xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-16 06:45:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Use ReplicatedLinear for SequenceClassification head (#23836)

Browse Source 
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Isotr0py 2025-08-29 12:41:20 +08:00 committed by GitHub
parent c8b3b299c9
commit 98ac0cb32d
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 4 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
tests/models/language/pooling/test_qwen3_reranker.py
View File

@ -96,8 +96,5 @@ def test_rerank_models_mteb_tp(vllm_runner,
"tensor_parallel_size": 2, "tensor_parallel_size": 2,
} }
mteb_test_rerank_models(Qwen3RerankerHfRunner, mteb_test_rerank_models(Qwen3RerankerHfRunner, vllm_runner, model_info,
vllm_runner, vllm_extra_kwargs)
model_info,
vllm_extra_kwargs,
atol=1.2e-2)

5
vllm/model_executor/models/adapters.py
View File

@ -248,7 +248,7 @@ def as_seq_cls_model(cls: _T) -> _T:
return cls return cls
# Lazy import # Lazy import
from vllm.model_executor.layers.linear import RowParallelLinear from vllm.model_executor.layers.linear import ReplicatedLinear
from vllm.model_executor.layers.pooler import (ClassifierPooler, from vllm.model_executor.layers.pooler import (ClassifierPooler,
DispatchPooler, Pooler, DispatchPooler, Pooler,
PoolingMethod, PoolingType) PoolingMethod, PoolingType)
@ -264,10 +264,9 @@ def as_seq_cls_model(cls: _T) -> _T:
config = vllm_config.model_config.hf_config config = vllm_config.model_config.hf_config
quant_config = vllm_config.quant_config quant_config = vllm_config.quant_config
self.score = RowParallelLinear( self.score = ReplicatedLinear(
config.hidden_size, config.hidden_size,
config.num_labels, config.num_labels,
input_is_parallel=False,
bias=False, bias=False,
params_dtype=torch.float32, params_dtype=torch.float32,
quant_config=quant_config, quant_config=quant_config,