mirror of
https://github.com/vllm-project/vllm.git
synced 2025-10-20 23:03:52 +08:00
2554b27baa
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu> Signed-off-by: Max de Bayser <mbayser@br.ibm.com> Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
65 lines
1.8 KiB
Python
65 lines
1.8 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
|
|
import weakref
|
|
|
|
import pytest
|
|
import torch
|
|
|
|
from vllm import LLM, PoolingParams
|
|
from vllm.distributed import cleanup_dist_env_and_memory
|
|
|
|
from ...models.utils import softmax
|
|
|
|
MODEL_NAME = "jason9693/Qwen2.5-1.5B-apeach"
|
|
|
|
prompts = ["The chef prepared a delicious meal."]
|
|
|
|
|
|
@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()
|
|
|
|
|
|
@pytest.mark.skip_global_cleanup
|
|
def test_pooling_params(llm: LLM):
|
|
|
|
def get_outputs(activation):
|
|
outputs = llm.classify(
|
|
prompts,
|
|
pooling_params=PoolingParams(activation=activation),
|
|
use_tqdm=False)
|
|
return torch.tensor([x.outputs.probs 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)."
|
|
|
|
|
|
def test_encode_api(llm: LLM):
|
|
err_msg = "pooling_task must be one of.+"
|
|
with pytest.raises(ValueError, match=err_msg):
|
|
llm.encode(prompts, use_tqdm=False)
|