[Refactor] Remove dead tests utils (#58803)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2026-09-26 09:16:53 -04:00
committed by GitHub
parent 2bb7605ae1
commit bf93e1f92c
16 changed files with 3 additions and 1336 deletions
@@ -6,15 +6,8 @@ from typing import TYPE_CHECKING, Any
import pytest
import torch
from mistral_common.multimodal import download_image
from mistral_common.protocol.instruct.chunk import ImageURLChunk
from mistral_common.protocol.instruct.request import ChatCompletionRequest
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
from mistral_common.tokens.tokenizers.multimodal import image_from_chunk
from transformers import AutoProcessor
from vllm import SamplingParams, TextPrompt, TokensPrompt
from vllm.inputs import MultiModalDataBuiltins
from vllm import SamplingParams
from vllm.logprobs import Logprob, SampleLogprobs
from vllm.model_executor.models.pixtral import _make_packed_sequence_metadata
from vllm.platforms import current_platform
@@ -56,59 +49,6 @@ def _create_msg_format(urls: list[str]) -> list[dict[str, Any]]:
]
def _create_msg_format_hf(urls: list[str]) -> list[dict[str, Any]]:
return [
{
"role": "user",
"content": [
{
"type": "text",
"content": PROMPT,
},
*({"type": "image", "image": download_image(url)} for url in urls),
],
}
]
def _create_engine_inputs(urls: list[str]) -> TokensPrompt:
msg = _create_msg_format(urls)
tokenizer = MistralTokenizer.from_model("pixtral")
request = ChatCompletionRequest(messages=msg) # type: ignore[type-var]
tokenized = tokenizer.encode_chat_completion(request)
engine_inputs = TokensPrompt(prompt_token_ids=tokenized.tokens)
images = []
for chunk in request.messages[0].content:
if isinstance(chunk, ImageURLChunk):
images.append(image_from_chunk(chunk))
mm_data = MultiModalDataBuiltins(image=images)
engine_inputs["multi_modal_data"] = mm_data
return engine_inputs
def _create_engine_inputs_hf(urls: list[str]) -> TextPrompt:
msg = _create_msg_format_hf(urls)
tokenizer = AutoProcessor.from_pretrained("mistral-community/pixtral-12b")
prompt = tokenizer.apply_chat_template(msg)
images = []
for chunk in msg[0]["content"]:
if chunk["type"] == "image":
images.append(chunk["image"])
mm_data = MultiModalDataBuiltins(image=images)
engine_inputs = TextPrompt(prompt=prompt, multi_modal_data=mm_data)
return engine_inputs
SAMPLING_PARAMS = SamplingParams(max_tokens=512, temperature=0.0, logprobs=5)
LIMIT_MM_PER_PROMPT = dict(image=4)
@@ -6,7 +6,6 @@ typically specific to a small subset of models.
"""
import types
from pathlib import PosixPath
import numpy as np
import numpy.typing as npt
@@ -29,7 +28,7 @@ from vllm.logprobs import SampleLogprobs
from vllm.platforms import current_platform
from vllm.utils.collection_utils import is_list_of
from .....conftest import HfRunner, ImageAsset, ImageTestAssets
from .....conftest import HfRunner, ImageTestAssets
from .types import RunnerOutput
@@ -48,17 +47,6 @@ def blip2_vllm_to_hf_output(vllm_output: RunnerOutput, model: str) -> RunnerOutp
return hf_output_ids, hf_output_str, out_logprobs
def qwen_vllm_to_hf_output(
vllm_output: RunnerOutput, model: str
) -> tuple[list[int], str, SampleLogprobs | None]:
"""Sanitize vllm output [qwen models] to be comparable with hf output."""
output_ids, output_str, out_logprobs = vllm_output
hf_output_str = output_str + "<|endoftext|>"
return output_ids, hf_output_str, out_logprobs
def qwen2_vllm_to_hf_output(
vllm_output: RunnerOutput, model: str
) -> tuple[list[int], str, SampleLogprobs | None]:
@@ -239,37 +227,6 @@ 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: 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
forward() call.
Args:
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.
assert prompt.count("<img></img>") == len(assets)
# Replace the placeholders with local paths to the exported assets
for asset in assets:
image_tmp_path = tmp_path / f"{asset.name}.jpg"
asset.pil_image.save(image_tmp_path)
prompt = prompt.replace(
"<img></img>",
f"<img>{image_tmp_path}</img>",
1,
)
return prompt
####### Model-specific HuggingFace runner patchers
def deepseekvl2_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
"""Patches and returns an instance of the HfRunner to use for GLM4."""