Add pydocstyle to the ruff rules (#52136)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Signed-off-by: Harry Mellor <hej.mellor@gmail.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5
parent
804e5377aa
commit
8c1557a79c
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Regression tests for Qwen2.5-Omni and Qwen3-Omni audio-in-video processor
|
||||
"""Regression tests for Qwen2.5-Omni and Qwen3-Omni audio-in-video processor
|
||||
caching.
|
||||
|
||||
Tests the use_audio_in_video feature where audio is extracted from video and
|
||||
@@ -50,8 +49,7 @@ def create_mm_data(num_videos: int) -> dict[str, list]:
|
||||
@pytest.mark.parametrize("model_id", MODELS)
|
||||
@pytest.mark.parametrize("num_videos", [1, 2])
|
||||
def test_audio_in_video_cache_correctness(model_id: str, num_videos: int) -> None:
|
||||
"""
|
||||
Regression test for https://github.com/vllm-project/vllm/pull/36800
|
||||
"""Regression test for https://github.com/vllm-project/vllm/pull/36800.
|
||||
|
||||
MultiModalProcessorSenderCache.get_and_update_item returns (None, updates)
|
||||
on a cache hit, so mm_kwargs["video"] items become None on the second call.
|
||||
@@ -117,8 +115,7 @@ def test_audio_in_video_cache_correctness(model_id: str, num_videos: int) -> Non
|
||||
|
||||
@pytest.mark.parametrize("model_id", MODELS)
|
||||
def test_use_audio_in_video_without_audio_track(model_id: str) -> None:
|
||||
"""
|
||||
A video with no audio track, combined with `use_audio_in_video=True`,
|
||||
"""A video with no audio track, combined with `use_audio_in_video=True`,
|
||||
must raise a clear `ValueError` naming the mismatch, rather than an
|
||||
opaque `StopIteration`-derived error from the underlying HF processor.
|
||||
"""
|
||||
|
||||
@@ -33,9 +33,7 @@ from ...registry import (
|
||||
|
||||
|
||||
def add_video_metadata(mm_data: MultiModalDataDict) -> MultiModalDataDict:
|
||||
"""
|
||||
Add metadata to video mm_data
|
||||
"""
|
||||
"""Add metadata to video mm_data."""
|
||||
|
||||
def create_metadata(frames: np.ndarray):
|
||||
num_frames = len(frames)
|
||||
@@ -61,8 +59,7 @@ def add_video_metadata(mm_data: MultiModalDataDict) -> MultiModalDataDict:
|
||||
|
||||
|
||||
def glmasr_patch_mm_data(mm_data: MultiModalDataDict) -> MultiModalDataDict:
|
||||
"""
|
||||
Patch the multimodal data for GLM-ASR model.
|
||||
"""Patch the multimodal data for GLM-ASR model.
|
||||
GLM-ASR requires text and audio to match 1:1, so we limit audio to 1.
|
||||
"""
|
||||
if "audio" in mm_data:
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Regression test for DeepSeek-OCR TensorSchema validation with empty images_crop.
|
||||
"""Regression test for DeepSeek-OCR TensorSchema validation with empty images_crop.
|
||||
|
||||
When using the Gundam preset (BASE_SIZE=1024, IMAGE_SIZE=640, CROP_MODE=True),
|
||||
images that are small enough to not require cropping produce an empty
|
||||
@@ -109,7 +108,8 @@ class TestDeepseekOCREmptyImagesCrop:
|
||||
|
||||
def test_mismatched_image_size_raises(self, processor):
|
||||
"""Deliberately wrong image_size binding should still be caught
|
||||
by TensorSchema validation."""
|
||||
by TensorSchema validation.
|
||||
"""
|
||||
small_image = Image.new("RGB", (100, 100), color="green")
|
||||
|
||||
result = processor(
|
||||
@@ -143,7 +143,8 @@ class TestDeepseekOCRInputValidation:
|
||||
def test_mismatched_image_count_raises_value_error(self, processor):
|
||||
"""Two ``<image>`` tokens with one image must raise ValueError, not
|
||||
AssertionError (which would yield HTTP 500 and, under ``-O``, silently
|
||||
drop the middle text segment)."""
|
||||
drop the middle text segment).
|
||||
"""
|
||||
image = Image.new("RGB", (100, 100), color="red")
|
||||
with pytest.raises(ValueError, match="does not match number of images"):
|
||||
processor(prompt="<image> and <image>", images=[image])
|
||||
|
||||
@@ -152,8 +152,7 @@ def test_video_loader_consistency(
|
||||
fps: int,
|
||||
backend: str,
|
||||
):
|
||||
"""
|
||||
Ensure dynamic video loader (pre-sampled by loader) and normal video
|
||||
"""Ensure dynamic video loader (pre-sampled by loader) and normal video
|
||||
loader (post-sampled by processor) produce same video processing outputs.
|
||||
"""
|
||||
ctx = build_model_context(
|
||||
|
||||
@@ -116,8 +116,7 @@ def _test_image_prompt_replacements(
|
||||
num_imgs: int,
|
||||
image_sizes: list[ImageSize],
|
||||
) -> None:
|
||||
"""
|
||||
Ensure LlavaNextMultiModalProcessor
|
||||
"""Ensure LlavaNextMultiModalProcessor
|
||||
handles prompt replacement properly for input images.
|
||||
"""
|
||||
failed_size_excs = list[tuple[ImageSize, Exception]]()
|
||||
|
||||
@@ -16,7 +16,8 @@ from ...utils import build_model_context
|
||||
class _StubModel:
|
||||
"""Carries only the attributes the token-count method reads from
|
||||
`self`, so the real methods can be exercised without constructing
|
||||
the full `nn.Module` (vision tower, language model, etc.)."""
|
||||
the full `nn.Module` (vision tower, language model, etc.).
|
||||
"""
|
||||
|
||||
patch_grid_length: int
|
||||
pooled_grid_length: int
|
||||
@@ -31,7 +32,8 @@ get_mm_lora_token_counts = (
|
||||
def test_num_mm_tokens_match_real_config(model_id):
|
||||
"""The stored grid lengths must match what `__init__` derives from
|
||||
the real HF config, and the LoRA token counts must preserve the
|
||||
frame-level scaling exactly."""
|
||||
frame-level scaling exactly.
|
||||
"""
|
||||
ctx = build_model_context(model_id, limit_mm_per_prompt={"video": 1})
|
||||
config = ctx.model_config.hf_config
|
||||
|
||||
|
||||
@@ -122,8 +122,7 @@ def _test_image_prompt_replacements(
|
||||
num_imgs: int,
|
||||
image_sizes: list[ImageSize],
|
||||
) -> None:
|
||||
"""
|
||||
Ensure LlavaOnevisionMultiModalProcessor
|
||||
"""Ensure LlavaOnevisionMultiModalProcessor
|
||||
handles prompt replacement properly for input images.
|
||||
"""
|
||||
failed_size_excs = list[tuple[ImageSize, Exception]]()
|
||||
|
||||
@@ -70,7 +70,8 @@ def test_dummy_video_num_frames_override_honors_min_of_two(
|
||||
):
|
||||
"""A ``num_frames`` override below 2 must be ignored (the model needs at
|
||||
least 2 frames), matching the "cannot be less than 2, will be ignored"
|
||||
warning."""
|
||||
warning.
|
||||
"""
|
||||
builder = object.__new__(Molmo2DummyInputsBuilder)
|
||||
builder.info = SimpleNamespace(
|
||||
get_hf_processor=lambda: SimpleNamespace(
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Unit tests for Qwen2.5-Omni embed_input_ids to verify embeddings are
|
||||
"""Unit tests for Qwen2.5-Omni embed_input_ids to verify embeddings are
|
||||
correctly assigned to audio/image/video token positions.
|
||||
|
||||
Regression test for: https://github.com/vllm-project/vllm/issues/34506
|
||||
@@ -40,8 +39,7 @@ def _mm_embed(shape: tuple[int, ...], value: float, modality: str) -> torch.Tens
|
||||
def make_token_seq(
|
||||
audio_n: int, image_n: int, video_n: int, text_prefix: int = 3, text_sep: int = 2
|
||||
):
|
||||
"""
|
||||
Build a flat token sequence:
|
||||
"""Build a flat token sequence:
|
||||
[text_prefix] [AUDIO * audio_n] [text_sep] [IMAGE * image_n]
|
||||
[text_sep] [VIDEO * video_n] [text_sep]
|
||||
Returns (input_ids tensor, is_multimodal mask, positions dict).
|
||||
@@ -67,9 +65,8 @@ def make_token_seq(
|
||||
def make_interleaved_seq(
|
||||
video_chunks: list[int], audio_chunks: list[int], text_prefix: int = 2
|
||||
):
|
||||
"""
|
||||
Build an interleaved sequence like use_audio_in_video:
|
||||
[text] [V*v0] [A*a0] [V*v1] [A*a1] ...
|
||||
"""Build an interleaved sequence like use_audio_in_video:
|
||||
[text] [V*v0] [A*a0] [V*v1] [A*a1] ...
|
||||
"""
|
||||
tokens = [TEXT_TOKEN_ID] * text_prefix
|
||||
for v, a in zip(video_chunks, audio_chunks):
|
||||
@@ -140,8 +137,7 @@ class TestCheckInterleavedAudioVideo:
|
||||
)
|
||||
|
||||
def test_batched_non_interleaved_no_false_positive(self):
|
||||
"""
|
||||
Regression test for https://github.com/vllm-project/vllm/issues/35394.
|
||||
"""Regression test for https://github.com/vllm-project/vllm/issues/35394.
|
||||
|
||||
5 identical non-interleaved mixed-modality requests batched together:
|
||||
each has [audio][image][video] in separate blocks with text between them.
|
||||
@@ -172,8 +168,7 @@ class TestCheckInterleavedAudioVideo:
|
||||
|
||||
|
||||
def make_mock_model(hidden: int = 8):
|
||||
"""
|
||||
Return a minimal mock of Qwen2_5OmniThinkerForConditionalGeneration
|
||||
"""Return a minimal mock of Qwen2_5OmniThinkerForConditionalGeneration
|
||||
that has enough structure to run embed_input_ids.
|
||||
"""
|
||||
from vllm.model_executor.models.qwen2_5_omni_thinker import (
|
||||
@@ -236,8 +231,7 @@ def make_mock_model(hidden: int = 8):
|
||||
def build_mm_embeds(
|
||||
audio_n, image_n, video_n, hidden, audio_val=10.0, image_val=20.0, video_val=30.0
|
||||
):
|
||||
"""
|
||||
Build multimodal_embeddings list in position order (audio, image, video).
|
||||
"""Build multimodal_embeddings list in position order (audio, image, video).
|
||||
Each embedding is filled with a distinct constant so we can verify placement.
|
||||
"""
|
||||
embs = []
|
||||
@@ -252,8 +246,7 @@ def build_mm_embeds(
|
||||
|
||||
class TestEmbedInputIds:
|
||||
def _run(self, audio_n, image_n, video_n, hidden=8):
|
||||
"""
|
||||
Run embed_input_ids for a non-interleaved mixed-modality sequence.
|
||||
"""Run embed_input_ids for a non-interleaved mixed-modality sequence.
|
||||
Returns (result_embeds, input_ids, is_multimodal).
|
||||
"""
|
||||
input_ids, is_multimodal = make_token_seq(audio_n, image_n, video_n)
|
||||
@@ -288,8 +281,7 @@ class TestEmbedInputIds:
|
||||
)
|
||||
|
||||
def test_mixed_modalities_audio_goes_to_audio_pos(self):
|
||||
"""
|
||||
Regression test for GitHub issue #34506:
|
||||
"""Regression test for GitHub issue #34506:
|
||||
With audio + image + video (non-interleaved), audio positions must
|
||||
receive audio embeddings (not image or video embeddings).
|
||||
"""
|
||||
@@ -329,8 +321,7 @@ class TestEmbedInputIds:
|
||||
)
|
||||
|
||||
def test_interleaved_use_audio_in_video(self):
|
||||
"""
|
||||
Interleaved (use_audio_in_video): video chunks interleaved with audio.
|
||||
"""Interleaved (use_audio_in_video): video chunks interleaved with audio.
|
||||
Video embeddings must go to video positions, audio to audio positions.
|
||||
"""
|
||||
hidden = 8
|
||||
|
||||
@@ -232,7 +232,6 @@ def test_mm_device_do_normalize(
|
||||
num_imgs: int,
|
||||
):
|
||||
"""Ensure that enable mm_device_do_normalize yields the correct result."""
|
||||
|
||||
ctx = build_model_context(
|
||||
model_id,
|
||||
limit_mm_per_prompt={"image": num_imgs},
|
||||
|
||||
@@ -25,8 +25,7 @@ def test_processor_with_audio_sample_rate(
|
||||
audio_sample_rate: int,
|
||||
audio_duration_sec: float,
|
||||
) -> None:
|
||||
"""
|
||||
Test that vLLM's processor generates expected outputs with audio_sample_rate.
|
||||
"""Test that vLLM's processor generates expected outputs with audio_sample_rate.
|
||||
|
||||
This validates that the processor correctly handles audio_sample_rate
|
||||
passed via hf_processor_mm_kwargs and generates audio tokens.
|
||||
@@ -70,8 +69,7 @@ def test_processor_with_audio_sample_rate(
|
||||
|
||||
@pytest.mark.parametrize("model_id", ["Qwen/Qwen3-Omni-30B-A3B-Instruct"])
|
||||
def test_longer_audio_generates_more_tokens(model_id: str) -> None:
|
||||
"""
|
||||
Test that longer audio generates more tokens than shorter audio.
|
||||
"""Test that longer audio generates more tokens than shorter audio.
|
||||
|
||||
This validates that audio_sample_rate is being used correctly by checking
|
||||
that audio duration affects token count as expected.
|
||||
|
||||
@@ -31,7 +31,8 @@ def _build_video_mm_data(
|
||||
|
||||
``total_num_frames`` is set equal to the ndarray frame count so
|
||||
that HF's ``sample_frames`` indices stay within bounds of the
|
||||
actual tensor that is passed."""
|
||||
actual tensor that is passed.
|
||||
"""
|
||||
video = np.zeros((num_frames, height, width, 3), dtype=np.uint8)
|
||||
metadata = {
|
||||
"fps": original_fps,
|
||||
@@ -237,7 +238,8 @@ def _build_video_embeds_mm_data(
|
||||
grid_thw: tuple[int, int, int] = (2, 4, 4),
|
||||
) -> dict[str, Any]:
|
||||
"""Create an embeds-only video item as an EPD consumer receives it:
|
||||
pre-computed embeddings plus the metadata published by the encoder."""
|
||||
pre-computed embeddings plus the metadata published by the encoder.
|
||||
"""
|
||||
import torch
|
||||
|
||||
t, h, w = grid_thw
|
||||
@@ -256,7 +258,8 @@ def test_processor_video_embeds_with_timestamps(model_id: str) -> None:
|
||||
"""Embeds-only video input must size the placeholder range from the
|
||||
grid and the real timestamps published by the encoder (EC consumer
|
||||
path); synthesized or missing timestamps would change the token count
|
||||
and break embedding merging downstream."""
|
||||
and break embedding merging downstream.
|
||||
"""
|
||||
# `build_model_context` forces enable_mm_embeds off, so build the
|
||||
# config directly; keep the same online-availability skip behavior.
|
||||
HF_EXAMPLE_MODELS.find_hf_info(model_id).check_available_online(on_fail="skip")
|
||||
@@ -301,7 +304,8 @@ def test_processor_video_embeds_with_timestamps(model_id: str) -> None:
|
||||
def test_processor_video_embeds_missing_timestamps(model_id: str) -> None:
|
||||
"""Timestamps are required metadata for video embeds: they size the
|
||||
placeholder range, so omitting them must fail loudly at parse time
|
||||
instead of silently producing a wrong prompt."""
|
||||
instead of silently producing a wrong prompt.
|
||||
"""
|
||||
ctx = build_model_context(
|
||||
model_id,
|
||||
limit_mm_per_prompt={"image": 0, "video": 1},
|
||||
|
||||
@@ -116,7 +116,8 @@ def test_audio_multimodal_processor(model_id, processor_cls):
|
||||
@pytest.mark.parametrize("separator", [" and ", ""])
|
||||
def test_audio_multiple_inputs(separator, processor_cls):
|
||||
"""Multiple audios per prompt are each detected as a separate placeholder
|
||||
and multi-modal item by the Transformers modelling backend."""
|
||||
and multi-modal item by the Transformers modelling backend.
|
||||
"""
|
||||
model_id = "ibm-granite/granite-speech-3.3-2b"
|
||||
mm_processor = create_processor(model_id, processor_cls)
|
||||
|
||||
|
||||
@@ -83,7 +83,8 @@ def _process_two_images(processor_cls, separator: str):
|
||||
@pytest.mark.parametrize("processor_cls", PROCESSOR_CLASSES)
|
||||
def test_image_multiple_inputs(processor_cls):
|
||||
"""Multiple images per prompt are each detected as a separate placeholder
|
||||
and multi-modal item by the Transformers modelling backend."""
|
||||
and multi-modal item by the Transformers modelling backend.
|
||||
"""
|
||||
result = _process_two_images(processor_cls, separator="\n and ")
|
||||
|
||||
assert len(result["mm_placeholders"]["image"]) == 2
|
||||
@@ -102,7 +103,8 @@ def test_image_adjacent_inputs(processor_cls):
|
||||
@pytest.mark.parametrize("processor_cls", PROCESSOR_CLASSES)
|
||||
def test_batch_padding_removed_from_image_items(processor_cls):
|
||||
"""Emu3 pads every image up to the largest in the batch, which would leave an
|
||||
item's data dependent on what it was processed with and so uncacheable."""
|
||||
item's data dependent on what it was processed with and so uncacheable.
|
||||
"""
|
||||
mm_processor = create_processor("BAAI/Emu3-Chat-hf", processor_cls)
|
||||
image_token = mm_processor.info.get_hf_processor().image_token
|
||||
|
||||
@@ -144,7 +146,8 @@ def _process_one_gemma3_image(processor_cls):
|
||||
@offsets_only
|
||||
def test_non_embedding_tokens_excluded_from_placeholders():
|
||||
"""Gemma3 wraps each image in text that carries no embeddings, which must be
|
||||
inside the placeholder range but masked out of it."""
|
||||
inside the placeholder range but masked out of it.
|
||||
"""
|
||||
_, result = _process_one_gemma3_image(OffsetsMultiModalProcessor)
|
||||
|
||||
(placeholder,) = result["mm_placeholders"]["image"]
|
||||
@@ -156,7 +159,8 @@ def test_legacy_placeholders_hold_only_image_tokens():
|
||||
"""The legacy path spans whatever `mm_token_type_ids` attributes to the image,
|
||||
which for Gemma3 excludes the text wrapping it, unlike the replacement the offsets
|
||||
path spans. Gemma3 is also the sharp case for the mask: its `image_token_id` is the
|
||||
marker in the unexpanded prompt, not the token the expansion repeats."""
|
||||
marker in the unexpanded prompt, not the token the expansion repeats.
|
||||
"""
|
||||
hf_processor, result = _process_one_gemma3_image(LegacyMultiModalProcessor)
|
||||
|
||||
(placeholder,) = result["mm_placeholders"]["image"]
|
||||
@@ -172,7 +176,8 @@ def test_tokens_structuring_an_image_are_masked_not_dropped(processor_cls):
|
||||
embeddings. Those belong inside the placeholder and masked out, because the token
|
||||
count the processor reports is over the whole span. Idefics3 also refuses a prompt
|
||||
holding `<image>` when no images are passed, which is how the offsets path has to
|
||||
tokenize it before splicing in the expansion."""
|
||||
tokenize it before splicing in the expansion.
|
||||
"""
|
||||
mm_processor = create_processor(
|
||||
"HuggingFaceTB/SmolVLM-256M-Instruct", processor_cls
|
||||
)
|
||||
@@ -192,7 +197,8 @@ def test_tokens_structuring_an_image_are_masked_not_dropped(processor_cls):
|
||||
@offsets_only
|
||||
def test_missing_replacement_offsets_names_the_processor():
|
||||
"""A processor that reports no replacement offsets cannot be served, which must
|
||||
be said plainly rather than surfacing later as a field config mismatch."""
|
||||
be said plainly rather than surfacing later as a field config mismatch.
|
||||
"""
|
||||
model_id = "llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
|
||||
mm_processor = create_processor(model_id, OffsetsMultiModalProcessor)
|
||||
hf_processor_cls = type(mm_processor.info.get_hf_processor())
|
||||
@@ -276,7 +282,8 @@ def test_repeated_image_hits_the_processor_cache():
|
||||
def test_spliced_prompt_matches_hf_expansion(model_id, prompt):
|
||||
"""The offsets path splices the expansion into a prompt tokenized without any
|
||||
multi-modal data, so its token ids have to come out the same as the ones the HF
|
||||
processor produces itself, which is what the legacy path returns."""
|
||||
processor produces itself, which is what the legacy path returns.
|
||||
"""
|
||||
prompt_ids = []
|
||||
for processor_cls in (LegacyMultiModalProcessor, OffsetsMultiModalProcessor):
|
||||
mm_processor = create_processor(model_id, processor_cls)
|
||||
@@ -298,7 +305,8 @@ def test_spliced_prompt_matches_hf_expansion(model_id, prompt):
|
||||
def test_nested_image_fields_split_per_image(processor_cls):
|
||||
"""Idefics3 returns image fields with a leading batch dimension, putting the rows
|
||||
belonging to each image one dimension further in. Slicing the batch dimension
|
||||
instead handed the first image every row and the second an empty tensor."""
|
||||
instead handed the first image every row and the second an empty tensor.
|
||||
"""
|
||||
mm_processor = create_processor(
|
||||
"HuggingFaceTB/SmolVLM-256M-Instruct", processor_cls
|
||||
)
|
||||
|
||||
@@ -29,7 +29,8 @@ PROCESSOR_CLASSES = [
|
||||
|
||||
def create_processor(model_id: str, processor_cls):
|
||||
"""Build a processor directly, because the registry only ever builds the one the
|
||||
installed transformers version selects, leaving the other path untested."""
|
||||
installed transformers version selects, leaving the other path untested.
|
||||
"""
|
||||
model_config = ModelConfig(model=model_id, model_impl="transformers")
|
||||
ctx = InputProcessingContext(
|
||||
model_config, cached_tokenizer_from_config(model_config)
|
||||
@@ -40,7 +41,8 @@ def create_processor(model_id: str, processor_cls):
|
||||
|
||||
def create_cached_processor(model_id: str, processor_cls):
|
||||
"""Build a processor backed by a real multi-modal processor cache, and hand the
|
||||
cache back so a test can check it was actually used."""
|
||||
cache back so a test can check it was actually used.
|
||||
"""
|
||||
model_config = ModelConfig(model=model_id, model_impl="transformers")
|
||||
model_config.multimodal_config.mm_processor_cache_gb = 4
|
||||
ctx = InputProcessingContext(
|
||||
|
||||
Reference in New Issue
Block a user