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
@@ -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
|
||||
|
||||
Reference in New Issue
Block a user