[Bugfix][Frontend] Do not modify caller params in LLM.score() (#59959)

Signed-off-by: 叮当猫 <2756823972@qq.com>
Signed-off-by: apex-mochen <2756823972@qq.com>
Co-authored-by: DeepSeek Harness <noreply@deepseek.com>
Co-authored-by: Taneem Ibrahim <taneem.ibrahim@gmail.com>
Co-authored-by: Codex <noreply@openai.com>

tests/entrypoints/pooling/test_io_processor.py

@@ -7,6 +7,8 @@ import pytest
 
 from vllm import PoolingParams
 from vllm.entrypoints.pooling.base.io_processor import PoolingIOProcessor
+from vllm.entrypoints.pooling.offline import PoolingOfflineMixin
+from vllm.entrypoints.pooling.scoring.io_processor import CrossEncoderIOProcessor
 from vllm.entrypoints.pooling.typing import OfflineEncodeInputsContext
 from vllm.exceptions import VLLMValidationError
 from vllm.renderers import TokenizeParams
@@ -221,3 +223,38 @@ def test_qwen3_reranker_no_warning_when_template_provided(monkeypatch):
     )
 
     mock_logger.warning.assert_not_called()
+
+
+def _make_score_llm(monkeypatch) -> PoolingOfflineMixin:
+    """A scoring LLM with no model behind it.
+
+    Input validation runs for real; only the rendering, engine and
+    post-processing boundaries are stubbed out. The pooling task is filled in
+    before any of them, so the mutation is observable without inference.
+    """
+    proc = CrossEncoderIOProcessor.__new__(CrossEncoderIOProcessor)
+    proc.is_multimodal_model = False
+    proc.architecture = "CrossEncoder"
+    monkeypatch.setattr(proc, "get_request_factory_offline", lambda ctx: (None, 0))
+    monkeypatch.setattr(proc, "post_process_offline", lambda ctx: [])
+
+    llm = PoolingOfflineMixin.__new__(PoolingOfflineMixin)
+    llm.runner_type = "pooling"
+    llm.pooling_task = "classify"  # SCORE_TYPE_MAP -> "cross-encoder"
+    llm.model_config = SimpleNamespace(hf_config=SimpleNamespace(num_labels=1))
+    llm.pooling_io_processors = {"cross-encoder": proc}
+    monkeypatch.setattr(llm, "_run_tiling_engine", lambda *args, **kwargs: [])
+    return llm
+
+
+def test_score_leaves_caller_pooling_params_untouched(monkeypatch):
+    llm = _make_score_llm(monkeypatch)
+
+    params = PoolingParams()
+    llm.score("query", "doc", pooling_params=params)
+
+    assert params.task is None, (
+        "LLM.score() wrote the model's own pooling task into the caller's "
+        "PoolingParams; reusing that object with another pooling model now "
+        "fails with 'You cannot overwrite ...' before any inference runs."
+    )

vllm/entrypoints/pooling/offline.py

@@ -362,6 +362,8 @@ class PoolingOfflineMixin(OfflineInferenceMixin):
 
         assert isinstance(pooling_params, PoolingParams)
         pooling_task = io_processor.pooling_task
+        # Clone to avoid modifying the caller's pooling parameters.
+        pooling_params = pooling_params.clone()
         if pooling_params.task is None:
             pooling_params.task = pooling_task
         elif pooling_params.task != pooling_task: