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 16, 03:15 UTC
[CI/Build] Move model-specific multi-modal processing tests (#11934)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Jan 11, 2025 at 13:50 UTC
7a3a83e3b87f50fe9c0985a5c5bcc1d4cf2e95cd
13 files changed
+251
-240
.buildkite/test-pipeline.yaml
+1
@@ -368,6 +368,7 @@ steps:
368
- tests/models/encoder_decoder/vision_language
369
commands:
370
- pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
371
+ - pytest -v -s models/multimodal
372
- pytest -v -s models/decoder_only/audio_language -m 'core_model or quant_model'
373
- pytest -v -s --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m 'core_model or quant_model'
374
- pytest -v -s models/embedding/vision_language -m core_model
tests/models/multimodal/__init__.py
renamed
tests/models/multimodal/processing/__init__.py
tests/models/multimodal/processing/test_common.py
new
+201
@@ -0,0 +1,201 @@
1
+from functools import partial
2
+
3
+import numpy as np
4
+import pytest
5
+from PIL import Image
6
+
7
+from vllm.config import ModelConfig
8
+from vllm.inputs import InputProcessingContext
9
+from vllm.multimodal import MULTIMODAL_REGISTRY
10
+from vllm.multimodal.processing import ProcessingCache
11
+from vllm.multimodal.utils import cached_get_tokenizer
12
+
13
+from ....multimodal.utils import random_audio, random_image, random_video
14
+
15
+
16
+def _test_processing_correctness(
17
+ model_id: str,
18
+ modalities: dict[str, bool],
19
+ hit_rate: float,
20
+ num_batches: int,
21
+ simplify_rate: float,
22
+):
23
+ if model_id == "TIGER-Lab/Mantis-8B-siglip-llama3":
24
+ hf_overrides = {"architectures": ["MantisForConditionalGeneration"]}
25
+ else:
26
+ hf_overrides = {}
27
+
28
+ limit_mm_per_prompt = {
29
+ modality: 3 if supports_multi else 1
30
+ for modality, supports_multi in modalities.items()
31
+ }
32
+
33
+ model_config = ModelConfig(
34
+ model_id,
35
+ task="auto",
36
+ tokenizer=model_id,
37
+ tokenizer_mode="auto",
38
+ trust_remote_code=True,
39
+ seed=0,
40
+ dtype="float16",
41
+ revision=None,
42
+ hf_overrides=hf_overrides,
43
+ limit_mm_per_prompt=limit_mm_per_prompt,
44
+ )
45
+
46
+ model_cls = MULTIMODAL_REGISTRY._get_model_cls(model_config)
47
+ factories = MULTIMODAL_REGISTRY._processor_factories[model_cls]
48
+ ctx = InputProcessingContext(
49
+ model_config,
50
+ tokenizer=cached_get_tokenizer(model_config.tokenizer),
51
+ )
52
+ # Ensure that it can fit all of the data
53
+ cache = ProcessingCache(capacity=1 << 30)
54
+
55
+ baseline_processor = factories.build_processor(ctx, cache=None)
56
+ cached_processor = factories.build_processor(ctx, cache=cache)
57
+ dummy_inputs = baseline_processor.dummy_inputs
58
+ tokenizer = baseline_processor.info.get_tokenizer()
59
+
60
+ rng = np.random.RandomState(0)
61
+
62
+ input_to_hit = {
63
+ "image": Image.new("RGB", size=(128, 128)),
64
+ "video": np.zeros((4, 128, 128, 3), dtype=np.uint8),
65
+ "audio": (np.zeros((512, )), 16000),
66
+ }
67
+ input_factory = {
68
+ "image":
69
+ partial(random_image, rng, min_wh=128, max_wh=256),
70
+ "video":
71
+ partial(random_video,
72
+ rng,
73
+ min_frames=2,
74
+ max_frames=8,
75
+ min_wh=128,
76
+ max_wh=256),
77
+ "audio":
78
+ partial(random_audio, rng, min_len=512, max_len=1024, sr=16000),
79
+ }
80
+
81
+ for batch_idx in range(num_batches):
82
+ mm_data = {
83
+ k:
84
+ [(input_to_hit[k] if rng.rand() < hit_rate else input_factory[k]())
85
+ for _ in range(rng.randint(limit_mm_per_prompt[k]))]
86
+ for k in modalities
87
+ }
88
+
89
+ mm_counts = {k: len(vs) for k, vs in mm_data.items()}
90
+ prompt = dummy_inputs.get_dummy_processor_inputs(
91
+ model_config.max_model_len,
92
+ mm_counts,
93
+ ).prompt_text
94
+
95
+ # Drop unnecessary keys and test single -> multi conversion
96
+ if rng.rand() < simplify_rate:
97
+ for k in list(mm_data.keys()):
98
+ if not mm_data[k]:
99
+ del mm_data[k]
100
+ elif len(mm_data[k]) == 1:
101
+ mm_data[k] = mm_data[k][0]
102
+
103
+ baseline_result = baseline_processor.apply(
104
+ prompt,
105
+ mm_data=mm_data,
106
+ hf_processor_mm_kwargs={},
107
+ )
108
+ cached_result = cached_processor.apply(
109
+ prompt,
110
+ mm_data=mm_data,
111
+ hf_processor_mm_kwargs={},
112
+ )
113
+
114
+ assert baseline_result == cached_result, (
115
+ f"Failed ({batch_idx=}, {prompt=}, {mm_data=})")
116
+
117
+ baseline_tokenized_result = baseline_processor.apply(
118
+ tokenizer.encode(prompt),
119
+ mm_data=mm_data,
120
+ hf_processor_mm_kwargs={},
121
+ )
122
+
123
+ assert baseline_result == baseline_tokenized_result, (
124
+ f"Failed ({batch_idx=}, {prompt=}, {mm_data=})")
125
+
126
+ cached_tokenized_result = cached_processor.apply(
127
+ tokenizer.encode(prompt),
128
+ mm_data=mm_data,
129
+ hf_processor_mm_kwargs={},
130
+ )
131
+
132
+ assert cached_result == cached_tokenized_result, (
133
+ f"Failed ({batch_idx=}, {prompt=}, {mm_data=})")
134
+
135
+
136
+# yapf: disable
137
+# True if the model supports multiple data items of the modality per request
138
+@pytest.mark.parametrize(("model_id", "modalities"), [
139
+ ("rhymes-ai/Aria", {"image": True}),
140
+ ("Salesforce/blip2-opt-2.7b", {"image": False}),
141
+ ("facebook/chameleon-7b", {"image": False}),
142
+ ("adept/fuyu-8b", {"image": False}),
143
+ ("llava-hf/llava-1.5-7b-hf", {"image": True}),
144
+ ("llava-hf/llava-v1.6-mistral-7b-hf", {"image": True}),
145
+ ("llava-hf/LLaVA-NeXT-Video-7B-hf", {"video": False}),
146
+ ("llava-hf/llava-onevision-qwen2-0.5b-ov-hf", {"image": True, "video": True}), # noqa: E501
147
+ ("TIGER-Lab/Mantis-8B-siglip-llama3", {"image": True}),
148
+ ("mistral-community/pixtral-12b", {"image": True}),
149
+ ("Qwen/Qwen2-VL-2B-Instruct", {"image": True, "video": True}),
150
+ ("Qwen/Qwen2-Audio-7B-Instruct", {"audio": True}),
151
+ ("fixie-ai/ultravox-v0_3", {"audio": True}),
152
+])
153
+@pytest.mark.parametrize("hit_rate", [0.3, 0.5, 1.0])
154
+@pytest.mark.parametrize("num_batches", [32])
155
+@pytest.mark.parametrize("simplify_rate", [1.0])
156
+# yapf: enable
157
+def test_processing_correctness(
158
+ model_id: str,
159
+ modalities: dict[str, bool],
160
+ hit_rate: float,
161
+ num_batches: int,
162
+ simplify_rate: float,
163
+):
164
+ _test_processing_correctness(
165
+ model_id,
166
+ modalities,
167
+ hit_rate=hit_rate,
168
+ num_batches=num_batches,
169
+ simplify_rate=simplify_rate,
170
+ )
171
+
172
+
173
+# yapf: disable
174
+@pytest.mark.parametrize(("model_id", "modalities"), [
175
+ ("microsoft/Phi-3-vision-128k-instruct", {"image": True}),
176
+])
177
+@pytest.mark.parametrize("hit_rate", [0.3, 0.5, 1.0])
178
+@pytest.mark.parametrize("num_batches", [32])
179
+@pytest.mark.parametrize("simplify_rate", [1.0])
180
+# yapf: enable
181
+def test_processing_correctness_phi3v(
182
+ model_id: str,
183
+ modalities: dict[str, bool],
184
+ hit_rate: float,
185
+ num_batches: int,
186
+ simplify_rate: float,
187
+):
188
+ # HACK - this is an attempted workaround for the following bug
189
+ # https://github.com/huggingface/transformers/issues/34307
190
+ from transformers import AutoImageProcessor # noqa: F401
191
+ from transformers import AutoProcessor # noqa: F401
192
+
193
+ AutoImageProcessor.from_pretrained(model_id, trust_remote_code=True)
194
+
195
+ _test_processing_correctness(
196
+ model_id,
197
+ modalities,
198
+ hit_rate=hit_rate,
199
+ num_batches=num_batches,
200
+ simplify_rate=simplify_rate,
201
+ )
tests/models/multimodal/processing/test_idefics3.py
renamed
+2
-2
@@ -8,8 +8,8 @@ from transformers import AutoImageProcessor, AutoTokenizer
8
from vllm.inputs import InputContext, token_inputs
9
from vllm.multimodal import MultiModalRegistry
10
11
-from .....conftest import _ImageAssets
12
-from ....utils import build_model_context
11
+from ....conftest import _ImageAssets
12
+from ...utils import build_model_context
13
14
models = ["HuggingFaceM4/Idefics3-8B-Llama3"]
15
tests/models/multimodal/processing/test_internvl.py
renamed
+2
-2
@@ -7,8 +7,8 @@ from transformers import AutoTokenizer
7
from vllm.inputs import InputContext, token_inputs
8
from vllm.multimodal import MultiModalRegistry
9
10
-from .....conftest import _ImageAssets
11
-from ....utils import build_model_context
10
+from ....conftest import _ImageAssets
11
+from ...utils import build_model_context
12
13
models = ["OpenGVLab/InternVL2-2B"]
14
tests/models/multimodal/processing/test_llava_next.py
renamed
+1
-1
@@ -10,7 +10,7 @@ from vllm.multimodal.parse import ImageSize
10
from vllm.multimodal.processing import BaseMultiModalProcessor
11
from vllm.multimodal.utils import cached_get_tokenizer
12
13
-from ....utils import build_model_context
13
+from ...utils import build_model_context
14
15
16
def _validate_image_prompt_replacements_one(
tests/models/multimodal/processing/test_llava_onevision.py
renamed
+1
-1
@@ -10,7 +10,7 @@ from vllm.multimodal.parse import ImageSize
10
from vllm.multimodal.processing import BaseMultiModalProcessor
11
from vllm.multimodal.utils import cached_get_tokenizer
12
13
-from ....utils import build_model_context
13
+from ...utils import build_model_context
14
15
16
def _validate_image_prompt_replacements_one(
tests/models/multimodal/processing/test_phi3v.py
renamed
+2
-2
@@ -4,8 +4,8 @@ import pytest
4
from vllm.multimodal import MULTIMODAL_REGISTRY
5
from vllm.multimodal.utils import cached_get_tokenizer
6
7
-from .....conftest import _ImageAssets
8
-from ....utils import build_model_context
7
+from ....conftest import _ImageAssets
8
+from ...utils import build_model_context
9
10
11
@pytest.mark.parametrize("model_id", ["microsoft/Phi-3.5-vision-instruct"])
tests/models/multimodal/processing/test_qwen.py
renamed
+2
-2
@@ -9,8 +9,8 @@ from vllm.inputs import InputContext, token_inputs
9
from vllm.multimodal import MultiModalKwargs
10
from vllm.multimodal.utils import cached_get_tokenizer
11
12
-from .....conftest import IMAGE_ASSETS
13
-from ....utils import build_model_context
12
+from ....conftest import IMAGE_ASSETS
13
+from ...utils import build_model_context
14
15
### Multimodal preprocessing tests
16
SAMPLE_IMAGE = IMAGE_ASSETS[0].pil_image
tests/models/multimodal/processing/test_qwen2_vl.py
renamed
+2
-2
@@ -3,8 +3,8 @@ import pytest
3
from vllm.multimodal import MULTIMODAL_REGISTRY
4
from vllm.multimodal.utils import cached_get_tokenizer
5
6
-from .....conftest import _ImageAssets
7
-from ....utils import build_model_context
6
+from ....conftest import _ImageAssets
7
+from ...utils import build_model_context
8
9
10
@pytest.mark.parametrize("model_id", ["Qwen/Qwen2-VL-2B-Instruct"])
tests/multimodal/test_processing.py
+4
-228
@@ -1,30 +1,25 @@
1
from contextlib import nullcontext
2
-from functools import partial
2
from typing import cast
3
from unittest.mock import MagicMock
4
5
import numpy as np
6
import pytest
8
-from PIL import Image
7
8
from vllm.config import ModelConfig
11
-from vllm.inputs import InputProcessingContext
9
from vllm.multimodal import MULTIMODAL_REGISTRY
13
-# yapf conflicts with isort for this block
14
-# yapf: disable
15
-from vllm.multimodal.processing import (PlaceholderInfo, ProcessingCache,
16
- PromptReplacement,
10
+from vllm.multimodal.processing import (PlaceholderInfo, PromptReplacement,
11
find_mm_placeholders,
12
find_text_matches, find_token_matches,
13
iter_token_matches,
14
replace_text_matches,
15
replace_token_matches)
22
-# yapf: enable
16
from vllm.multimodal.profiling import MultiModalProfiler
17
from vllm.multimodal.utils import cached_get_tokenizer
18
from vllm.transformers_utils.tokenizer import AnyTokenizer
19
from vllm.utils import full_groupby
20
21
+from .utils import random_image
22
+
23
24
# yapf: disable
25
@pytest.mark.parametrize(
@@ -531,37 +526,6 @@ def test_find_mm_placeholders(
526
assert result == expected
527
528
534
-def _rand_img(rng: np.random.RandomState, min_wh: int, max_wh: int):
535
- w, h = rng.randint(min_wh, max_wh, size=(2, ))
536
- arr = rng.randint(0, 255, size=(w, h, 3), dtype=np.uint8)
537
- return Image.fromarray(arr)
538
-
539
-
540
-def _rand_video(
541
- rng: np.random.RandomState,
542
- min_frames: int,
543
- max_frames: int,
544
- min_wh: int,
545
- max_wh: int,
546
-):
547
- # Temporary workaround for https://github.com/huggingface/transformers/issues/35412
548
- num_frames = rng.randint(min_frames, max_frames)
549
- num_frames = (num_frames // 2) * 2
550
-
551
- w, h = rng.randint(min_wh, max_wh, size=(2, ))
552
- return rng.randint(0, 255, size=(num_frames, w, h, 3), dtype=np.uint8)
553
-
554
-
555
-def _rand_audio(
556
- rng: np.random.RandomState,
557
- min_len: int,
558
- max_len: int,
559
- sr: int,
560
-):
561
- audio_len = rng.randint(min_len, max_len)
562
- return rng.rand(audio_len), sr
563
-
564
-
529
@pytest.mark.parametrize("model_id", ["llava-hf/llava-v1.6-mistral-7b-hf"])
530
@pytest.mark.parametrize(
531
("limit", "num_supported", "is_valid"),
@@ -628,7 +592,7 @@ def test_limit_mm_per_prompt_apply(model_id, num_images, limit, is_valid):
592
)
593
594
rng = np.random.RandomState(0)
631
- image = _rand_img(rng, min_wh=128, max_wh=256)
595
+ image = random_image(rng, min_wh=128, max_wh=256)
596
if num_images == 0:
597
mm_data = {}
598
elif num_images == 1:
@@ -647,191 +611,3 @@ def test_limit_mm_per_prompt_apply(model_id, num_images, limit, is_valid):
611
mm_data=mm_data,
612
hf_processor_mm_kwargs={},
613
)
650
-
651
-
652
-def _test_processing_correctness(
653
- model_id: str,
654
- modalities: dict[str, bool],
655
- hit_rate: float,
656
- num_batches: int,
657
- simplify_rate: float,
658
-):
659
- if model_id == "TIGER-Lab/Mantis-8B-siglip-llama3":
660
- hf_overrides = {"architectures": ["MantisForConditionalGeneration"]}
661
- else:
662
- hf_overrides = {}
663
-
664
- limit_mm_per_prompt = {
665
- modality: 3 if supports_multi else 1
666
- for modality, supports_multi in modalities.items()
667
- }
668
-
669
- model_config = ModelConfig(
670
- model_id,
671
- task="auto",
672
- tokenizer=model_id,
673
- tokenizer_mode="auto",
674
- trust_remote_code=True,
675
- seed=0,
676
- dtype="float16",
677
- revision=None,
678
- hf_overrides=hf_overrides,
679
- limit_mm_per_prompt=limit_mm_per_prompt,
680
- )
681
-
682
- model_cls = MULTIMODAL_REGISTRY._get_model_cls(model_config)
683
- factories = MULTIMODAL_REGISTRY._processor_factories[model_cls]
684
- ctx = InputProcessingContext(
685
- model_config,
686
- tokenizer=cached_get_tokenizer(model_config.tokenizer),
687
- )
688
- # Ensure that it can fit all of the data
689
- cache = ProcessingCache(capacity=1 << 30)
690
-
691
- baseline_processor = factories.build_processor(ctx, cache=None)
692
- cached_processor = factories.build_processor(ctx, cache=cache)
693
- dummy_inputs = baseline_processor.dummy_inputs
694
- tokenizer = baseline_processor.info.get_tokenizer()
695
-
696
- rng = np.random.RandomState(0)
697
-
698
- input_to_hit = {
699
- "image": Image.new("RGB", size=(128, 128)),
700
- "video": np.zeros((4, 128, 128, 3), dtype=np.uint8),
701
- "audio": (np.zeros((512, )), 16000),
702
- }
703
- input_factory = {
704
- "image":
705
- partial(_rand_img, rng, min_wh=128, max_wh=256),
706
- "video":
707
- partial(_rand_video,
708
- rng,
709
- min_frames=2,
710
- max_frames=8,
711
- min_wh=128,
712
- max_wh=256),
713
- "audio":
714
- partial(_rand_audio, rng, min_len=512, max_len=1024, sr=16000),
715
- }
716
-
717
- for batch_idx in range(num_batches):
718
- mm_data = {
719
- k:
720
- [(input_to_hit[k] if rng.rand() < hit_rate else input_factory[k]())
721
- for _ in range(rng.randint(limit_mm_per_prompt[k]))]
722
- for k in modalities
723
- }
724
-
725
- mm_counts = {k: len(vs) for k, vs in mm_data.items()}
726
- prompt = dummy_inputs.get_dummy_processor_inputs(
727
- model_config.max_model_len,
728
- mm_counts,
729
- ).prompt_text
730
-
731
- # Drop unnecessary keys and test single -> multi conversion
732
- if rng.rand() < simplify_rate:
733
- for k in list(mm_data.keys()):
734
- if not mm_data[k]:
735
- del mm_data[k]
736
- elif len(mm_data[k]) == 1:
737
- mm_data[k] = mm_data[k][0]
738
-
739
- baseline_result = baseline_processor.apply(
740
- prompt,
741
- mm_data=mm_data,
742
- hf_processor_mm_kwargs={},
743
- )
744
- cached_result = cached_processor.apply(
745
- prompt,
746
- mm_data=mm_data,
747
- hf_processor_mm_kwargs={},
748
- )
749
-
750
- assert baseline_result == cached_result, (
751
- f"Failed ({batch_idx=}, {prompt=}, {mm_data=})")
752
-
753
- baseline_tokenized_result = baseline_processor.apply(
754
- tokenizer.encode(prompt),
755
- mm_data=mm_data,
756
- hf_processor_mm_kwargs={},
757
- )
758
-
759
- assert baseline_result == baseline_tokenized_result, (
760
- f"Failed ({batch_idx=}, {prompt=}, {mm_data=})")
761
-
762
- cached_tokenized_result = cached_processor.apply(
763
- tokenizer.encode(prompt),
764
- mm_data=mm_data,
765
- hf_processor_mm_kwargs={},
766
- )
767
-
768
- assert cached_result == cached_tokenized_result, (
769
- f"Failed ({batch_idx=}, {prompt=}, {mm_data=})")
770
-
771
-
772
-# yapf: disable
773
-# True if the model supports multiple data items of the modality per request
774
-@pytest.mark.parametrize(("model_id", "modalities"), [
775
- ("rhymes-ai/Aria", {"image": True}),
776
- ("Salesforce/blip2-opt-2.7b", {"image": False}),
777
- ("facebook/chameleon-7b", {"image": False}),
778
- ("adept/fuyu-8b", {"image": False}),
779
- ("llava-hf/llava-1.5-7b-hf", {"image": True}),
780
- ("llava-hf/llava-v1.6-mistral-7b-hf", {"image": True}),
781
- ("llava-hf/LLaVA-NeXT-Video-7B-hf", {"video": False}),
782
- ("llava-hf/llava-onevision-qwen2-0.5b-ov-hf", {"image": True, "video": True}), # noqa: E501
783
- ("TIGER-Lab/Mantis-8B-siglip-llama3", {"image": True}),
784
- ("mistral-community/pixtral-12b", {"image": True}),
785
- ("Qwen/Qwen2-VL-2B-Instruct", {"image": True, "video": True}),
786
- ("Qwen/Qwen2-Audio-7B-Instruct", {"audio": True}),
787
- ("fixie-ai/ultravox-v0_3", {"audio": True}),
788
-])
789
-@pytest.mark.parametrize("hit_rate", [0.3, 0.5, 1.0])
790
-@pytest.mark.parametrize("num_batches", [32])
791
-@pytest.mark.parametrize("simplify_rate", [1.0])
792
-# yapf: enable
793
-def test_processing_correctness(
794
- model_id: str,
795
- modalities: dict[str, bool],
796
- hit_rate: float,
797
- num_batches: int,
798
- simplify_rate: float,
799
-):
800
- _test_processing_correctness(
801
- model_id,
802
- modalities,
803
- hit_rate=hit_rate,
804
- num_batches=num_batches,
805
- simplify_rate=simplify_rate,
806
- )
807
-
808
-
809
-# yapf: disable
810
-@pytest.mark.parametrize(("model_id", "modalities"), [
811
- ("microsoft/Phi-3-vision-128k-instruct", {"image": True}),
812
-])
813
-@pytest.mark.parametrize("hit_rate", [0.3, 0.5, 1.0])
814
-@pytest.mark.parametrize("num_batches", [32])
815
-@pytest.mark.parametrize("simplify_rate", [1.0])
816
-# yapf: enable
817
-def test_processing_correctness_phi3v(
818
- model_id: str,
819
- modalities: dict[str, bool],
820
- hit_rate: float,
821
- num_batches: int,
822
- simplify_rate: float,
823
-):
824
- # HACK - this is an attempted workaround for the following bug
825
- # https://github.com/huggingface/transformers/issues/34307
826
- from transformers import AutoImageProcessor # noqa: F401
827
- from transformers import AutoProcessor # noqa: F401
828
-
829
- AutoImageProcessor.from_pretrained(model_id, trust_remote_code=True)
830
-
831
- _test_processing_correctness(
832
- model_id,
833
- modalities,
834
- hit_rate=hit_rate,
835
- num_batches=num_batches,
836
- simplify_rate=simplify_rate,
837
- )
tests/multimodal/utils.py
new
+33
@@ -0,0 +1,33 @@
1
+import numpy as np
2
+from PIL import Image
3
+
4
+
5
+def random_image(rng: np.random.RandomState, min_wh: int, max_wh: int):
6
+ w, h = rng.randint(min_wh, max_wh, size=(2, ))
7
+ arr = rng.randint(0, 255, size=(w, h, 3), dtype=np.uint8)
8
+ return Image.fromarray(arr)
9
+
10
+
11
+def random_video(
12
+ rng: np.random.RandomState,
13
+ min_frames: int,
14
+ max_frames: int,
15
+ min_wh: int,
16
+ max_wh: int,
17
+):
18
+ # Temporary workaround for https://github.com/huggingface/transformers/issues/35412
19
+ num_frames = rng.randint(min_frames, max_frames)
20
+ num_frames = (num_frames // 2) * 2
21
+
22
+ w, h = rng.randint(min_wh, max_wh, size=(2, ))
23
+ return rng.randint(0, 255, size=(num_frames, w, h, 3), dtype=np.uint8)
24
+
25
+
26
+def random_audio(
27
+ rng: np.random.RandomState,
28
+ min_len: int,
29
+ max_len: int,
30
+ sr: int,
31
+):
32
+ audio_len = rng.randint(min_len, max_len)
33
+ return rng.rand(audio_len), sr