[Bugfix] Fix tensor parallel issue in Qwen3 reranker weight loading

tests/models/language/pooling/test_qwen3_reranker.py

@@ -6,6 +6,7 @@
 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 +88,26 @@ 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)

vllm/model_executor/models/adapters.py

@@ -322,15 +322,15 @@ def load_weights_using_from_2_way_softmax(
     # refer to https://huggingface.co/Qwen/Qwen3-Reranker-0.6B/discussions/3
     from vllm.model_executor.layers.vocab_parallel_embedding import (
         ParallelLMHead)
+    from vllm.model_executor.model_loader.weight_utils import (
+        default_weight_loader)
     from vllm.model_executor.models.utils import AutoWeightsLoader
 
     model_config = model.vllm_config.model_config
     tokens = getattr(model.config, "classifier_from_token", [])
     tokens = cast(list[int], tokens)
     assert len(tokens) == 2
 
-    device = model.score.weight.device
-
     if model.config.tie_word_embeddings:
         model.lm_head = model.model.embed_tokens
     else:
@@ -349,10 +349,13 @@ def load_weights_using_from_2_way_softmax(
 
     false_id = tokenizer.convert_tokens_to_ids(tokens[0])
     true_id = tokenizer.convert_tokens_to_ids(tokens[1])
-    weight = model.lm_head.weight.data[true_id].to(device).to(
-        torch.float32) - model.lm_head.weight.data[false_id].to(device).to(
+    weight = model.lm_head.weight.data[[true_id]].to(
+        torch.float32) - model.lm_head.weight.data[[false_id]].to(
             torch.float32)
-    model.score.weight.data.copy_(weight)
+
+    param = model.score.weight
+    weight_loader = getattr(param, "weight_loader", default_weight_loader)
+    weight_loader(param, weight)
 
     del model.lm_head
     loaded_weights.add("score.weight")