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:
Harry Mellor
2026-09-16 16:44:29 +00:00
committed by GitHub
co-authored by Claude Opus 5
parent 804e5377aa
commit 8c1557a79c
2002 changed files with 15024 additions and 14269 deletions
@@ -23,11 +23,9 @@ NONINTERLEAVED_PROMPT = base_prompt("<image><image><video>\n")
@pytest.mark.parametrize("dtype", ["float16"])
@pytest.mark.parametrize("max_tokens", [128])
def test_models(vllm_runner, model, dtype: str, max_tokens: int) -> None:
"""
This is a simple test to check if interleaved and non-interleaved prompts
"""This is a simple test to check if interleaved and non-interleaved prompts
give the same result.
"""
image_cherry = convert_image_mode(ImageAsset("cherry_blossom").pil_image, "RGB")
image_stop = convert_image_mode(ImageAsset("stop_sign").pil_image, "RGB")
images = [image_cherry, image_stop]
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Create a reduced-layer version of the Maverick model for testing purposes.
"""Create a reduced-layer version of the Maverick model for testing purposes.
This script creates a new model with fewer layers by:
1. Loading the original Maverick model configuration
@@ -39,7 +38,6 @@ def run_maverick_serving(model: str):
"""Test Llama-4-Maverick model with vLLM LLM class using CLI equivalent
options with reduced layers.
"""
try:
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
@@ -71,8 +69,7 @@ def run_maverick_serving(model: str):
def get_rope_layers_config(model_path: str) -> list[int]:
"""
Get the interleaved RoPE configuration from HuggingFace config
"""Get the interleaved RoPE configuration from HuggingFace config.
Args:
model_path: Path to the local directory containing the reduced
@@ -81,6 +78,7 @@ def get_rope_layers_config(model_path: str) -> list[int]:
Returns:
List of 0 or 1 indicating whether each layer uses RoPE and local attn
0 indicates that RoPE is not used while 1 indicates that RoPE is used.
"""
config_path = Path(model_path) / "config.json"
model_config = json.loads(config_path.read_text())
@@ -98,8 +96,7 @@ def create_reduced_maverick_model(
vision_layers: int = 2,
force_recreate: bool = False,
) -> str:
"""
Create a reduced-layer version of the Maverick model.
"""Create a reduced-layer version of the Maverick model.
Args:
original_model_name: Name of the original Maverick model
@@ -111,8 +108,8 @@ def create_reduced_maverick_model(
Returns:
Path to the created reduced model directory
"""
"""
print(
f"Creating reduced Maverick model with {text_layers} text layers and "
f"{vision_layers} vision layers..."
@@ -178,7 +175,6 @@ def create_reduced_config(
original_config: Any, text_layers: int, num_experts: int, vision_layers: int
) -> dict[str, Any]:
"""Create a reduced configuration based on the original."""
# Convert config to dictionary
config_dict = original_config.to_dict()
@@ -220,7 +216,6 @@ def create_reduced_config(
def copy_tokenizer_files(original_model_name: str, output_path: Path) -> None:
"""Copy tokenizer files from the original model."""
try:
tokenizer = AutoTokenizer.from_pretrained(
original_model_name, trust_remote_code=True
@@ -233,7 +228,6 @@ def copy_tokenizer_files(original_model_name: str, output_path: Path) -> None:
def create_preprocessor_config(original_config: Any, output_path: Path) -> None:
"""Create preprocessor_config.json for multimodal model."""
# Try to load the original preprocessor config
try:
processor = AutoProcessor.from_pretrained(
@@ -253,7 +247,6 @@ def create_reduced_safetensors(
original_config: Any, reduced_config: dict[str, Any], output_path: Path
) -> None:
"""Create safetensors files with weights for the reduced model."""
print("Generating synthetic weights for reduced model...")
text_config = reduced_config["text_config"]
@@ -276,7 +269,6 @@ def create_reduced_safetensors(
def create_text_model_weights(text_config: dict[str, Any]) -> dict[str, torch.Tensor]:
"""Create synthetic weights for the text model with MoE structure."""
weights = {}
vocab_size = text_config["vocab_size"]
@@ -405,7 +397,6 @@ def create_vision_model_weights(
vision_config: dict[str, Any],
) -> dict[str, torch.Tensor]:
"""Create synthetic weights for the vision model."""
weights = {}
hidden_size = vision_config["hidden_size"]
@@ -473,8 +464,7 @@ def create_vision_model_weights(
def create_shared_weights(
text_config: dict[str, Any], vision_config: dict[str, Any]
) -> dict[str, torch.Tensor]:
"""Create weights for shared components (vision-language connector)"""
"""Create weights for shared components (vision-language connector)."""
weights = {}
text_hidden_size = text_config["hidden_size"]
@@ -492,7 +482,6 @@ def save_weights_to_safetensors(
weights: dict[str, torch.Tensor], output_path: Path
) -> None:
"""Save weights to safetensors files and create index."""
# Determine how to shard the weights
max_shard_size = 5 * 1024 * 1024 * 1024 # 5GB per shard
@@ -655,7 +644,6 @@ def test_dummy_maverick(
def main():
"""Main function to create and test the reduced model."""
import argparse
parser = argparse.ArgumentParser(
@@ -33,7 +33,8 @@ NUM_TOKENS = 3
class _StubGate(torch.nn.Module):
"""Router stand-in, so the test does not depend on platform-specific
``ReplicatedLinear`` weight post-processing."""
``ReplicatedLinear`` weight post-processing.
"""
def __init__(self, hidden_size: int, num_experts: int):
super().__init__()
@@ -31,8 +31,7 @@ def mask_bbox_tokens(
output: tuple[list[int], str, SampleLogprobs],
tokenizer: TokenizerLike,
) -> tuple[list[int], str, SampleLogprobs]:
"""
Always pass check_logprobs_close check for bounding box tokens
"""Always pass check_logprobs_close check for bounding box tokens
because it is reasonable for them to differ slightly.
"""
ignore_pattern = r"<[xy]_[\d.]+>"
@@ -97,7 +97,7 @@ def batch_make_image_embeddings(
processor,
llm: VllmRunner,
) -> list[Qwen2VLPromptImageEmbeddingInput]:
"""batched image embeddings for Qwen2-VL
"""Batched image embeddings for Qwen2-VL.
This will infer all images' embeddings in a single batch,
and split the result according to input batches.
@@ -108,7 +108,6 @@ def batch_make_image_embeddings(
returns: `list[Qwen2VLPromptImageEmbeddingInput]`
"""
image_batches_: list[Any] = image_batches[:]
# convert single-image batches to multiple-image batches
@@ -183,7 +182,7 @@ def batch_make_image_embeddings(
def batch_make_video_embeddings(
video_batches: PromptVideoInput, processor, llm: VllmRunner
) -> list[Qwen2VLPromptVideoEmbeddingInput]:
"""batched video embeddings for Qwen2-VL
"""Batched video embeddings for Qwen2-VL.
A NDArray represents a single video's all frames.
@@ -194,7 +193,6 @@ def batch_make_video_embeddings(
- Single-video batches: `list[NDArray]`
- Multiple-video batches: `list[list[NDArray]]`
"""
video_batches_: list[Any] = video_batches[:]
for idx in range(len(video_batches_)):
@@ -81,7 +81,8 @@ def test_transformers_audio_generation(
under pytest can cause (possibly flaky) hangs, so they are grouped under
the same config. Using VLLM_WORKER_MULTIPROC_METHOD=spawn avoids the
deadlock and allows worker processes to terminate cleanly, and release
GPU memory between test runs until the issue is fixed."""
GPU memory between test runs until the issue is fixed.
"""
# TODO: Remove monkeypatch once
# https://github.com/vllm-project/vllm/issues/17676 is fixed.
disable_envs_cache()
@@ -204,7 +204,6 @@ def test_variable_length_audio_batching(
@pytest.mark.asyncio
async def test_online_serving(client, audio_assets: AudioTestAssets):
"""Exercises online serving with/without chunked prefill enabled."""
messages = [
{
"role": "user",
@@ -78,7 +78,6 @@ def test_online_serving(vllm_runner, audio_assets: AudioTestAssets):
Steps run sequentially so each releases the GPU before the next starts.
"""
question = f"What's happening in these {len(audio_assets)} audio clips?"
max_tokens = 10
audio_data = [asset.audio_and_sample_rate for asset in audio_assets]
@@ -294,7 +294,8 @@ def sample_frames_with_video_metadata(
def apply_image_size_scaling(image, size: float | tuple[int, int], size_type: SizeType):
"""Applies a size scaler to one image; this can be an image size factor,
which scales the image while maintaining the aspect ratio"""
which scales the image while maintaining the aspect ratio
"""
# Special case for embeddings; if it's a tensor, it's only valid if we
# are considering size factors at constant scale, i.e., we just clone
# the tensor
@@ -156,6 +156,7 @@ def get_wrapped_test_sizes(
Args:
test_info: Test configuration to be expanded.
test_type: The type of test being filtered for.
"""
# If it is an embedding test, we always use the EMBEDDING_SIZE_FACTORS
if test_type == VLMTestType.EMBEDDING:
@@ -200,7 +200,7 @@ def process_runner_outputs(
def process_outputs(output_processor, model, outputs_per_image):
"""Applies a model specific post-processor function to a runner's output"""
"""Applies a model specific post-processor function to a runner's output."""
return [
[output_processor(res, model) for res in outputs]
for outputs in outputs_per_image
@@ -22,6 +22,7 @@ def multi_image_multi_aspect_ratio_inputs(formatter: Callable[[str], str]):
Args:
formatter: model-specific prompt formatter.
"""
stop_sign = IMAGE_ASSETS[0].pil_image
cherry_blossom = IMAGE_ASSETS[1].pil_image
@@ -65,6 +66,7 @@ def multi_video_multi_aspect_ratio_inputs(
Args:
formatter: model-specific prompt formatter.
"""
video = sample_frames_from_video(VIDEO_ASSETS[0].np_ndarrays, num_frames)
# Apply the selected formatter to the base prompts
@@ -252,6 +252,7 @@ def qwen_prompt_path_encoder(
tmp_path: Tempdir for test under consideration.
prompt: Prompt with image placeholders.
assets: list of image assets whose len equals the num placeholders.
"""
# Ensure that the number of placeholders matches the number of assets;
# If this is not true, the test is probably written incorrectly.
@@ -507,15 +508,12 @@ def isaac_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
2) Ensure IsaacModel.forward returns hidden_states
for compatibility with hidden_states_to_seq_logprobs()
"""
from perceptron.tensorstream import TextType
from perceptron.tensorstream.ops import compute_mrope_pos_tensor, modality_mask
from transformers.modeling_outputs import BaseModelOutputWithPast
def compute_position_ids_input_ids(input_ids: torch.Tensor) -> torch.Tensor:
"""
Create 3D positional indices for token input.
"""
"""Create 3D positional indices for token input."""
batch_size, seq_length = input_ids.shape
position_ids = torch.arange(seq_length, device=input_ids.device)
position_ids = position_ids.view(1, -1).expand(batch_size, -1)
@@ -570,8 +568,7 @@ def isaac_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
cache_position=None,
**kwargs,
):
"""
Forward pass with MRoPE position embeddings.
"""Forward pass with MRoPE position embeddings.
Computes position embeddings once and passes them through all layers.
"""
output_hidden_states = (
@@ -1184,7 +1181,6 @@ def voxtral_patch_hf_runner(hf_model: "HfRunner") -> "HfRunner":
HfRunner.generate calls batch_decode on the full sequence (prompt +
generated).
"""
import io
import pybase64 as base64
@@ -87,7 +87,7 @@ def _run_test(
*,
dtype: str,
) -> None:
"""SET PYTHONPATH"""
"""SET PYTHONPATH."""
# NOTE: take care of the order. run vLLM first, and then run HF.
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
@@ -108,8 +108,7 @@ def _normalize_image(image_val: str) -> str:
def create_score_multimodal_param(
content_parts: list[dict],
) -> list[ScoreMultiModalParam]:
"""
Create a ScoreMultiModalParam from a list of content dictionaries.
"""Create a ScoreMultiModalParam from a list of content dictionaries.
Each dict supports the following formats:
- Text: {'text': 'content'}
@@ -257,7 +256,7 @@ def test_model_text_image(
model: str,
dtype: str,
) -> None:
"""Visual Documents Reranking"""
"""Visual Documents Reranking."""
_run_test(
hf_runner,
vllm_runner,
@@ -280,7 +279,7 @@ def test_model_text_text(
model: str,
dtype: str,
) -> None:
"""Textual Documents Reranking"""
"""Textual Documents Reranking."""
_run_test(
hf_runner,
vllm_runner,
@@ -303,7 +302,7 @@ def test_model_image_text(
model: str,
dtype: str,
) -> None:
"""Image Querying for Textual Documents"""
"""Image Querying for Textual Documents."""
_run_test(
hf_runner,
vllm_runner,
@@ -326,7 +325,7 @@ def test_model_image_image(
model: str,
dtype: str,
) -> None:
"""Image Querying for Image Documents"""
"""Image Querying for Image Documents."""
_run_test(
hf_runner,
vllm_runner,
@@ -349,7 +348,7 @@ def test_model_text_mixed_documents(
model: str,
dtype: str,
) -> None:
"""Text Query for Mixed Text and Image Documents"""
"""Text Query for Mixed Text and Image Documents."""
_run_test(
hf_runner,
vllm_runner,
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Tests for the LlamaNemotronVL model family:
"""Tests for the LlamaNemotronVL model family:
- nvidia/llama-nemotron-embed-vl-1b-v2 (LlamaNemotronVLForCausalLM / embed)
- nvidia/llama-nemotron-rerank-vl-1b-v2
(LlamaNemotronVLForSequenceClassification / rerank)
@@ -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(
+2 -1
View File
@@ -18,7 +18,8 @@ from vllm.model_executor.models.cohere_asr import CohereAsrForConditionalGenerat
def test_get_num_audio_tokens_streaming_estimate(audio_duration_s, expected_tokens):
"""The duration-based estimate must convert ``window_stride`` (seconds) to a
sample hop and divide by the encoder subsampling factor. Values are pinned to
concrete numbers so a rounding regression is actually caught."""
concrete numbers so a rounding regression is actually caught.
"""
model_config = SimpleNamespace(
hf_config=SimpleNamespace(
preprocessor={"window_stride": 0.01, "sample_rate": 16000},
+5 -8
View File
@@ -78,7 +78,8 @@ def test_cosmos3_modelopt_quantizer_weights_mapper():
"""ModelOpt/Diffusers FP8 checkpoints ship native fake-quant buffers
(``*_quantizer._amax`` / ``._scale``) alongside the vLLM-consumable
``weight_scale`` / ``input_scale`` sidecars. vLLM must drop the former
(it has no parameter for them) while keeping the latter."""
(it has no parameter for them) while keeping the latter.
"""
from vllm.model_executor.models.cosmos3 import Cosmos3ForConditionalGeneration
mapper = Cosmos3ForConditionalGeneration.hf_to_vllm_mapper
@@ -215,7 +216,7 @@ def test_bailing_vl_mapper_handles_module_and_parameter_names():
def create_repo_dummy_weights(repo: str) -> Iterable[tuple[str, torch.Tensor]]:
"""Create weights from safetensors checkpoint metadata"""
"""Create weights from safetensors checkpoint metadata."""
metadata = try_get_safetensors_metadata(repo)
weight_names = list(metadata.weight_map.keys())
with torch.device("meta"):
@@ -223,9 +224,7 @@ def create_repo_dummy_weights(repo: str) -> Iterable[tuple[str, torch.Tensor]]:
def create_dummy_base_model(repo: str, model_arch: str) -> PreTrainedModel:
"""
Create weights from a dummy meta deserialized hf base model with name conversion
"""
"""Create weights from a dummy meta-deserialized HF base model, with renaming."""
config = AutoConfig.from_pretrained(repo)
with torch.device("meta"):
model = AutoModel.from_config(config)
@@ -233,9 +232,7 @@ def create_dummy_base_model(repo: str, model_arch: str) -> PreTrainedModel:
def create_dummy_model(repo: str, model_arch: str) -> PreTrainedModel:
"""
Create weights from a dummy meta deserialized hf model with name conversion
"""
"""Create weights from a dummy meta deserialized hf model with name conversion."""
model_cls: PreTrainedModel = getattr(transformers, model_arch)
config = AutoConfig.from_pretrained(repo)
with torch.device("meta"):
@@ -81,7 +81,8 @@ class _FakeTensor:
"""Sentinel stand-in for torch.Tensor in load_weights tests. Supports the
.detach().clone() chain used by load_weights for buffered mm weights;
both methods return self so identity (and the existing equality
assertions) are preserved through cloning."""
assertions) are preserved through cloning.
"""
def detach(self):
return self
@@ -165,7 +166,8 @@ def _make_mm_items_with_video_bytes(
def test_extract_audio_from_videos_passes_max_duration():
"""_extract_audio_from_videos must forward VLLM_MAX_AUDIO_DECODE_DURATION_S
to load_audio_pyav so decompression-bomb audio is rejected."""
to load_audio_pyav so decompression-bomb audio is rejected.
"""
dummy_audio = (np.zeros(16000, dtype=np.float32), 16000.0)
mm_items = _make_mm_items_with_video_bytes(b"\x00" * 64)
@@ -185,7 +187,8 @@ def test_extract_audio_from_videos_passes_max_duration():
def test_extract_audio_from_videos_rejects_oversized_audio():
"""When load_audio_pyav raises due to duration limit the video is
marked as having no audio instead of crashing the server."""
marked as having no audio instead of crashing the server.
"""
mm_items = _make_mm_items_with_video_bytes(b"\x00" * 64)
processor = object.__new__(NanoNemotronVLMultiModalProcessor)