[Model] Let more models to support the score template. (#31335)

Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
Signed-off-by: wang.yuqi <noooop@126.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>

tests/models/language/pooling_mteb_test/test_qwen3_reranker.py

 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# ruff: noqa: E501
 from typing import Any
 
+import mteb
+import numpy as np
 import pytest
 import torch
+from torch.utils.data import DataLoader
 
 from tests.conftest import HfRunner
 from tests.models.utils import RerankModelInfo
 from tests.utils import multi_gpu_test
 
-from .mteb_score_utils import mteb_test_rerank_models
+from .mteb_score_utils import MtebCrossEncoderMixin, mteb_test_rerank_models
 
 qwen3_reranker_hf_overrides = {
     "architectures": ["Qwen3ForSequenceClassification"],
@@ -21,51 +25,71 @@ RERANK_MODELS = [
     RerankModelInfo(
         "Qwen/Qwen3-Reranker-0.6B",
         architecture="Qwen3ForSequenceClassification",
-        mteb_score=0.25736,
         hf_overrides=qwen3_reranker_hf_overrides,
+        chat_template_name="qwen3_reranker.jinja",
         pooling_type="LAST",
         attn_type="decoder",
         is_prefix_caching_supported=True,
         is_chunked_prefill_supported=True,
+        mteb_score=0.33459,
         enable_test=True,
     ),
     RerankModelInfo(
         "Qwen/Qwen3-Reranker-4B",
         architecture="Qwen3ForSequenceClassification",
+        chat_template_name="qwen3_reranker.jinja",
         hf_overrides=qwen3_reranker_hf_overrides,
         enable_test=False,
     ),
 ]
 
 
-class Qwen3RerankerHfRunner(HfRunner):
+class Qwen3RerankerHfRunner(MtebCrossEncoderMixin, HfRunner):
     def __init__(
         self, model_name: str, dtype: str = "auto", *args: Any, **kwargs: Any
     ) -> None:
         from transformers import AutoModelForCausalLM, AutoTokenizer
 
-        super().__init__(model_name, dtype, auto_cls=AutoModelForCausalLM)
+        HfRunner.__init__(
+            self,
+            model_name=model_name,
+            auto_cls=AutoModelForCausalLM,
+            dtype=dtype,
+            **kwargs,
+        )
 
         self.tokenizer = AutoTokenizer.from_pretrained(model_name, padding_side="left")
         self.token_false_id = self.tokenizer.convert_tokens_to_ids("no")
         self.token_true_id = self.tokenizer.convert_tokens_to_ids("yes")
-
-    def predict(self, prompts: list[list[str]], *args, **kwargs) -> torch.Tensor:
-        def process_inputs(pairs):
-            inputs = self.tokenizer(
-                pairs,
-                padding=False,
-                truncation="longest_first",
-                return_attention_mask=False,
+        self.max_length = 40960
+
+    @torch.no_grad
+    def predict(
+        self,
+        inputs1: DataLoader[mteb.types.BatchedInput],
+        inputs2: DataLoader[mteb.types.BatchedInput],
+        *args,
+        **kwargs,
+    ) -> np.ndarray:
+        queries = [text for batch in inputs1 for text in batch["text"]]
+        corpus = [text for batch in inputs2 for text in batch["text"]]
+
+        tokenizer = self.tokenizer
+        prompts = []
+        for query, document in zip(queries, corpus):
+            conversation = [
+                {"role": "query", "content": query},
+                {"role": "document", "content": document},
+            ]
+
+            prompt = tokenizer.apply_chat_template(
+                conversation=conversation,
+                tools=None,
+                chat_template=self.chat_template,
+                tokenize=False,
             )
-            for i, ele in enumerate(inputs["input_ids"]):
-                inputs["input_ids"][i] = ele
-            inputs = self.tokenizer.pad(inputs, padding=True, return_tensors="pt")
-            for key in inputs:
-                inputs[key] = inputs[key].to(self.model.device)
-            return inputs
-
-        @torch.no_grad()
+            prompts.append(prompt)
+
         def compute_logits(inputs):
             batch_scores = self.model(**inputs).logits[:, -1, :]
             true_vector = batch_scores[:, self.token_true_id]
@@ -76,9 +100,9 @@ class Qwen3RerankerHfRunner(HfRunner):
             return scores
 
         scores = []
-        for query, doc, *_ in prompts:
-            pairs = [(query, doc)]
-            inputs = process_inputs(pairs)
+        for prompt in prompts:
+            inputs = tokenizer([prompt], return_tensors="pt")
+            inputs = self.wrap_device(inputs)
             score = compute_logits(inputs)
             scores.append(score[0].item())
         return torch.Tensor(scores)
@@ -86,7 +110,7 @@ class Qwen3RerankerHfRunner(HfRunner):
 
 @pytest.mark.parametrize("model_info", RERANK_MODELS)
 def test_rerank_models_mteb(vllm_runner, model_info: RerankModelInfo) -> None:
-    mteb_test_rerank_models(Qwen3RerankerHfRunner, vllm_runner, model_info)
+    mteb_test_rerank_models(vllm_runner, model_info, hf_runner=Qwen3RerankerHfRunner)
 
 
 @pytest.mark.parametrize("model_info", RERANK_MODELS)
@@ -99,5 +123,8 @@ def test_rerank_models_mteb_tp(vllm_runner, model_info: RerankModelInfo) -> None
     }
 
     mteb_test_rerank_models(
-        Qwen3RerankerHfRunner, vllm_runner, model_info, vllm_extra_kwargs
+        vllm_runner,
+        model_info,
+        vllm_extra_kwargs=vllm_extra_kwargs,
+        hf_runner=Qwen3RerankerHfRunner,
     )

vllm/model_executor/models/config.py

@@ -251,7 +251,7 @@ class Qwen3ForSequenceClassificationConfig(VerifyAndUpdateConfig):
         tokens = getattr(config, "classifier_from_token", None)
         assert tokens is not None and len(tokens) == 2, (
             "Try loading the original Qwen3 Reranker?, see: "
-            "https://github.com/vllm-project/vllm/tree/main/examples/offline_inference/offline_reranker.py"
+            "https://github.com/vllm-project/vllm/tree/main/examples/pooling/score/qwen3_reranker_offline.py"
         )
         model_config.hf_config.method = "from_2_way_softmax"