biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Move PoolerConfig from config/__init__.py to config/pooler.py (#25181)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-09-19 12:02:55 +01:00
committed by GitHub
parent 1dfea5f4a9
commit 058525b997
14 changed files with 193 additions and 179 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tests/models/language/pooling/test_embedding.py
View File

@@ -58,7 +58,7 @@ def test_models(
vllm_extra_kwargs = {}
if model == "ssmits/Qwen2-7B-Instruct-embed-base":
vllm_extra_kwargs["override_pooler_config"] = \
vllm_extra_kwargs["pooler_config"] = \
PoolerConfig(pooling_type="MEAN", normalize=False)
max_model_len: Optional[int] = 512

3
tests/models/language/pooling/test_mm_classifier_conversion.py
View File

@@ -1,6 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from vllm.config.pooler import PoolerConfig
from vllm.platforms import current_platform
@@ -99,7 +100,7 @@ def test_gemma_multimodal(
convert="classify",
load_format="auto",
hf_overrides=update_config,
override_pooler_config={"pooling_type": "LAST"},
pooler_config=PoolerConfig(pooling_type="LAST"),
max_model_len=512,
enforce_eager=True,
tensor_parallel_size=1,

62
tests/models/language/pooling/test_override_pooler_config.py → tests/models/language/pooling/test_pooler_config_init_behaviour.py
View File

@@ -24,18 +24,18 @@ def test_classify_models_using_activation(
dtype: str,
) -> None:
with vllm_runner(model,
max_model_len=512,
dtype=dtype,
override_pooler_config=PoolerConfig(
activation=False)) as vllm_model:
with vllm_runner(
model,
max_model_len=512,
dtype=dtype,
pooler_config=PoolerConfig(activation=False)) as vllm_model:
wo_activation_out = vllm_model.classify(example_prompts)
with vllm_runner(model,
max_model_len=512,
dtype=dtype,
override_pooler_config=PoolerConfig(
activation=True)) as vllm_model:
with vllm_runner(
model,
max_model_len=512,
dtype=dtype,
pooler_config=PoolerConfig(activation=True)) as vllm_model:
w_activation_out = vllm_model.classify(example_prompts)
for wo_activation, w_activation in zip(wo_activation_out,
@@ -43,9 +43,8 @@ def test_classify_models_using_activation(
wo_activation = torch.tensor(wo_activation)
w_activation = torch.tensor(w_activation)
assert not torch.allclose(
wo_activation, w_activation,
atol=1e-2), "override_pooler_config is not working"
assert not torch.allclose(wo_activation, w_activation,
atol=1e-2), "pooler_config is not working"
assert torch.allclose(softmax(wo_activation), w_activation,
1e-3 if dtype == "float" else 1e-2)
@@ -65,23 +64,22 @@ def test_embed_models_using_normalize(
dtype: str,
) -> None:
with vllm_runner(model,
max_model_len=512,
dtype=dtype,
override_pooler_config=PoolerConfig(
normalize=False)) as vllm_model:
wo_normalize = torch.tensor(vllm_model.embed(example_prompts))
with vllm_runner(
model,
max_model_len=512,
dtype=dtype,
override_pooler_config=PoolerConfig(normalize=True)) as vllm_model:
pooler_config=PoolerConfig(normalize=False)) as vllm_model:
wo_normalize = torch.tensor(vllm_model.embed(example_prompts))
with vllm_runner(model,
max_model_len=512,
dtype=dtype,
pooler_config=PoolerConfig(normalize=True)) as vllm_model:
w_normalize = torch.tensor(vllm_model.embed(example_prompts))
assert not torch.allclose(
wo_normalize, w_normalize,
atol=1e-2), "override_pooler_config normalize is not working"
atol=1e-2), "pooler_config normalize is not working"
assert torch.allclose(
F.normalize(wo_normalize, p=2, dim=-1), w_normalize,
atol=1e-2), "w_normal should be close to normal(wo_normal)."
@@ -102,18 +100,16 @@ def test_reward_models_using_softmax(
dtype: str,
) -> None:
with vllm_runner(
model,
max_model_len=1024,
dtype=dtype,
override_pooler_config=PoolerConfig(softmax=False)) as vllm_model:
with vllm_runner(model,
max_model_len=1024,
dtype=dtype,
pooler_config=PoolerConfig(softmax=False)) as vllm_model:
wo_softmax = vllm_model.encode(example_prompts)
with vllm_runner(
model,
max_model_len=1024,
dtype=dtype,
override_pooler_config=PoolerConfig(softmax=True)) as vllm_model:
with vllm_runner(model,
max_model_len=1024,
dtype=dtype,
pooler_config=PoolerConfig(softmax=True)) as vllm_model:
w_softmax = vllm_model.encode(example_prompts)
for wo, w in zip(wo_softmax, w_softmax):
@@ -121,7 +117,7 @@ def test_reward_models_using_softmax(
w = torch.tensor(w)
assert not torch.allclose(
wo, w, atol=1e-2), "override_pooler_config softmax is not working"
wo, w, atol=1e-2), "pooler_config softmax is not working"
assert torch.allclose(
softmax(wo), w,
atol=1e-2), "w_softmax should be close to softmax(wo_softmax)."