[Misc] Rename assets for testing (#17575)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-05-02 18:29:25 +08:00
committed by GitHub
parent c777df79f7
commit d7543862bd
28 changed files with 145 additions and 131 deletions

View File

@@ -11,7 +11,7 @@ from vllm.multimodal.image import rescale_image_size
from vllm.multimodal.video import (rescale_video_size, resize_video,
sample_frames_from_video)
from .....conftest import _ImageAssets, _VideoAssets
from .....conftest import ImageTestAssets, VideoTestAssets
from .types import (SINGLE_IMAGE_BASE_PROMPTS, TEST_IMG_PLACEHOLDER,
TEST_VIDEO_PLACEHOLDER, VIDEO_BASE_PROMPT,
ImageSizeWrapper, SizeType, VLMTestInfo)
@@ -69,7 +69,7 @@ def get_model_prompts(base_prompts: Iterable[str],
def build_single_image_inputs_from_test_info(
test_info: VLMTestInfo,
image_assets: _ImageAssets,
image_assets: ImageTestAssets,
size_wrapper: ImageSizeWrapper,
tmp_path: Optional[PosixPath] = None):
if test_info.prompt_formatter is None:
@@ -116,7 +116,7 @@ def build_single_image_inputs(images, model_prompts,
def build_multi_image_inputs_from_test_info(
test_info: VLMTestInfo,
image_assets: _ImageAssets,
image_assets: ImageTestAssets,
size_wrapper: ImageSizeWrapper,
tmp_path: Optional[PosixPath] = None):
if test_info.prompt_formatter is None:
@@ -159,7 +159,7 @@ def build_multi_image_inputs(image_lists, model_prompts,
def build_embedding_inputs_from_test_info(
test_info: VLMTestInfo,
image_assets: _ImageAssets,
image_assets: ImageTestAssets,
size_wrapper: ImageSizeWrapper,
):
# These conditions will always be true if invoked through filtering,
@@ -192,7 +192,7 @@ def build_embedding_inputs_from_test_info(
def build_video_inputs_from_test_info(
test_info: VLMTestInfo,
video_assets: _VideoAssets,
video_assets: VideoTestAssets,
size_wrapper: ImageSizeWrapper,
num_frames: int,
):

View File

@@ -16,7 +16,7 @@ from transformers import (AutoConfig, AutoTokenizer, BatchFeature,
from vllm.sequence import SampleLogprobs
from vllm.transformers_utils.tokenizer import patch_padding_side
from .....conftest import HfRunner, ImageAsset, _ImageAssets
from .....conftest import HfRunner, ImageAsset, ImageTestAssets
from .types import RunnerOutput
@@ -238,14 +238,14 @@ def minimax_vl_01_hf_output(hf_output: RunnerOutput,
####### Functions for converting image assets to embeddings
def get_llava_embeddings(image_assets: _ImageAssets):
def get_llava_embeddings(image_assets: ImageTestAssets):
return [asset.image_embeds for asset in image_assets]
####### Prompt path encoders for models that need models on disk
def qwen_prompt_path_encoder(
tmp_path: PosixPath, prompt: str, assets: Union[list[ImageAsset],
_ImageAssets]) -> str:
tmp_path: PosixPath, prompt: str,
assets: Union[list[ImageAsset], ImageTestAssets]) -> str:
"""Given a temporary dir path, export one or more image assets into the
tempdir & replace its contents with the local path to the string so that
the HF version of Qwen-VL can resolve the path and load the image in its

View File

@@ -4,7 +4,8 @@ types / modalities.
"""
from pathlib import PosixPath
from .....conftest import HfRunner, VllmRunner, _ImageAssets, _VideoAssets
from .....conftest import (HfRunner, ImageTestAssets, VideoTestAssets,
VllmRunner)
from . import builders, core
from .types import ExpandableVLMTestArgs, VLMTestInfo
@@ -14,7 +15,7 @@ def run_single_image_test(*, tmp_path: PosixPath, model_test_info: VLMTestInfo,
test_case: ExpandableVLMTestArgs,
hf_runner: type[HfRunner],
vllm_runner: type[VllmRunner],
image_assets: _ImageAssets):
image_assets: ImageTestAssets):
assert test_case.size_wrapper is not None
inputs = builders.build_single_image_inputs_from_test_info(
model_test_info, image_assets, test_case.size_wrapper, tmp_path)
@@ -37,7 +38,7 @@ def run_multi_image_test(*, tmp_path: PosixPath, model_test_info: VLMTestInfo,
test_case: ExpandableVLMTestArgs,
hf_runner: type[HfRunner],
vllm_runner: type[VllmRunner],
image_assets: _ImageAssets):
image_assets: ImageTestAssets):
assert test_case.size_wrapper is not None
inputs = builders.build_multi_image_inputs_from_test_info(
model_test_info, image_assets, test_case.size_wrapper, tmp_path)
@@ -60,7 +61,7 @@ def run_embedding_test(*, model_test_info: VLMTestInfo,
test_case: ExpandableVLMTestArgs,
hf_runner: type[HfRunner],
vllm_runner: type[VllmRunner],
image_assets: _ImageAssets):
image_assets: ImageTestAssets):
assert test_case.size_wrapper is not None
inputs, vllm_embeddings = builders.build_embedding_inputs_from_test_info(
model_test_info, image_assets, test_case.size_wrapper)
@@ -86,7 +87,7 @@ def run_video_test(
test_case: ExpandableVLMTestArgs,
hf_runner: type[HfRunner],
vllm_runner: type[VllmRunner],
video_assets: _VideoAssets,
video_assets: VideoTestAssets,
):
assert test_case.size_wrapper is not None
assert test_case.num_video_frames is not None

View File

@@ -15,7 +15,7 @@ from vllm.config import TaskOption
from vllm.sequence import SampleLogprobs
from vllm.transformers_utils.tokenizer import AnyTokenizer
from .....conftest import IMAGE_ASSETS, HfRunner, ImageAsset, _ImageAssets
from .....conftest import IMAGE_ASSETS, HfRunner, ImageAsset, ImageTestAssets
from ....utils import check_logprobs_close
# meta image tag; will be replaced by the appropriate tag for the model
@@ -85,7 +85,7 @@ class VLMTestInfo(NamedTuple):
# Function for converting ImageAssets to image embeddings;
# We need to define this explicitly for embedding tests
convert_assets_to_embeddings: Optional[Callable[[_ImageAssets],
convert_assets_to_embeddings: Optional[Callable[[ImageTestAssets],
torch.Tensor]] = None
# Exposed options for vLLM runner; we change these in a several tests,
@@ -141,7 +141,7 @@ class VLMTestInfo(NamedTuple):
# for Qwen-VL, which requires encoding the image path / url into the prompt
# for HF runner
prompt_path_encoder: Optional[
Callable[[PosixPath, str, Union[list[ImageAsset], _ImageAssets]],
Callable[[PosixPath, str, Union[list[ImageAsset], ImageTestAssets]],
str]] = None # noqa: E501
# Allows configuring a test to run with custom inputs