[Bugfix] Fix tensor parallel issue in Qwen3 reranker weight loading (#20682)

Signed-off-by: Isotr0py <2037008807@qq.com>
Co-authored-by: Isotr0py <2037008807@qq.com>

tests/models/language/pooling/mteb_utils.py

@@ -268,7 +268,8 @@ def mteb_test_rerank_models(hf_runner,
                             model_info: RerankModelInfo,
                             vllm_extra_kwargs=None,
                             hf_model_callback=None,
-                            vllm_mteb_encoder=VllmMtebEncoder):
+                            vllm_mteb_encoder=VllmMtebEncoder,
+                            atol=MTEB_RERANK_TOL):
     if not model_info.enable_test:
         # A model family has many models with the same architecture,
         # and we don't need to test each one.
@@ -301,4 +302,4 @@ def mteb_test_rerank_models(hf_runner,
     print("SentenceTransformers:", st_dtype, st_main_score)
     print("Difference:", st_main_score - vllm_main_score)
 
-    assert st_main_score == pytest.approx(vllm_main_score, abs=MTEB_RERANK_TOL)
+    assert st_main_score == pytest.approx(vllm_main_score, abs=atol)

tests/models/language/pooling/test_qwen3_reranker.py

@@ -6,6 +6,7 @@ import pytest
 import torch
 
 from tests.conftest import HfRunner
+from tests.utils import multi_gpu_test
 
 from .mteb_utils import RerankModelInfo, mteb_test_rerank_models
 
@@ -87,3 +88,29 @@ def test_rerank_models_mteb(vllm_runner, model_info: RerankModelInfo) -> None:
 
     mteb_test_rerank_models(Qwen3RerankerHfRunner, vllm_runner, model_info,
                             vllm_extra_kwargs)
+
+
+@pytest.mark.parametrize("model_info", RERANK_MODELS)
+@multi_gpu_test(num_gpus=2)
+def test_rerank_models_mteb_tp(vllm_runner,
+                               model_info: RerankModelInfo) -> None:
+
+    assert model_info.architecture == "Qwen3ForSequenceClassification"
+
+    vllm_extra_kwargs: dict[str, Any] = {
+        "hf_overrides": {
+            "architectures": ["Qwen3ForSequenceClassification"],
+            "classifier_from_token": ["no", "yes"],
+            "is_original_qwen3_reranker": True,
+        },
+        "tensor_parallel_size": 2,
+    }
+
+    if model_info.name == "Qwen/Qwen3-Reranker-4B":
+        vllm_extra_kwargs["max_num_seqs"] = 1
+
+    mteb_test_rerank_models(Qwen3RerankerHfRunner,
+                            vllm_runner,
+                            model_info,
+                            vllm_extra_kwargs,
+                            atol=1.2e-2)