[CI/Build] Reduce more duplicate runner startup in tests (#52570)

Signed-off-by: Isotr0py <Isotr0py@outlook.com>
This commit is contained in:
Isotr0py
2026-08-17 05:19:07 +00:00
committed by GitHub
parent 311b3513af
commit 53e211d292
10 changed files with 647 additions and 916 deletions
+66 -97
View File
@@ -26,8 +26,7 @@ MODELS = ["openai/clip-vit-base-patch32"]
def _run_test(
hf_runner: type[HfRunner],
vllm_runner: type[VllmRunner],
input_texts: list[str],
input_images: PromptImageInput,
input_cases: list[tuple[list[str], PromptImageInput]],
model: str,
*,
dtype: str,
@@ -39,105 +38,75 @@ def _run_test(
with vllm_runner(
model, runner="pooling", dtype=dtype, enforce_eager=True, max_model_len=77
) as vllm_model:
vllm_outputs = vllm_model.embed(input_texts, images=input_images)
vllm_outputs_per_case = [
vllm_model.embed(input_texts, images=input_images)
for input_texts, input_images in input_cases
]
with hf_runner(model, dtype=dtype, auto_cls=CLIPModel) as hf_model:
all_inputs = hf_model.get_inputs(input_texts, images=input_images)
all_outputs = []
for inputs in all_inputs:
inputs = hf_model.wrap_device(inputs)
if "pixel_values" in inputs:
pooled_output = hf_model.model.get_image_features(
pixel_values=inputs.pixel_values,
)
else:
pooled_output = hf_model.model.get_text_features(
input_ids=inputs.input_ids,
attention_mask=inputs.attention_mask,
)
if not isinstance(pooled_output, torch.Tensor):
pooled_output = pooled_output.pooler_output
pooled_output = pooled_output.squeeze(0)
all_outputs.append(pooled_output.tolist())
hf_outputs = all_outputs
check_embeddings_close(
embeddings_0_lst=hf_outputs,
embeddings_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["float"])
def test_models_text(
hf_runner,
vllm_runner,
image_assets,
model: str,
dtype: str,
) -> None:
input_texts_images = [(text, None) for text in HF_TEXT_PROMPTS]
input_texts = [text for text, _ in input_texts_images]
input_images = [image for _, image in input_texts_images]
_run_test(
hf_runner,
vllm_runner,
input_texts,
input_images, # type: ignore
model,
dtype=dtype,
)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["float"])
def test_models_image(
hf_runner,
vllm_runner,
image_assets,
model: str,
dtype: str,
) -> None:
input_texts_images = [
(text, asset.pil_image) for text, asset in zip(HF_IMAGE_PROMPTS, image_assets)
]
input_texts = [text for text, _ in input_texts_images]
input_images = [image for _, image in input_texts_images]
_run_test(
hf_runner,
vllm_runner,
input_texts,
input_images,
model,
dtype=dtype,
)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["float"])
def test_models_text_image_no_crash(
vllm_runner,
image_assets,
model: str,
dtype: str,
) -> None:
texts = [HF_TEXT_PROMPTS[0]]
images = [image_assets[0].pil_image]
with vllm_runner(
model, runner="pooling", dtype=dtype, enforce_eager=True, max_model_len=77
) as vllm_model:
texts = [HF_TEXT_PROMPTS[0]]
images = [input_cases[1][1][0]]
with pytest.raises(ValueError, match="not both"):
vllm_model.embed(texts, images=images)
# Should still be able to run subsequent requests
vllm_model.embed(texts)
vllm_model.embed([""], images=images)
with hf_runner(model, dtype=dtype, auto_cls=CLIPModel) as hf_model:
hf_outputs_per_case = []
for input_texts, input_images in input_cases:
all_inputs = hf_model.get_inputs(input_texts, images=input_images)
hf_outputs = []
for inputs in all_inputs:
inputs = hf_model.wrap_device(inputs)
if "pixel_values" in inputs:
pooled_output = hf_model.model.get_image_features(
pixel_values=inputs.pixel_values,
)
else:
pooled_output = hf_model.model.get_text_features(
input_ids=inputs.input_ids,
attention_mask=inputs.attention_mask,
)
if not isinstance(pooled_output, torch.Tensor):
pooled_output = pooled_output.pooler_output
pooled_output = pooled_output.squeeze(0)
hf_outputs.append(pooled_output.tolist())
hf_outputs_per_case.append(hf_outputs)
for hf_outputs, vllm_outputs in zip(hf_outputs_per_case, vllm_outputs_per_case):
check_embeddings_close(
embeddings_0_lst=hf_outputs,
embeddings_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["float"])
def test_models(
hf_runner,
vllm_runner,
image_assets,
model: str,
dtype: str,
) -> None:
text_images = [None] * len(HF_TEXT_PROMPTS)
images = [asset.pil_image for asset in image_assets]
input_cases = [
(HF_TEXT_PROMPTS, text_images),
(HF_IMAGE_PROMPTS, images),
]
_run_test(
hf_runner,
vllm_runner,
input_cases, # type: ignore[arg-type]
model,
dtype=dtype,
)