list[str]:

104 - return [prompts["sample_demo_1"]]

93 + def prompts(self, prompts: VideoAssetPrompts) -> list[str]:

94 + return [prompts["baby_reading"]]

95

96

107 -class _AudioAssetPrompts(TypedDict):

97 +class AudioAssetPrompts(TypedDict):

98 mary_had_lamb: str

99 winning_call: str

100

101

112 -class _AudioAssetsBase(UserList[AudioAsset]):

113 - pass

114 -

115 -

116 -class _AudioAssets(_AudioAssetsBase):

102 +class AudioTestAssets(list[AudioAsset]):

103

104 def __init__(self) -> None:

105 super().__init__([

@@ -121,16 +107,16 @@ class _AudioAssets(_AudioAssetsBase):

107 AudioAsset("winning_call"),

108 ])

109

124 - def prompts(self, prompts: _AudioAssetPrompts) -> list[str]:

110 + def prompts(self, prompts: AudioAssetPrompts) -> list[str]:

111 return [prompts["mary_had_lamb"], prompts["winning_call"]]

112

113

128 -IMAGE_ASSETS = _ImageAssets()

129 -"""Singleton instance of :class:`_ImageAssets`."""

130 -VIDEO_ASSETS = _VideoAssets()

131 -"""Singleton instance of :class:`_VideoAssets`."""

132 -AUDIO_ASSETS = _AudioAssets()

133 -"""Singleton instance of :class:`_AudioAssets`."""

114 +IMAGE_ASSETS = ImageTestAssets()

115 +"""Singleton instance of :class:`ImageTestAssets`."""

116 +VIDEO_ASSETS = VideoTestAssets()

117 +"""Singleton instance of :class:`VideoTestAssets`."""

118 +AUDIO_ASSETS = AudioTestAssets()

119 +"""Singleton instance of :class:`AudioTestAssets`."""

120

121

122 @pytest.fixture(scope="function", autouse=True)

@@ -278,17 +264,17 @@ def example_long_prompts() -> list[str]:

264

265

266 @pytest.fixture(scope="session")

281 -def image_assets() -> _ImageAssets:

267 +def image_assets() -> ImageTestAssets:

268 return IMAGE_ASSETS

269

270

271 @pytest.fixture(scope="session")

286 -def video_assets() -> _VideoAssets:

272 +def video_assets() -> VideoTestAssets:

273 return VIDEO_ASSETS

274

275

276 @pytest.fixture(scope="session")

291 -def audio_assets() -> _AudioAssets:

277 +def audio_assets() -> AudioTestAssets:

278 return AUDIO_ASSETS

279

280

tests/models/multimodal/generation/test_common.py

+11 -10

@@ -13,8 +13,8 @@ from transformers import AutoModelForImageTextToText, AutoModelForVision2Seq

13 from vllm.platforms import current_platform

14 from vllm.utils import identity

15

16 -from ....conftest import (IMAGE_ASSETS, HfRunner, VllmRunner, _ImageAssets,

17 - _VideoAssets)

16 +from ....conftest import (IMAGE_ASSETS, HfRunner, ImageTestAssets,

17 + VideoTestAssets, VllmRunner)

18 from ....utils import (create_new_process_for_each_test, large_gpu_mark,

19 multi_gpu_marks)

20 from ...utils import check_outputs_equal

@@ -691,7 +691,7 @@ def test_single_image_models(tmp_path: PosixPath, model_type: str,

691 test_case: ExpandableVLMTestArgs,

692 hf_runner: type[HfRunner],

693 vllm_runner: type[VllmRunner],

694 - image_assets: _ImageAssets, monkeypatch):

694 + image_assets: ImageTestAssets, monkeypatch):

695 if model_type in REQUIRES_V0_MODELS:

696 monkeypatch.setenv("VLLM_USE_V1", "0")

697 model_test_info = VLM_TEST_SETTINGS[model_type]

@@ -716,7 +716,7 @@ def test_multi_image_models(tmp_path: PosixPath, model_type: str,

716 test_case: ExpandableVLMTestArgs,

717 hf_runner: type[HfRunner],

718 vllm_runner: type[VllmRunner],

719 - image_assets: _ImageAssets, monkeypatch):

719 + image_assets: ImageTestAssets, monkeypatch):

720 if model_type in REQUIRES_V0_MODELS:

721 monkeypatch.setenv("VLLM_USE_V1", "0")

722 model_test_info = VLM_TEST_SETTINGS[model_type]

@@ -741,7 +741,7 @@ def test_image_embedding_models(model_type: str,

741 test_case: ExpandableVLMTestArgs,

742 hf_runner: type[HfRunner],

743 vllm_runner: type[VllmRunner],

744 - image_assets: _ImageAssets, monkeypatch):

744 + image_assets: ImageTestAssets, monkeypatch):

745 if model_type in REQUIRES_V0_MODELS:

746 monkeypatch.setenv("VLLM_USE_V1", "0")

747 model_test_info = VLM_TEST_SETTINGS[model_type]

@@ -763,7 +763,7 @@ def test_image_embedding_models(model_type: str,

763 ))

764 def test_video_models(model_type: str, test_case: ExpandableVLMTestArgs,

765 hf_runner: type[HfRunner], vllm_runner: type[VllmRunner],

766 - video_assets: _VideoAssets, monkeypatch):

766 + video_assets: VideoTestAssets, monkeypatch):

767 if model_type in REQUIRES_V0_MODELS:

768 monkeypatch.setenv("VLLM_USE_V1", "0")

769 model_test_info = VLM_TEST_SETTINGS[model_type]

@@ -814,7 +814,7 @@ def test_single_image_models_heavy(tmp_path: PosixPath, model_type: str,

814 test_case: ExpandableVLMTestArgs,

815 hf_runner: type[HfRunner],

816 vllm_runner: type[VllmRunner],

817 - image_assets: _ImageAssets, monkeypatch):

817 + image_assets: ImageTestAssets, monkeypatch):

818 if model_type in REQUIRES_V0_MODELS:

819 monkeypatch.setenv("VLLM_USE_V1", "0")

820 model_test_info = VLM_TEST_SETTINGS[model_type]

@@ -840,7 +840,7 @@ def test_multi_image_models_heavy(tmp_path: PosixPath, model_type: str,

840 test_case: ExpandableVLMTestArgs,

841 hf_runner: type[HfRunner],

842 vllm_runner: type[VllmRunner],

843 - image_assets: _ImageAssets, monkeypatch):

843 + image_assets: ImageTestAssets, monkeypatch):

844 if model_type in REQUIRES_V0_MODELS:

845 monkeypatch.setenv("VLLM_USE_V1", "0")

846 model_test_info = VLM_TEST_SETTINGS[model_type]

@@ -866,7 +866,8 @@ def test_image_embedding_models_heavy(model_type: str,

866 test_case: ExpandableVLMTestArgs,

867 hf_runner: type[HfRunner],

868 vllm_runner: type[VllmRunner],

869 - image_assets: _ImageAssets, monkeypatch):

869 + image_assets: ImageTestAssets,

870 + monkeypatch):

871 if model_type in REQUIRES_V0_MODELS:

872 monkeypatch.setenv("VLLM_USE_V1", "0")

873 model_test_info = VLM_TEST_SETTINGS[model_type]

@@ -889,7 +890,7 @@ def test_image_embedding_models_heavy(model_type: str,

890 def test_video_models_heavy(model_type: str, test_case: ExpandableVLMTestArgs,

891 hf_runner: type[HfRunner],

892 vllm_runner: type[VllmRunner],

892 - video_assets: _VideoAssets, monkeypatch):

893 + video_assets: VideoTestAssets, monkeypatch):

894 if model_type in REQUIRES_V0_MODELS:

895 monkeypatch.setenv("VLLM_USE_V1", "0")

896 model_test_info = VLM_TEST_SETTINGS[model_type]

tests/models/multimodal/generation/test_florence2.py

+2 -2

@@ -9,7 +9,7 @@ from vllm.inputs.data import ExplicitEncoderDecoderPrompt, TextPrompt

9 from vllm.multimodal.image import rescale_image_size

10 from vllm.sequence import SampleLogprobs

11

12 -from ....conftest import IMAGE_ASSETS, HfRunner, VllmRunner, _ImageAssets

12 +from ....conftest import IMAGE_ASSETS, HfRunner, ImageTestAssets, VllmRunner

13 from ...utils import check_logprobs_close

14

15 MODELS = ["microsoft/Florence-2-base"]

@@ -118,7 +118,7 @@ def run_test(

118 @pytest.mark.parametrize("max_tokens", [64])

119 @pytest.mark.parametrize("num_logprobs", [5])

120 def test_models(hf_runner: type[HfRunner], vllm_runner: type[VllmRunner],

121 - image_assets: _ImageAssets, model: str,

121 + image_assets: ImageTestAssets, model: str,

122 size_factors: list[int], dtype: str, max_tokens: int,

123 num_logprobs: int) -> None:

124 images = [asset.pil_image for asset in image_assets]

tests/models/multimodal/generation/test_granite_speech.py

+5 -4

@@ -9,7 +9,8 @@ from transformers import AutoModelForSpeechSeq2Seq

9 from vllm.lora.request import LoRARequest

10 from vllm.sequence import SampleLogprobs

11

12 -from ....conftest import HfRunner, PromptAudioInput, VllmRunner, _AudioAssets

12 +from ....conftest import (AudioTestAssets, HfRunner, PromptAudioInput,

13 + VllmRunner)

14 from ...registry import HF_EXAMPLE_MODELS

15 from ...utils import check_logprobs_close

16

@@ -116,9 +117,9 @@ def run_test(

117 @pytest.mark.parametrize("max_model_len", [2048])

118 @pytest.mark.parametrize("max_tokens", [128])

119 @pytest.mark.parametrize("num_logprobs", [10])

119 -def test_models(hf_runner, vllm_runner, model: str, audio_assets: _AudioAssets,

120 - dtype: str, max_model_len: int, max_tokens: int,

121 - num_logprobs: int) -> None:

120 +def test_models(hf_runner, vllm_runner, model: str,

121 + audio_assets: AudioTestAssets, dtype: str, max_model_len: int,

122 + max_tokens: int, num_logprobs: int) -> None:

123 model_info = HF_EXAMPLE_MODELS.find_hf_info(model)

124 model_info.check_available_online(on_fail="skip")

125 model_info.check_transformers_version(on_fail="skip")

tests/models/multimodal/generation/test_interleaved.py

+1 -1

@@ -29,7 +29,7 @@ def test_models(vllm_runner, model, dtype: str, max_tokens: int) -> None:

29 image_cherry = ImageAsset("cherry_blossom").pil_image.convert("RGB")

30 image_stop = ImageAsset("stop_sign").pil_image.convert("RGB")

31 images = [image_cherry, image_stop]

32 - video = VideoAsset(name="sample_demo_1", num_frames=16).np_ndarrays

32 + video = VideoAsset(name="baby_reading", num_frames=16).np_ndarrays

33

34 inputs = [

35 (

tests/models/multimodal/generation/test_mllama.py

+8 -8

@@ -14,8 +14,8 @@ from vllm.model_executor.models.mllama import MllamaForConditionalGeneration

14 from vllm.multimodal.image import rescale_image_size

15 from vllm.sequence import SampleLogprobs

16

17 -from ....conftest import (IMAGE_ASSETS, HfRunner, PromptImageInput, VllmRunner,

18 - _ImageAssets)

17 +from ....conftest import (IMAGE_ASSETS, HfRunner, ImageTestAssets,

18 + PromptImageInput, VllmRunner)

19 from ....quantization.utils import is_quant_method_supported

20 from ....utils import (create_new_process_for_each_test, large_gpu_test,

21 multi_gpu_test)

@@ -90,7 +90,7 @@ def vllm_to_hf_output(vllm_output: tuple[list[int], str,

90

91

92 def _get_inputs(

93 - image_assets: _ImageAssets,

93 + image_assets: ImageTestAssets,

94 *,

95 size_factors: Optional[list[float]] = None,

96 sizes: Optional[list[tuple[int, int]]] = None,

@@ -126,7 +126,7 @@ def _get_inputs(

126 def run_test(

127 hf_runner: type[HfRunner],

128 vllm_runner: type[VllmRunner],

129 - image_assets: _ImageAssets,

129 + image_assets: ImageTestAssets,

130 model: str,

131 *,

132 size_factors: list[float],

@@ -143,7 +143,7 @@ def run_test(

143 def run_test(

144 hf_runner: type[HfRunner],

145 vllm_runner: type[VllmRunner],

146 - image_assets: _ImageAssets,

146 + image_assets: ImageTestAssets,

147 model: str,

148 *,

149 sizes: list[tuple[int, int]],

@@ -159,7 +159,7 @@ def run_test(

159 def run_test(

160 hf_runner: type[HfRunner],

161 vllm_runner: type[VllmRunner],

162 - image_assets: _ImageAssets,

162 + image_assets: ImageTestAssets,

163 model: str,

164 *,

165 size_factors: Optional[list[float]] = None,

@@ -433,7 +433,7 @@ def test_models_distributed(

433 @pytest.mark.skipif(not is_quant_method_supported("bitsandbytes"),

434 reason='bitsandbytes is not supported on this GPU type.')

435 def test_bnb_regression(

436 - image_assets: _ImageAssets,

436 + image_assets: ImageTestAssets,

437 model: str,

438 dtype: str,

439 max_tokens: int,

@@ -473,7 +473,7 @@ def test_bnb_regression(

473 @pytest.mark.parametrize("dtype", ["bfloat16"])

474 @pytest.mark.parametrize("max_tokens", [32])

475 def test_explicit_implicit_prompt(

476 - image_assets: _ImageAssets,

476 + image_assets: ImageTestAssets,

477 model: str,

478 dtype: str,

479 max_tokens: int,

tests/models/multimodal/generation/test_qwen2_vl.py

+1 -1

@@ -50,7 +50,7 @@ IMAGE_PROMPTS = IMAGE_ASSETS.prompts({

50 })

51

52 VIDEO_PROMPTS = VIDEO_ASSETS.prompts({

53 - "sample_demo_1":

53 + "baby_reading":

54 qwen2_vl_chat_template(

55 VIDEO_PLACEHOLDER,

56 "Describe this video with a short sentence ",

tests/models/multimodal/generation/test_ultravox.py

+25 -19

@@ -11,13 +11,22 @@ from transformers import AutoModel, AutoTokenizer

11 from vllm.multimodal.audio import resample_audio_librosa

12 from vllm.sequence import SampleLogprobs

13

14 -from ....conftest import HfRunner, VllmRunner, _AudioAssets

14 +from ....conftest import AUDIO_ASSETS, AudioTestAssets, HfRunner, VllmRunner

15 from ....utils import RemoteOpenAIServer

16 from ...registry import HF_EXAMPLE_MODELS

17 from ...utils import check_logprobs_close

18

19 MODEL_NAME = "fixie-ai/ultravox-v0_5-llama-3_2-1b"

20

21 +AUDIO_PROMPTS = AUDIO_ASSETS.prompts({

22 + "mary_had_lamb":

23 + "Transcribe this into English.",

24 + "winning_call":

25 + "What is happening in this audio clip?",

26 +})

27 +

28 +MULTI_AUDIO_PROMPT = "Describe each of the audios above."

29 +

30 AudioTuple = tuple[np.ndarray, int]

31

32 VLLM_PLACEHOLDER = "<|audio|>"

@@ -31,12 +40,6 @@ CHUNKED_PREFILL_KWARGS = {

40 }

41

42

34 -@pytest.fixture(scope="module", params=("mary_had_lamb", "winning_call"))

35 -def audio(request):

36 - from vllm.assets.audio import AudioAsset

37 - return AudioAsset(request.param)

38 -

39 -

43 def params_kwargs_to_cli_args(params_kwargs: dict[str, Any]) -> list[str]:

44 """Convert kwargs to CLI args."""

45 args = []

@@ -53,7 +56,7 @@ def params_kwargs_to_cli_args(params_kwargs: dict[str, Any]) -> list[str]:

56 pytest.param({}, marks=pytest.mark.cpu_model),

57 pytest.param(CHUNKED_PREFILL_KWARGS),

58 ])

56 -def server(request, audio_assets: _AudioAssets):

59 +def server(request, audio_assets: AudioTestAssets):

60 args = [

61 "--dtype", "bfloat16", "--max-model-len", "4096", "--enforce-eager",

62 "--limit-mm-per-prompt",

@@ -199,15 +202,19 @@ def run_multi_audio_test(

202 pytest.param({}, marks=pytest.mark.cpu_model),

203 pytest.param(CHUNKED_PREFILL_KWARGS),

204 ])

202 -def test_models(hf_runner, vllm_runner, audio, dtype: str, max_tokens: int,

203 - num_logprobs: int, vllm_kwargs: dict) -> None:

205 +def test_models(hf_runner, vllm_runner, audio_assets: AudioTestAssets,

206 + dtype: str, max_tokens: int, num_logprobs: int,

207 + vllm_kwargs: dict) -> None:

208 + audio_inputs = [(

209 + _get_prompt(1, audio, VLLM_PLACEHOLDER),

210 + _get_prompt(1, audio, HF_PLACEHOLDER),

211 + audio.audio_and_sample_rate,

212 + ) for audio in audio_assets]

213

205 - vllm_prompt = _get_prompt(1, "Describe the audio above.", VLLM_PLACEHOLDER)

206 - hf_prompt = _get_prompt(1, "Describe the audio above.", HF_PLACEHOLDER)

214 run_test(

215 hf_runner,

216 vllm_runner,

210 - [(vllm_prompt, hf_prompt, audio.audio_and_sample_rate)],

217 + audio_inputs,

218 MODEL_NAME,

219 dtype=dtype,

220 max_tokens=max_tokens,

@@ -224,13 +231,12 @@ def test_models(hf_runner, vllm_runner, audio, dtype: str, max_tokens: int,

231 pytest.param({}, marks=pytest.mark.cpu_model),

232 pytest.param(CHUNKED_PREFILL_KWARGS),

233 ])

227 -def test_models_with_multiple_audios(vllm_runner, audio_assets: _AudioAssets,

228 - dtype: str, max_tokens: int,

229 - num_logprobs: int,

234 +def test_models_with_multiple_audios(vllm_runner,

235 + audio_assets: AudioTestAssets, dtype: str,

236 + max_tokens: int, num_logprobs: int,

237 vllm_kwargs: dict) -> None:

238

232 - vllm_prompt = _get_prompt(len(audio_assets),

233 - "Describe each of the audios above.",

239 + vllm_prompt = _get_prompt(len(audio_assets), MULTI_AUDIO_PROMPT,

240 VLLM_PLACEHOLDER)

241 run_multi_audio_test(

242 vllm_runner,

@@ -245,7 +251,7 @@ def test_models_with_multiple_audios(vllm_runner, audio_assets: _AudioAssets,

251

252

253 @pytest.mark.asyncio

248 -async def test_online_serving(client, audio_assets: _AudioAssets):

254 +async def test_online_serving(client, audio_assets: AudioTestAssets):

255 """Exercises online serving with/without chunked prefill enabled."""

256

257 messages = [{

tests/models/multimodal/generation/vlm_utils/builders.py

+5 -5

@@ -11,7 +11,7 @@ from vllm.multimodal.image import rescale_image_size

11 from vllm.multimodal.video import (rescale_video_size, resize_video,

12 sample_frames_from_video)

13

14 -from .....conftest import _ImageAssets, _VideoAssets

14 +from .....conftest import ImageTestAssets, VideoTestAssets

15 from .types import (SINGLE_IMAGE_BASE_PROMPTS, TEST_IMG_PLACEHOLDER,

16 TEST_VIDEO_PLACEHOLDER, VIDEO_BASE_PROMPT,

17 ImageSizeWrapper, SizeType, VLMTestInfo)

@@ -69,7 +69,7 @@ def get_model_prompts(base_prompts: Iterable[str],

69

70 def build_single_image_inputs_from_test_info(

71 test_info: VLMTestInfo,

72 - image_assets: _ImageAssets,

72 + image_assets: ImageTestAssets,

73 size_wrapper: ImageSizeWrapper,

74 tmp_path: Optional[PosixPath] = None):

75 if test_info.prompt_formatter is None:

@@ -116,7 +116,7 @@ def build_single_image_inputs(images, model_prompts,

116

117 def build_multi_image_inputs_from_test_info(

118 test_info: VLMTestInfo,

119 - image_assets: _ImageAssets,

119 + image_assets: ImageTestAssets,

120 size_wrapper: ImageSizeWrapper,

121 tmp_path: Optional[PosixPath] = None):

122 if test_info.prompt_formatter is None:

@@ -159,7 +159,7 @@ def build_multi_image_inputs(image_lists, model_prompts,

159

160 def build_embedding_inputs_from_test_info(

161 test_info: VLMTestInfo,

162 - image_assets: _ImageAssets,

162 + image_assets: ImageTestAssets,

163 size_wrapper: ImageSizeWrapper,

164 ):

165 # These conditions will always be true if invoked through filtering,

@@ -192,7 +192,7 @@ def build_embedding_inputs_from_test_info(

192

193 def build_video_inputs_from_test_info(

194 test_info: VLMTestInfo,

195 - video_assets: _VideoAssets,

195 + video_assets: VideoTestAssets,

196 size_wrapper: ImageSizeWrapper,

197 num_frames: int,

198 ):

tests/models/multimodal/generation/vlm_utils/model_utils.py

+4 -4

@@ -16,7 +16,7 @@ from transformers import (AutoConfig, AutoTokenizer, BatchFeature,

16 from vllm.sequence import SampleLogprobs

17 from vllm.transformers_utils.tokenizer import patch_padding_side

18

19 -from .....conftest import HfRunner, ImageAsset, _ImageAssets

19 +from .....conftest import HfRunner, ImageAsset, ImageTestAssets

20 from .types import RunnerOutput

21

22

@@ -238,14 +238,14 @@ def minimax_vl_01_hf_output(hf_output: RunnerOutput,

238

239

240 ####### Functions for converting image assets to embeddings

241 -def get_llava_embeddings(image_assets: _ImageAssets):

241 +def get_llava_embeddings(image_assets: ImageTestAssets):

242 return [asset.image_embeds for asset in image_assets]

243

244

245 ####### Prompt path encoders for models that need models on disk

246 def qwen_prompt_path_encoder(

247 - tmp_path: PosixPath, prompt: str, assets: Union[list[ImageAsset],

248 - _ImageAssets]) -> str:

247 + tmp_path: PosixPath, prompt: str,

248 + assets: Union[list[ImageAsset], ImageTestAssets]) -> str:

249 """Given a temporary dir path, export one or more image assets into the

250 tempdir & replace its contents with the local path to the string so that

251 the HF version of Qwen-VL can resolve the path and load the image in its

tests/models/multimodal/generation/vlm_utils/runners.py

+6 -5

@@ -4,7 +4,8 @@ types / modalities.

4 """

5 from pathlib import PosixPath

6

7 -from .....conftest import HfRunner, VllmRunner, _ImageAssets, _VideoAssets

7 +from .....conftest import (HfRunner, ImageTestAssets, VideoTestAssets,

8 + VllmRunner)

9 from . import builders, core

10 from .types import ExpandableVLMTestArgs, VLMTestInfo

11

@@ -14,7 +15,7 @@ def run_single_image_test(*, tmp_path: PosixPath, model_test_info: VLMTestInfo,

15 test_case: ExpandableVLMTestArgs,

16 hf_runner: type[HfRunner],

17 vllm_runner: type[VllmRunner],

17 - image_assets: _ImageAssets):

18 + image_assets: ImageTestAssets):

19 assert test_case.size_wrapper is not None

20 inputs = builders.build_single_image_inputs_from_test_info(

21 model_test_info, image_assets, test_case.size_wrapper, tmp_path)

@@ -37,7 +38,7 @@ def run_multi_image_test(*, tmp_path: PosixPath, model_test_info: VLMTestInfo,

38 test_case: ExpandableVLMTestArgs,

39 hf_runner: type[HfRunner],

40 vllm_runner: type[VllmRunner],

40 - image_assets: _ImageAssets):

41 + image_assets: ImageTestAssets):

42 assert test_case.size_wrapper is not None

43 inputs = builders.build_multi_image_inputs_from_test_info(

44 model_test_info, image_assets, test_case.size_wrapper, tmp_path)

@@ -60,7 +61,7 @@ def run_embedding_test(*, model_test_info: VLMTestInfo,

61 test_case: ExpandableVLMTestArgs,

62 hf_runner: type[HfRunner],

63 vllm_runner: type[VllmRunner],

63 - image_assets: _ImageAssets):

64 + image_assets: ImageTestAssets):

65 assert test_case.size_wrapper is not None

66 inputs, vllm_embeddings = builders.build_embedding_inputs_from_test_info(

67 model_test_info, image_assets, test_case.size_wrapper)

@@ -86,7 +87,7 @@ def run_video_test(

87 test_case: ExpandableVLMTestArgs,

88 hf_runner: type[HfRunner],

89 vllm_runner: type[VllmRunner],

89 - video_assets: _VideoAssets,

90 + video_assets: VideoTestAssets,

91 ):

92 assert test_case.size_wrapper is not None

93 assert test_case.num_video_frames is not None

tests/models/multimodal/generation/vlm_utils/types.py

+3 -3

@@ -15,7 +15,7 @@ from vllm.config import TaskOption

15 from vllm.sequence import SampleLogprobs

16 from vllm.transformers_utils.tokenizer import AnyTokenizer

17

18 -from .....conftest import IMAGE_ASSETS, HfRunner, ImageAsset, _ImageAssets

18 +from .....conftest import IMAGE_ASSETS, HfRunner, ImageAsset, ImageTestAssets

19 from ....utils import check_logprobs_close

20

21 # meta image tag; will be replaced by the appropriate tag for the model

@@ -85,7 +85,7 @@ class VLMTestInfo(NamedTuple):

85

86 # Function for converting ImageAssets to image embeddings;

87 # We need to define this explicitly for embedding tests

88 - convert_assets_to_embeddings: Optional[Callable[[_ImageAssets],

88 + convert_assets_to_embeddings: Optional[Callable[[ImageTestAssets],

89 torch.Tensor]] = None

90

91 # Exposed options for vLLM runner; we change these in a several tests,

@@ -141,7 +141,7 @@ class VLMTestInfo(NamedTuple):

141 # for Qwen-VL, which requires encoding the image path / url into the prompt

142 # for HF runner

143 prompt_path_encoder: Optional[

144 - Callable[[PosixPath, str, Union[list[ImageAsset], _ImageAssets]],

144 + Callable[[PosixPath, str, Union[list[ImageAsset], ImageTestAssets]],

145 str]] = None # noqa: E501

146

147 # Allows configuring a test to run with custom inputs

tests/models/multimodal/pooling/test_intern_vit.py

+2 -2

@@ -7,7 +7,7 @@ from transformers import AutoConfig, AutoModel, CLIPImageProcessor

7

8 from vllm.distributed import cleanup_dist_env_and_memory

9

10 -from ....conftest import _ImageAssets

10 +from ....conftest import ImageTestAssets

11

12 # we use snapshot_download to prevent conflicts between

13 # dynamic_module and trust_remote_code for hf_runner

@@ -15,7 +15,7 @@ DOWNLOAD_PATTERN = ["*.json", "*.py", "*.safetensors", "*.txt", "*.model"]

15

16

17 def run_intern_vit_test(

18 - image_assets: _ImageAssets,

18 + image_assets: ImageTestAssets,

19 model_id: str,

20 *,

21 dtype: str,

tests/models/multimodal/processing/test_h2ovl.py

+2 -2

@@ -11,7 +11,7 @@ from vllm.multimodal import MULTIMODAL_REGISTRY

11 from vllm.multimodal.image import rescale_image_size

12 from vllm.multimodal.processing import BaseMultiModalProcessor

13

14 -from ....conftest import _ImageAssets

14 +from ....conftest import ImageTestAssets

15 from ...utils import build_model_context

16

17

@@ -137,7 +137,7 @@ def _run_check(

137 @pytest.mark.parametrize("kwargs_on_init", [True, False])

138 def test_processor_override(

139 model_id: str,

140 - image_assets: _ImageAssets,

140 + image_assets: ImageTestAssets,

141 size_factors: list[int],

142 min_dynamic_patch: int,

143 max_dynamic_patch: int,

tests/models/multimodal/processing/test_idefics3.py

+2 -2

@@ -5,7 +5,7 @@ from transformers import Idefics3Config

5

6 from vllm.multimodal import MULTIMODAL_REGISTRY

7

8 -from ....conftest import _ImageAssets

8 +from ....conftest import ImageTestAssets

9 from ...utils import build_model_context

10

11

@@ -21,7 +21,7 @@ from ...utils import build_model_context

21 @pytest.mark.parametrize("num_imgs", [1, 2])

22 @pytest.mark.parametrize("kwargs_on_init", [True, False])

23 def test_processor_override(

24 - image_assets: _ImageAssets,

24 + image_assets: ImageTestAssets,

25 model_id: str,

26 mm_processor_kwargs: dict[str, object],

27 expected_toks_per_img: int,

tests/models/multimodal/processing/test_internvl.py

+2 -2

@@ -11,7 +11,7 @@ from vllm.multimodal import MULTIMODAL_REGISTRY

11 from vllm.multimodal.image import rescale_image_size

12 from vllm.multimodal.processing import BaseMultiModalProcessor

13

14 -from ....conftest import _ImageAssets

14 +from ....conftest import ImageTestAssets

15 from ...utils import build_model_context

16

17

@@ -94,7 +94,7 @@ def _run_check(

94 @pytest.mark.parametrize("kwargs_on_init", [True, False])

95 def test_processor_override(

96 model_id: str,

97 - image_assets: _ImageAssets,

97 + image_assets: ImageTestAssets,

98 size_factors: list[int],

99 min_dynamic_patch: int,

100 max_dynamic_patch: int,

tests/models/multimodal/processing/test_llama4.py

+2 -2

@@ -6,7 +6,7 @@ import pytest

6 from vllm.multimodal import MULTIMODAL_REGISTRY

7 from vllm.transformers_utils.tokenizer import encode_tokens

8

9 -from ....conftest import _ImageAssets

9 +from ....conftest import ImageTestAssets

10 from ...utils import build_model_context

11

12

@@ -17,7 +17,7 @@ from ...utils import build_model_context

17 @pytest.mark.parametrize("disable_mm_preprocessor_cache", [True, False])

18 @pytest.mark.parametrize("tokenized_prompt", [True, False])

19 def test_processor_override(

20 - image_assets: _ImageAssets,

20 + image_assets: ImageTestAssets,

21 model_id: str,

22 mm_processor_kwargs: dict,

23 num_imgs: int,

tests/models/multimodal/processing/test_minimax_vl_01.py

+2 -2

@@ -7,14 +7,14 @@ from vllm.multimodal import MULTIMODAL_REGISTRY

7 from vllm.multimodal.parse import ImageSize

8 from vllm.multimodal.processing import BaseMultiModalProcessor

9

10 -from ....conftest import _ImageAssets

10 +from ....conftest import ImageTestAssets

11 from ...utils import build_model_context

12

13

14 @pytest.mark.parametrize("model_id", ["MiniMaxAI/MiniMax-VL-01"])

15 @pytest.mark.parametrize("num_imgs", [1, 2])

16 def test_processor_override(

17 - image_assets: _ImageAssets,

17 + image_assets: ImageTestAssets,

18 model_id: str,

19 num_imgs: int,

20 ):

tests/models/multimodal/processing/test_phi3v.py

+2 -2

@@ -4,7 +4,7 @@ import pytest

4

5 from vllm.multimodal import MULTIMODAL_REGISTRY

6

7 -from ....conftest import _ImageAssets

7 +from ....conftest import ImageTestAssets

8 from ...utils import build_model_context

9

10

@@ -22,7 +22,7 @@ from ...utils import build_model_context

22 @pytest.mark.parametrize("num_imgs", [1, 2])

23 @pytest.mark.parametrize("kwargs_on_init", [True, False])

24 def test_processor_override(

25 - image_assets: _ImageAssets,

25 + image_assets: ImageTestAssets,

26 model_id: str,

27 mm_processor_kwargs: dict[str, int],

28 expected_toks_per_img: int,

tests/models/multimodal/processing/test_phi4mm.py

+2 -2

@@ -4,7 +4,7 @@ import pytest

4

5 from vllm.multimodal import MULTIMODAL_REGISTRY

6

7 -from ....conftest import _ImageAssets

7 +from ....conftest import ImageTestAssets

8 from ...utils import build_model_context

9

10

@@ -22,7 +22,7 @@ from ...utils import build_model_context

22 @pytest.mark.parametrize("num_imgs", [1, 2])

23 @pytest.mark.parametrize("kwargs_on_init", [True, False])

24 def test_processor_override(

25 - image_assets: _ImageAssets,

25 + image_assets: ImageTestAssets,

26 model_id: str,

27 mm_processor_kwargs: dict[str, int],

28 expected_toks_per_img: int,

tests/models/multimodal/processing/test_qwen2_vl.py

+2 -2

@@ -4,7 +4,7 @@ import pytest

4

5 from vllm.multimodal import MULTIMODAL_REGISTRY

6

7 -from ....conftest import _ImageAssets

7 +from ....conftest import ImageTestAssets

8 from ...utils import build_model_context

9

10

@@ -19,7 +19,7 @@ from ...utils import build_model_context

19 @pytest.mark.parametrize("num_imgs", [1, 2])

20 @pytest.mark.parametrize("kwargs_on_init", [True, False])

21 def test_processor_override(

22 - image_assets: _ImageAssets,

22 + image_assets: ImageTestAssets,

23 model_id: str,

24 mm_processor_kwargs: dict[str, object],

25 expected_toks_per_img: int,

tests/models/multimodal/processing/test_smolvlm.py

+2 -2

@@ -5,7 +5,7 @@ from transformers import SmolVLMConfig

5

6 from vllm.multimodal import MULTIMODAL_REGISTRY

7

8 -from ....conftest import _ImageAssets

8 +from ....conftest import ImageTestAssets

9 from ...utils import build_model_context

10

11

@@ -21,7 +21,7 @@ from ...utils import build_model_context

21 @pytest.mark.parametrize("num_imgs", [1, 2])

22 @pytest.mark.parametrize("kwargs_on_init", [True, False])

23 def test_processor_override(

24 - image_assets: _ImageAssets,

24 + image_assets: ImageTestAssets,

25 model_id: str,

26 mm_processor_kwargs: dict[str, object],

27 expected_toks_per_img: int,

tests/models/quantization/test_awq.py

+2 -2

@@ -7,7 +7,7 @@ import torch

7

8 from vllm.multimodal.image import rescale_image_size

9

10 -from ...conftest import IMAGE_ASSETS, VllmRunner, _ImageAssets

10 +from ...conftest import IMAGE_ASSETS, ImageTestAssets, VllmRunner

11 from ..utils import check_logprobs_close

12

13 HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({

@@ -20,7 +20,7 @@ HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({

20

21 def run_awq_test(

22 vllm_runner: type[VllmRunner],

23 - image_assets: _ImageAssets,

23 + image_assets: ImageTestAssets,

24 source_model: str,

25 quant_model: str,

26 *,

vllm/assets/audio.py

+9 -3

@@ -18,19 +18,25 @@ except ImportError:

18

19 ASSET_DIR = "multimodal_asset"

20

21 +AudioAssetName = Literal["winning_call", "mary_had_lamb"]

22 +

23

24 @dataclass(frozen=True)

25 class AudioAsset:

24 - name: Literal["winning_call", "mary_had_lamb"]

26 + name: AudioAssetName

27 +

28 + @property

29 + def filename(self) -> str:

30 + return f"{self.name}.ogg"

31

32 @property

33 def audio_and_sample_rate(self) -> tuple[npt.NDArray, float]:

28 - audio_path = get_vllm_public_assets(filename=f"{self.name}.ogg",

34 + audio_path = get_vllm_public_assets(filename=self.filename,

35 s3_prefix=ASSET_DIR)

36 return librosa.load(audio_path, sr=None)

37

38 def get_local_path(self) -> Path:

33 - return get_vllm_public_assets(filename=f"{self.name}.ogg",

39 + return get_vllm_public_assets(filename=self.filename,

40 s3_prefix=ASSET_DIR)

41

42 @property

vllm/assets/image.py

+3 -1

@@ -10,10 +10,12 @@ from .base import get_vllm_public_assets

10

11 VLM_IMAGES_DIR = "vision_model_images"

12

13 +ImageAssetName = Literal["stop_sign", "cherry_blossom"]

14 +

15

16 @dataclass(frozen=True)

17 class ImageAsset:

16 - name: Literal["stop_sign", "cherry_blossom"]

18 + name: ImageAssetName

19

20 @property

21 def pil_image(self) -> Image.Image:

vllm/assets/video.py

+16 -5

@@ -2,7 +2,7 @@

2

3 from dataclasses import dataclass

4 from functools import lru_cache

5 -from typing import Literal, Optional

5 +from typing import ClassVar, Literal, Optional

6

7 import cv2

8 import numpy as np

@@ -76,20 +76,31 @@ def video_to_pil_images_list(path: str,

76 ]

77

78

79 +VideoAssetName = Literal["baby_reading"]

80 +

81 +

82 @dataclass(frozen=True)

83 class VideoAsset:

81 - name: Literal["sample_demo_1"]

84 + name: VideoAssetName

85 num_frames: int = -1

86

87 + _NAME_TO_FILE: ClassVar[dict[VideoAssetName, str]] = {

88 + "baby_reading": "sample_demo_1.mp4",

89 + }

90 +

91 + @property

92 + def filename(self) -> str:

93 + return self._NAME_TO_FILE[self.name]

94 +

95 @property

96 def pil_images(self) -> list[Image.Image]:

86 - video_path = download_video_asset(self.name + ".mp4")

97 + video_path = download_video_asset(self.filename)

98 ret = video_to_pil_images_list(video_path, self.num_frames)

99 return ret

100

101 @property

102 def np_ndarrays(self) -> npt.NDArray:

92 - video_path = download_video_asset(self.name + ".mp4")

103 + video_path = download_video_asset(self.filename)

104 ret = video_to_ndarrays(video_path, self.num_frames)

105 return ret

106

@@ -99,5 +110,5 @@ class VideoAsset:

110

111 See also: examples/offline_inference/qwen2_5_omni/only_thinker.py

112 """

102 - video_path = download_video_asset(self.name + ".mp4")

113 + video_path = download_video_asset(self.filename)

114 return librosa.load(video_path, sr=sampling_rate)[0]

[Misc] Rename assets for testing (#17575)