A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Bugfix] Use `ReplicatedLinear` for SequenceClassification head (#23836)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Isotr0py committed
Aug 29, 2025 at 12:41 UTC
98ac0cb32d9462e50bd998f9f2eb6e4c09232c95
2 files changed
+4
-8
tests/models/language/pooling/test_qwen3_reranker.py
+2
-5
@@ -96,8 +96,5 @@ def test_rerank_models_mteb_tp(vllm_runner,
96
"tensor_parallel_size": 2,
97
}
98
99
- mteb_test_rerank_models(Qwen3RerankerHfRunner,
100
- vllm_runner,
101
- model_info,
102
- vllm_extra_kwargs,
103
- atol=1.2e-2)
99
+ mteb_test_rerank_models(Qwen3RerankerHfRunner, vllm_runner, model_info,
100
+ vllm_extra_kwargs)
vllm/model_executor/models/adapters.py
+2
-3
@@ -248,7 +248,7 @@ def as_seq_cls_model(cls: _T) -> _T:
248
return cls
249
250
# Lazy import
251
- from vllm.model_executor.layers.linear import RowParallelLinear
251
+ from vllm.model_executor.layers.linear import ReplicatedLinear
252
from vllm.model_executor.layers.pooler import (ClassifierPooler,
253
DispatchPooler, Pooler,
254
PoolingMethod, PoolingType)
@@ -264,10 +264,9 @@ def as_seq_cls_model(cls: _T) -> _T:
264
config = vllm_config.model_config.hf_config
265
quant_config = vllm_config.quant_config
266
267
- self.score = RowParallelLinear(
267
+ self.score = ReplicatedLinear(
268
config.hidden_size,
269
config.num_labels,
270
- input_is_parallel=False,
270
bias=False,
271
params_dtype=torch.float32,
272
quant_config=quant_config,