[Bugfix][Frontend] Honor the constructor chat template in offline scoring (#61990)

Signed-off-by: mottopanikeiku <176798723+mottopanikeiku@users.noreply.github.com>
Co-authored-by: mottopanikeiku <176798723+mottopanikeiku@users.noreply.github.com>

docs/models/pooling_models/scoring.md

@@ -455,6 +455,8 @@ Score templates are supported for **cross-encoder** models only. If you are usin
 
 Some scoring models require a specific prompt format to work correctly. You can specify a custom score template using the `--chat-template` parameter (see [Chat Template](../../serving/online_serving/README.md#chat-template)).
 
+For offline scoring, pass `chat_template` to the `LLM` constructor to use it for every `LLM.score` call. A `chat_template` passed to `LLM.score` overrides the constructor's template for that call.
+
 Like chat templates, the score template receives a `messages` list. For scoring, each message has a `role` attribute—either `"query"` or `"document"`. For the usual kind of point-wise cross-encoder, you can expect exactly two messages: one query and one document. To access the query and document content, use Jinja's `selectattr` filter:
 
 - **Query**: `{{ (messages | selectattr("role", "eq", "query") | first).content }}`

tests/entrypoints/pooling/scoring/test_io_processor_unit.py

@@ -10,7 +10,7 @@ import torch
 from tokenizers import Tokenizer, models, pre_tokenizers, processors
 from transformers import TokenizersBackend
 
-from vllm import TokensPrompt
+from vllm import PoolingParams, TokensPrompt
 from vllm.entrypoints.pooling.scoring.io_processor import (
     BiEncoderIOProcessor,
     CrossEncoderIOProcessor,
@@ -18,7 +18,9 @@ from vllm.entrypoints.pooling.scoring.io_processor import (
     LateInteractionIOProcessor,
     _apply_post_tokenization_to_token_type_ids,
 )
+from vllm.entrypoints.pooling.scoring.typing import ScoringData
 from vllm.entrypoints.pooling.scoring.utils import compress_token_type_ids
+from vllm.entrypoints.pooling.typing import OfflineScoringInputsContext
 from vllm.outputs import PoolingOutput, PoolingRequestOutput, RequestError
 from vllm.renderers import TokenizeParams
 
@@ -81,6 +83,49 @@ def test_llm_reranker_tokenization_is_independent_of_nonbinding_doc_limit(
     assert nonbinding["prompt_token_ids"] == expected
 
 
+@pytest.mark.parametrize(
+    "default_template,request_template,expected",
+    [
+        ("{{ messages[0].content }} {{ messages[1].content }}", None, "a bc"),
+        (
+            "{{ messages[0].content }} {{ messages[1].content }}",
+            "{{ messages[1].content }} {{ messages[0].content }}",
+            "bc a",
+        ),
+        (None, "{{ messages[1].content }} {{ messages[0].content }}", "bc a"),
+        (None, None, "abc"),
+    ],
+)
+def test_offline_scoring_uses_default_template_unless_overridden(
+    llm_reranker_processor, default_template, request_template, expected
+):
+    processor = llm_reranker_processor
+    processor.chat_template = default_template
+    processor.renderer = SimpleNamespace(
+        default_cmpl_tok_params=TokenizeParams(max_total_tokens=32)
+    )
+    ctx = OfflineScoringInputsContext(
+        pooling_task="classify",
+        tokenization_kwargs=None,
+        lora_request=None,
+        priorities=None,
+        scoring_data=ScoringData(data_1=["a"], data_2=["bc"]),
+        chat_template=request_template,
+        pooling_params=PoolingParams(),
+    )
+    factory, _ = processor.get_request_factory_offline(ctx)
+    (request,) = factory()
+    text, prompt = processor.get_score_prompt(
+        request["data_1"],
+        request["data_2"],
+        request["tok_params"].get_encode_kwargs(),
+        chat_template=request["chat_template"],
+    )
+
+    assert text == expected
+    assert prompt["prompt_token_ids"] == processor.tokenizer.encode(expected)
+
+
 def test_token_type_ids_stay_aligned_with_a_truncated_padded_prompt():
     """The cross-encoder segment boundary must survive truncate + pad.
 

vllm/entrypoints/pooling/offline.py

@@ -325,8 +325,8 @@ class PoolingOfflineMixin(OfflineInferenceMixin):
                 it is used to create the progress bar.
                 If `False`, no progress bar is created.
             lora_request: LoRA request to use for generation, if any.
-            chat_template: The chat template to use for the scoring. If None, we
-                use the model's default chat template.
+            chat_template: The chat template to use for scoring. If None, use
+                the chat template passed to the LLM constructor.
             tokenization_kwargs: Overrides for `tokenizer.encode`.
 
         Returns:

vllm/entrypoints/pooling/scoring/io_processor.py

@@ -548,7 +548,11 @@ class CrossEncoderIOProcessor(ScoringIOProcessor):
                 yield ScoringRenderParams(
                     data_1=data_1[i],
                     data_2=data_2[i],
-                    chat_template=ctx.chat_template,
+                    chat_template=(
+                        self.chat_template
+                        if ctx.chat_template is None
+                        else ctx.chat_template
+                    ),
                     max_tokens_per_query=max_tokens_per_query,
                     max_tokens_per_doc=max_tokens_per_doc,
                     tok_params=tok_params,