biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Convert formatting to use ruff instead of yapf + isort (#26247)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-10-05 15:06:22 +01:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115244 additions and 94146 deletions
Download Patch File Download Diff File Expand all files Collapse all files

155
tests/models/multimodal/generation/test_phi4_multimodal.py
View File

@@ -14,26 +14,35 @@ from vllm.lora.request import LoRARequest
from vllm.multimodal.image import rescale_image_size
from vllm.platforms import current_platform
from ....conftest import (IMAGE_ASSETS, HfRunner, PromptAudioInput,
PromptImageInput, VllmRunner)
from ....conftest import (
IMAGE_ASSETS,
HfRunner,
PromptAudioInput,
PromptImageInput,
VllmRunner,
)
from ....utils import large_gpu_test
from ...utils import check_logprobs_close
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign":
"<|user|>\n<|image|>\nWhat's the content of the image?<|end|>\n<|assistant|>\n", # noqa: E501
"cherry_blossom":
"<|user|>\n<|image|>\nPlease infer the season with reason in details.<|end|>\n<|assistant|>\n", # noqa: E501
})
HF_MULTIIMAGE_IMAGE_PROMPT = "<|user|>\n<|image|>\n<|image|>\nDescribe these images.<|end|>\n<|assistant|>\n" # noqa: E501
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts(
{
"stop_sign": "<|user|>\n<|image|>\nWhat's the content of the image?<|end|>\n<|assistant|>\n", # noqa: E501
"cherry_blossom": "<|user|>\n<|image|>\nPlease infer the season with reason in details.<|end|>\n<|assistant|>\n", # noqa: E501
}
)
HF_MULTIIMAGE_IMAGE_PROMPT = (
"<|user|>\n<|image|>\n<|image|>\nDescribe these images.<|end|>\n<|assistant|>\n" # noqa: E501
)
model_path = snapshot_download("microsoft/Phi-4-multimodal-instruct",
revision="refs/pr/70")
model_path = snapshot_download(
"microsoft/Phi-4-multimodal-instruct", revision="refs/pr/70"
)
# Since the vision-lora and speech-lora co-exist with the base model,
# we have to manually specify the path of the lora weights.
vision_lora_path = os.path.join(model_path, "vision-lora")
speech_question = os.path.join(model_path, "examples",
"what_is_shown_in_this_image.wav")
speech_question = os.path.join(
model_path, "examples", "what_is_shown_in_this_image.wav"
)
models = [model_path]
target_dtype = "half"
@@ -48,8 +57,7 @@ if current_platform.is_rocm():
def run_test(
hf_runner: type[HfRunner],
vllm_runner: type[VllmRunner],
inputs: Sequence[tuple[list[str], PromptImageInput,
Optional[PromptAudioInput]]],
inputs: Sequence[tuple[list[str], PromptImageInput, Optional[PromptAudioInput]]],
model: str,
*,
max_model_len: int,
@@ -75,28 +83,30 @@ def run_test(
# will hurt multiprocessing backend with fork method (the default method).
# max_model_len should be greater than image_feature_size
with vllm_runner(
model,
task="generate",
max_model_len=max_model_len,
max_num_seqs=2,
dtype=dtype,
limit_mm_per_prompt={"image": mm_limit},
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enable_lora=True,
max_lora_rank=320,
gpu_memory_utilization=0.8, # set to 0.8 to avoid OOM in CI
enforce_eager=True,
trust_remote_code=False,
model,
task="generate",
max_model_len=max_model_len,
max_num_seqs=2,
dtype=dtype,
limit_mm_per_prompt={"image": mm_limit},
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enable_lora=True,
max_lora_rank=320,
gpu_memory_utilization=0.8, # set to 0.8 to avoid OOM in CI
enforce_eager=True,
trust_remote_code=False,
) as vllm_model:
lora_request = LoRARequest("vision", 1, vision_lora_path)
vllm_outputs_per_case = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images,
audios=audios,
lora_request=lora_request)
vllm_model.generate_greedy_logprobs(
prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images,
audios=audios,
lora_request=lora_request,
)
for prompts, images, audios in inputs
]
@@ -108,17 +118,18 @@ def run_test(
hf_processor = hf_model.processor
eos_token_id = hf_processor.tokenizer.eos_token_id
hf_outputs_per_case = [
hf_model.generate_greedy_logprobs_limit(prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images,
audios=audios,
eos_token_id=eos_token_id)
hf_model.generate_greedy_logprobs_limit(
prompts,
max_tokens,
num_logprobs=num_logprobs,
images=images,
audios=audios,
eos_token_id=eos_token_id,
)
for prompts, images, audios in inputs
]
for hf_outputs, vllm_outputs in zip(hf_outputs_per_case,
vllm_outputs_per_case):
for hf_outputs, vllm_outputs in zip(hf_outputs_per_case, vllm_outputs_per_case):
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_outputs,
@@ -145,16 +156,27 @@ def run_test(
@pytest.mark.parametrize("max_model_len", [12800])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [10])
def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
dtype: str, max_model_len: int, max_tokens: int,
num_logprobs: int) -> None:
def test_models(
hf_runner,
vllm_runner,
image_assets,
model,
size_factors,
dtype: str,
max_model_len: int,
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],
None,
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
inputs_per_image = [
(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
None,
)
for image, prompt in zip(images, HF_IMAGE_PROMPTS)
]
run_test(
hf_runner,
@@ -189,16 +211,26 @@ def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
@pytest.mark.parametrize("max_model_len", [25600])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [10])
def test_multi_images_models(hf_runner, vllm_runner, image_assets, model,
size_factors, dtype: str, max_model_len: int,
max_tokens: int, num_logprobs: int) -> None:
def test_multi_images_models(
hf_runner,
vllm_runner,
image_assets,
model,
size_factors,
dtype: str,
max_model_len: int,
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],
[
[rescale_image_size(image, factor) for image in images]
for factor in size_factors
],
None,
),
]
@@ -222,10 +254,15 @@ def test_multi_images_models(hf_runner, vllm_runner, image_assets, model,
@pytest.mark.parametrize("max_model_len", [12800])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [10])
def test_vision_speech_models(hf_runner, vllm_runner, model, dtype: str,
max_model_len: int, max_tokens: int,
num_logprobs: int) -> None:
def test_vision_speech_models(
hf_runner,
vllm_runner,
model,
dtype: str,
max_model_len: int,
max_tokens: int,
num_logprobs: int,
) -> None:
# use the example speech question so that the model outputs are reasonable
audio = librosa.load(speech_question, sr=16000)
image = ImageAsset("cherry_blossom").pil_image.convert("RGB")