mirror of
https://github.com/vllm-project/vllm.git
synced 2025-12-06 15:04:47 +08:00
62 lines
1.7 KiB
Python
62 lines
1.7 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
|
|
import weakref
|
|
|
|
import pytest
|
|
import torch
|
|
|
|
from tests.models.utils import softmax
|
|
from vllm import LLM, PoolingParams
|
|
from vllm.distributed import cleanup_dist_env_and_memory
|
|
|
|
MODEL_NAME = "tomaarsen/Qwen3-Reranker-0.6B-seq-cls"
|
|
|
|
|
|
@pytest.fixture(scope="module")
|
|
def llm():
|
|
# pytest caches the fixture so we use weakref.proxy to
|
|
# enable garbage collection
|
|
llm = LLM(
|
|
model=MODEL_NAME,
|
|
max_num_batched_tokens=32768,
|
|
tensor_parallel_size=1,
|
|
gpu_memory_utilization=0.75,
|
|
enforce_eager=True,
|
|
seed=0,
|
|
)
|
|
|
|
yield weakref.proxy(llm)
|
|
|
|
del llm
|
|
|
|
cleanup_dist_env_and_memory()
|
|
|
|
|
|
def test_pooling_params(llm: LLM):
|
|
def get_outputs(activation):
|
|
text_1 = "What is the capital of France?"
|
|
text_2 = "The capital of France is Paris."
|
|
|
|
outputs = llm.score(
|
|
text_1,
|
|
text_2,
|
|
pooling_params=PoolingParams(activation=activation),
|
|
use_tqdm=False,
|
|
)
|
|
return torch.tensor([x.outputs.score for x in outputs])
|
|
|
|
default = get_outputs(activation=None)
|
|
w_activation = get_outputs(activation=True)
|
|
wo_activation = get_outputs(activation=False)
|
|
|
|
assert torch.allclose(default, w_activation, atol=1e-2), (
|
|
"Default should use activation."
|
|
)
|
|
assert not torch.allclose(w_activation, wo_activation, atol=1e-2), (
|
|
"wo_activation should not use activation."
|
|
)
|
|
assert torch.allclose(softmax(wo_activation), w_activation, atol=1e-2), (
|
|
"w_activation should be close to activation(wo_activation)."
|
|
)
|