Skip to content

GitLab

Commit 4851c202 authored by zhuwenwen's avatar zhuwenwen
Browse files

Merge tag 'v0.6.1' into v0.6.1-dev

parents 9b902f9e 3fd2b0d2
Expand all Show whitespace changes
Inline Side-by-side
Showing with 1439 additions and 936 deletions
tests/models/test_aqlm.py
View file @ 4851c202
......@@ -7,26 +7,6 @@ import pytest
from tests.quantization.utils import is_quant_method_supported
# In this test we hardcode prompts and generations for the model so we don't
# need to require the AQLM package as a dependency
example_prompts = [
'vLLM is a high-throughput and memory-efficient inference and serving '
'engine for LLMs.\n',
'Briefly describe the major milestones in the development of artificial '
'intelligence from 1950 to 2020.\n',
'Compare and contrast artificial intelligence with human intelligence in '
'terms of processing information.\n',
'Describe the basic components of a neural network and how it can be '
'trained.\n',
'Write a short story about a robot that dreams for the first time.\n',
'Analyze the impact of the COVID-19 pandemic on global economic structures '
'and future business models.\n',
'Explain the cultural significance of the Mona Lisa painting, and how its '
'perception might vary in Western versus Eastern societies.\n',
"Translate the following English sentence into Japanese, French, and "
"Swahili: 'The early bird catches the worm.'\n"
]
# These ground truth generations were generated using `transformers==4.38.1
# aqlm==1.1.0 torch==2.2.0`
# and the below code:
......
tests/models/test_internvl.py
View file @ 4851c202
import types
from typing import List, Optional, Tuple, Type
from typing import List, Optional, Tuple, Type, Union
import pytest
import torch
......@@ -9,7 +9,8 @@ from transformers import AutoConfig
from vllm.multimodal.utils import rescale_image_size
from vllm.utils import is_cpu
from ..conftest import IMAGE_ASSETS, HfRunner, VllmRunner, _ImageAssets
from ..conftest import (IMAGE_ASSETS, HfRunner, PromptImageInput, VllmRunner,
_ImageAssets)
from .utils import check_logprobs_close
pytestmark = pytest.mark.vlm
......@@ -20,6 +21,7 @@ HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"cherry_blossom":
"<|im_start|>User\n<image>\nWhat is the season?<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501
})
HF_MULTIIMAGE_IMAGE_PROMPT = "<|im_start|>User\nImage-1: <image>\nImage-2: <image>\nDescribe the two images in detail.<|im_end|>\n<|im_start|>Assistant\n" # noqa: E501
models = [
"OpenGVLab/InternVL2-1B",
......@@ -64,13 +66,13 @@ def generate(
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
image_assets: _ImageAssets,
inputs: List[Tuple[List[str], PromptImageInput]],
model: str,
*,
size_factors: List[float],
dtype: str,
max_tokens: int,
num_logprobs: int,
mm_limit: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
......@@ -83,12 +85,6 @@ def run_test(
Note, the text input is also adjusted to abide by vllm contract.
The text output is sanitized to be able to compare with hf.
"""
images = [asset.pil_image for asset in image_assets]
inputs_per_image = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
# NOTE: take care of the order. run vLLM first, and then run HF.
# vLLM needs a fresh new process without cuda initialization.
......@@ -110,13 +106,21 @@ def run_test(
self.max_num = self.config.max_dynamic_patch
self.image_size = self.vision_config.image_size
def __call__(self, text: str, images: Image, **kwargs):
def __call__(self, text: str, images: Union[Image, List[Image]],
**kwargs):
from vllm.model_executor.models.internvl import (
IMG_CONTEXT, IMG_END, IMG_START, image_to_pixel_values)
pixel_values = image_to_pixel_values(
images, self.image_size, self.min_num, self.max_num,
images = [images] if isinstance(images, Image) else images
pixel_values = [
image_to_pixel_values(image, self.image_size, self.min_num,
self.max_num,
self.use_thumbnail).to(self.dtype)
num_patches_list = [pixel_values.shape[0]]
for image in images
]
num_patches_list = [
pixel_value.shape[0] for pixel_value in pixel_values
]
pixel_values = torch.cat(pixel_values, dim=0)
for num_patches in num_patches_list:
context_tokens = IMG_CONTEXT * self.num_image_token \
* num_patches
......@@ -130,6 +134,7 @@ def run_test(
with vllm_runner(model,
max_model_len=4096,
dtype=dtype,
limit_mm_per_prompt={"image": mm_limit},
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True) as vllm_model:
......@@ -138,7 +143,7 @@ def run_test(
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs_per_image
for prompts, images in inputs
]
with hf_runner(model, dtype=dtype) as hf_model:
......@@ -156,7 +161,7 @@ def run_test(
num_logprobs=num_logprobs,
images=hf_images,
eos_token_id=eos_token_id)
for prompts, hf_images in inputs_per_image
for prompts, hf_images in inputs
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_image,
......@@ -264,15 +269,64 @@ if is_cpu():
@torch.inference_mode()
def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
dtype: str, max_tokens: int, num_logprobs: int) -> None:
images = [asset.pil_image for asset in image_assets]
inputs_per_image = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
run_test(
hf_runner,
vllm_runner,
image_assets,
inputs_per_image,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
mm_limit=1,
tensor_parallel_size=1,
)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"size_factors",
[
# No image
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
# Multi-scale
[0.5, 0.75, 1.0],
],
)
@pytest.mark.parametrize("dtype", [target_dtype])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
@torch.inference_mode()
def test_multi_images_models(hf_runner, vllm_runner, image_assets, model,
size_factors, dtype: str, max_tokens: int,
num_logprobs: int) -> None:
images = [asset.pil_image for asset in image_assets]
inputs_per_case = [
([HF_MULTIIMAGE_IMAGE_PROMPT for _ in size_factors],
[[rescale_image_size(image, factor) for image in images]
for factor in size_factors])
]
run_test(
hf_runner,
vllm_runner,
inputs_per_case,
model,
size_factors=size_factors,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
mm_limit=2,
tensor_parallel_size=1,
)
......
tests/models/test_llava.py
View file @ 4851c202
from typing import List, Optional, Tuple, Type
from typing import List, Optional, Tuple, Type, overload
import pytest
from transformers import (AutoConfig, AutoModelForVision2Seq, AutoTokenizer,
......@@ -8,11 +8,14 @@ from vllm.multimodal.utils import rescale_image_size
from vllm.sequence import SampleLogprobs
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE
from ..conftest import IMAGE_ASSETS, HfRunner, VllmRunner, _ImageAssets
from ..conftest import (IMAGE_ASSETS, HfRunner, PromptImageInput, VllmRunner,
_ImageAssets)
from .utils import check_logprobs_close
pytestmark = pytest.mark.vlm
_LIMIT_IMAGE_PER_PROMPT = 4
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign":
"USER: <image>\nWhat's the content of the image?\nASSISTANT:",
......@@ -52,6 +55,7 @@ def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
return hf_output_ids, hf_output_str, out_logprobs
@overload
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
......@@ -64,6 +68,78 @@ def run_test(
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
...
@overload
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
image_assets: _ImageAssets,
model: str,
*,
sizes: List[Tuple[int, int]],
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
...
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
image_assets: _ImageAssets,
model: str,
*,
size_factors: Optional[List[float]] = None,
sizes: Optional[List[Tuple[int, int]]] = None,
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
images = [asset.pil_image for asset in image_assets]
if size_factors is not None:
inputs_per_image = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
elif sizes is not None:
inputs_per_image = [(
[prompt for _ in sizes],
[image.resize(size) for size in sizes],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
else:
raise ValueError("You must provide either `size_factors` or `sizes`")
_run_test(hf_runner,
vllm_runner,
inputs_per_image,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend)
def _run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
inputs: List[Tuple[List[str], PromptImageInput]],
model: str,
*,
dtype: str,
max_tokens: int,
num_logprobs: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
"""Inference result should be the same between hf and vllm.
......@@ -85,13 +161,6 @@ def run_test(
else:
mantis_processor = None
images = [asset.pil_image for asset in image_assets]
inputs_per_image = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
# 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
......@@ -100,15 +169,18 @@ def run_test(
# max_model_len should be greater than image_feature_size
with vllm_runner(model,
dtype=dtype,
max_model_len=4096,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True) as vllm_model:
enforce_eager=True,
limit_mm_per_prompt={"image": _LIMIT_IMAGE_PER_PROMPT
}) as vllm_model:
vllm_outputs_per_image = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs_per_image
for prompts, images in inputs
]
if mantis_processor is not None:
......@@ -131,7 +203,7 @@ def run_test(
max_tokens,
num_logprobs=num_logprobs,
images=images)
for prompts, images in inputs_per_image
for prompts, images in inputs
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_image,
......@@ -181,6 +253,51 @@ def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models_multiple_image_inputs(hf_runner, vllm_runner, image_assets,
model, dtype, max_tokens,
num_logprobs) -> None:
stop_sign = image_assets[0].pil_image
cherry_blossom = image_assets[1].pil_image
inputs = [(
[
"USER: <image><image>\nDescribe 2 images.\nASSISTANT:",
"USER: <image><image>\nDescribe 2 images.\nASSISTANT:",
"USER: <image><image><image><image>\nDescribe 4 images.\nASSISTANT:", # noqa: E501
"USER: <image>\nWhat is the season?\nASSISTANT:",
],
[
[stop_sign, cherry_blossom],
# Images with different sizes and aspect-ratios
[
rescale_image_size(stop_sign, 0.1),
stop_sign,
],
[
stop_sign,
rescale_image_size(stop_sign, 0.25),
cherry_blossom.resize((183, 488)),
cherry_blossom.resize((488, 183))
],
cherry_blossom,
])]
_run_test(
hf_runner,
vllm_runner,
inputs,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@pytest.mark.parametrize("model", models)
def test_context_length_too_short(vllm_runner, image_assets, model):
images = [asset.pil_image for asset in image_assets]
......
tests/models/test_llava_next_video.py 0 → 100644
View file @ 4851c202
from typing import List, Optional, Tuple, Type, overload
import pytest
import transformers
from transformers import AutoConfig, AutoModelForVision2Seq, AutoTokenizer
from vllm.multimodal.utils import (rescale_video_size, resize_video,
sample_frames_from_video)
from vllm.sequence import SampleLogprobs
from ..conftest import VIDEO_ASSETS, HfRunner, VllmRunner, _VideoAssets
from .utils import check_logprobs_close
pytestmark = pytest.mark.vlm
_PREFACE = (
"A chat between a curious human and an artificial intelligence assistant. "
"The assistant gives helpful, detailed, and polite answers to the human's "
"questions.")
HF_VIDEO_PROMPTS = VIDEO_ASSETS.prompts({
"sample_demo_1":
f"{_PREFACE}USER: <video>\nWhy is this video funny? ASSISTANT:"
})
models = ["llava-hf/LLaVA-NeXT-Video-7B-hf"]
def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
Optional[SampleLogprobs]],
model: str):
"""Sanitize vllm output to be comparable with hf output."""
output_ids, output_str, out_logprobs = vllm_output
config = AutoConfig.from_pretrained(model)
video_token_id = config.video_token_index
tokenizer = AutoTokenizer.from_pretrained(model)
eos_token_id = tokenizer.eos_token_id
hf_output_ids = [
token_id for idx, token_id in enumerate(output_ids)
if token_id != video_token_id or output_ids[idx - 1] != video_token_id
]
assert output_str[0] == " "
hf_output_str = output_str[1:]
if hf_output_ids[-1] == eos_token_id:
hf_output_str = hf_output_str + tokenizer.decode(eos_token_id)
return hf_output_ids, hf_output_str, out_logprobs
@overload
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
video_assets: _VideoAssets,
model: str,
*,
size_factors: List[float],
dtype: str,
max_tokens: int,
num_logprobs: int,
num_frames: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
...
@overload
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
video_assets: _VideoAssets,
model: str,
*,
sizes: List[Tuple[int, int]],
dtype: str,
max_tokens: int,
num_logprobs: int,
num_frames: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
...
def run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
video_assets: _VideoAssets,
model: str,
*,
size_factors: Optional[List[float]] = None,
sizes: Optional[List[Tuple[int, int]]] = None,
dtype: str,
max_tokens: int,
num_logprobs: int,
num_frames: int,
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
videos = [
sample_frames_from_video(asset.np_ndarrays, num_frames)
for asset in video_assets
]
for video in videos:
print(video.shape)
if size_factors is not None:
inputs_per_video = [(
[prompt for _ in size_factors],
[rescale_video_size(video, factor) for factor in size_factors],
) for video, prompt in zip(videos, HF_VIDEO_PROMPTS)]
elif sizes is not None:
inputs_per_video = [(
[prompt for _ in sizes],
[resize_video(video, size) for size in sizes],
) for video, prompt in zip(videos, HF_VIDEO_PROMPTS)]
else:
raise ValueError("You must provide either `size_factors` or `sizes`")
# max_model_len should be greater than image_feature_size
with vllm_runner(model,
dtype=dtype,
max_model_len=4096,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True) as vllm_model:
vllm_outputs_per_video = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs=num_logprobs,
videos=videos)
for prompts, videos in inputs_per_video
]
with hf_runner(model, dtype=dtype,
auto_cls=AutoModelForVision2Seq) as hf_model:
hf_outputs_per_video = [
hf_model.generate_greedy_logprobs_limit(prompts,
max_tokens,
num_logprobs=num_logprobs,
videos=videos)
for prompts, videos in inputs_per_video
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_video,
vllm_outputs_per_video):
# TODO: Check whether using original CLIPVisionModel can improve
# consistency against HF
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=[
vllm_to_hf_output(vllm_output, model)
for vllm_output in vllm_outputs
],
name_0="hf",
name_1="vllm",
)
@pytest.mark.skipif(transformers.__version__ < "4.45",
reason="Waiting for next transformers release")
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"size_factors",
[
# No video
[],
# Single-scale
[1.0],
# Single-scale, batched
[1.0, 1.0, 1.0],
# Multi-scale
[0.25, 0.5, 1.0],
],
)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
@pytest.mark.parametrize("num_frames", [16])
def test_models(hf_runner, vllm_runner, video_assets, model, size_factors,
dtype, max_tokens, num_logprobs, num_frames) -> None:
"""Inference result should be the same between hf and vllm.
All the image fixtures for the test is under tests/videos.
For huggingface runner, we provide the np.ndarray as input.
For vllm runner, we provide MultiModalDataDict objects
and corresponding MultiModalConfig as input.
Note, the text input is also adjusted to abide by vllm contract.
The text output is sanitized to be able to compare with hf.
"""
run_test(
hf_runner,
vllm_runner,
video_assets,
model,
size_factors=size_factors,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
num_frames=num_frames,
tensor_parallel_size=1,
)
@pytest.mark.skipif(transformers.__version__ < "4.45",
reason="Waiting for next transformers release")
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"sizes",
[[(1669, 2560), (2560, 1669), (183, 488), (488, 183)]],
)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
@pytest.mark.parametrize("num_frames", [16])
def test_models_fixed_sizes(hf_runner, vllm_runner, video_assets, model, sizes,
dtype, max_tokens, num_logprobs,
num_frames) -> None:
run_test(
hf_runner,
vllm_runner,
video_assets,
model,
sizes=sizes,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
num_frames=num_frames,
tensor_parallel_size=1,
)
tests/models/test_mistral.py
View file @ 4851c202
......@@ -41,3 +41,43 @@ def test_models(
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", MODELS[1:])
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [64])
@pytest.mark.parametrize("num_logprobs", [5])
def test_mistral_format(
vllm_runner,
example_prompts,
model: str,
dtype: str,
max_tokens: int,
num_logprobs: int,
) -> None:
with vllm_runner(
model,
dtype=dtype,
tokenizer_mode="auto",
load_format="safetensors",
config_format="hf",
) as hf_format_model:
hf_format_outputs = hf_format_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
with vllm_runner(
model,
dtype=dtype,
tokenizer_mode="mistral",
load_format="mistral",
config_format="mistral",
) as mistral_format_model:
mistral_format_outputs = mistral_format_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
check_logprobs_close(
outputs_0_lst=hf_format_outputs,
outputs_1_lst=mistral_format_outputs,
name_0="hf",
name_1="mistral",
)
tests/models/test_modelopt.py 0 → 100644
View file @ 4851c202
This diff is collapsed.
tests/models/test_phi3v.py
View file @ 4851c202
This diff is collapsed.
tests/models/test_pixtral.py 0 → 100644
View file @ 4851c202
This diff is collapsed.
tests/models/test_qwen.py
View file @ 4851c202
This diff is collapsed.
tests/models/test_registry.py
View file @ 4851c202
import pytest
import transformers
from vllm.model_executor.models import _MODELS, ModelRegistry
@pytest.mark.parametrize("model_cls", _MODELS)
def test_registry_imports(model_cls):
if (model_cls == "Qwen2VLForConditionalGeneration"
and transformers.__version__ < "4.45"):
pytest.skip("Waiting for next transformers release")
# Ensure all model classes can be imported successfully
ModelRegistry.resolve_model_cls([model_cls])
tests/multi_step/test_correctness_llm.py
View file @ 4851c202
......@@ -57,7 +57,7 @@ def test_multi_step_llm(
GPU -> CPU output transfer
num_prompts: number of example prompts under test
num_logprobs: corresponds to the `logprobs` argument to the OpenAI
completions endpoint; `None` -> no logprobs
completions endpoint; `None` -> 1 logprob returned.
"""
prompts = example_prompts
......
tests/spec_decode/e2e/conftest.py
View file @ 4851c202
This diff is collapsed.
tests/spec_decode/e2e/test_integration.py
View file @ 4851c202
This diff is collapsed.
tests/spec_decode/e2e/test_logprobs.py
View file @ 4851c202
This diff is collapsed.
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment