[Model] Pooling model activation supports per request control by PoolingParams (#20538)

Signed-off-by: wang.yuqi <noooop@126.com>
This commit is contained in:
wang.yuqi
2025-08-05 00:37:00 -07:00
committed by GitHub
parent 811ac13d03
commit 586f286789
21 changed files with 948 additions and 173 deletions
+67
View File
@@ -0,0 +1,67 @@
# 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(autouse=True)
def v1(run_with_both_engines):
# Simple autouse wrapper to run both engines for each test
# This can be promoted up to conftest.py to run for every
# test in a package
pass
@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)
with llm.deprecate_legacy_api():
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)."
+56
View File
@@ -0,0 +1,56 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import weakref
import pytest
import torch
import torch.nn.functional as F
from vllm import LLM, PoolingParams
from vllm.distributed import cleanup_dist_env_and_memory
MODEL_NAME = "intfloat/multilingual-e5-small"
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)
with llm.deprecate_legacy_api():
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(normalize):
outputs = llm.embed(prompts,
pooling_params=PoolingParams(normalize=normalize),
use_tqdm=False)
return torch.tensor([x.outputs.embedding for x in outputs])
default = get_outputs(normalize=None)
w_normal = get_outputs(normalize=True)
wo_normal = get_outputs(normalize=False)
assert torch.allclose(default, w_normal,
atol=1e-2), "Default should use normal."
assert not torch.allclose(w_normal, wo_normal,
atol=1e-2), "wo_normal should not use normal."
assert torch.allclose(
w_normal, F.normalize(wo_normal, p=2, dim=-1),
atol=1e-2), "w_normal should be close to normal(wo_normal)."
+66
View File
@@ -0,0 +1,66 @@
# 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 = "internlm/internlm2-1_8b-reward"
prompts = ["The chef prepared a delicious meal."]
@pytest.fixture(autouse=True)
def v1(run_with_both_engines):
# Simple autouse wrapper to run both engines for each test
# This can be promoted up to conftest.py to run for every
# test in a package
pass
@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,
trust_remote_code=True,
seed=0)
with llm.deprecate_legacy_api():
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(softmax):
outputs = llm.reward(prompts,
pooling_params=PoolingParams(softmax=softmax),
use_tqdm=False)
return torch.cat([x.outputs.data for x in outputs])
default = get_outputs(softmax=None)
w_softmax = get_outputs(softmax=True)
wo_softmax = get_outputs(softmax=False)
assert torch.allclose(default, w_softmax,
atol=1e-2), "Default should use softmax."
assert not torch.allclose(w_softmax, wo_softmax,
atol=1e-2), "wo_softmax should not use softmax."
assert torch.allclose(
softmax(wo_softmax), w_softmax,
atol=1e-2), "w_softmax should be close to softmax(wo_softmax)."
+69
View File
@@ -0,0 +1,69 @@
# 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 = "tomaarsen/Qwen3-Reranker-0.6B-seq-cls"
@pytest.fixture(autouse=True)
def v1(run_with_both_engines):
# Simple autouse wrapper to run both engines for each test
# This can be promoted up to conftest.py to run for every
# test in a package
pass
@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)
with llm.deprecate_legacy_api():
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):
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)."