Decoder-only Scoring API (#6460)

Co-authored-by: Chanh Nguyen <cnguyen@linkedin.com>

python/sglang/srt/entrypoints/engine.py

@@ -472,6 +472,79 @@ class Engine(EngineBase):
     def save_sharded_model(self, **kwargs):
         self.collective_rpc("save_sharded_model", **kwargs)
 
+    def score(
+        self,
+        query: Optional[Union[str, List[int]]] = None,
+        items: Optional[Union[str, List[str], List[List[int]]]] = None,
+        label_token_ids: Optional[List[int]] = None,
+        apply_softmax: bool = False,
+        item_first: bool = False,
+    ) -> List[List[float]]:
+        """
+        Score the probability of specified token IDs appearing after the given (query + item) pair. For example:
+        query = "<|user|>Is the following city the capital of France? "
+        items = ["Paris <|assistant|>", "London <|assistant|>", "Berlin <|assistant|>"]
+        label_token_ids = [2332, 1223] # Token IDs for "Yes" and "No"
+        item_first = False
+
+        This would pass the following prompts to the model:
+        "<|user|>Is the following city the capital of France? Paris <|assistant|>"
+        "<|user|>Is the following city the capital of France? London <|assistant|>"
+        "<|user|>Is the following city the capital of France? Berlin <|assistant|>"
+        The api would then return the probabilities of the model producing "Yes" and "No" as the next token.
+        The output would look like:
+        [[0.9, 0.1], [0.2, 0.8], [0.1, 0.9]]
+
+
+        Args:
+            query: The query text or pre-tokenized query token IDs. Must be provided.
+            items: The item text(s) or pre-tokenized item token IDs. Must be provided.
+            label_token_ids: List of token IDs to compute probabilities for. If None, no token probabilities will be computed.
+            apply_softmax: Whether to normalize probabilities using softmax.
+            item_first: If True, prepend items to query. Otherwise append items to query.
+
+        Returns:
+            List of dictionaries mapping token IDs to their probabilities for each item.
+            Each dictionary in the list corresponds to one item input.
+
+        Raises:
+            ValueError: If query is not provided, or if items is not provided,
+                      or if token IDs are out of vocabulary, or if logprobs are not available for the specified tokens.
+        """
+        loop = asyncio.get_event_loop()
+        return loop.run_until_complete(
+            self.tokenizer_manager.score_request(
+                query=query,
+                items=items,
+                label_token_ids=label_token_ids,
+                apply_softmax=apply_softmax,
+                item_first=item_first,
+                request=None,
+            )
+        )
+
+    async def async_score(
+        self,
+        query: Optional[Union[str, List[int]]] = None,
+        items: Optional[Union[str, List[str], List[List[int]]]] = None,
+        label_token_ids: Optional[List[int]] = None,
+        apply_softmax: bool = False,
+        item_first: bool = False,
+    ) -> List[List[float]]:
+        """
+        Asynchronous version of score method.
+
+        See score() for detailed documentation.
+        """
+        return await self.tokenizer_manager.score_request(
+            query=query,
+            items=items,
+            label_token_ids=label_token_ids,
+            apply_softmax=apply_softmax,
+            item_first=item_first,
+            request=None,
+        )
+
 
 def _set_envs_and_config(server_args: ServerArgs):
     # Set global environments

python/sglang/srt/entrypoints/http_server.py

@@ -82,6 +82,7 @@ from sglang.srt.openai_api.adapter import (
     v1_retrieve_batch,
     v1_retrieve_file,
     v1_retrieve_file_content,
+    v1_score,
 )
 from sglang.srt.openai_api.protocol import ModelCard, ModelList
 from sglang.srt.reasoning_parser import ReasoningParser
@@ -720,6 +721,12 @@ async def vertex_generate(vertex_req: VertexGenerateReqInput, raw_request: Reque
     return ORJSONResponse({"predictions": ret})
 
 
+@app.post("/v1/score")
+async def v1_score_request(raw_request: Request):
+    """Endpoint for the decoder-only scoring API. See Engine.score() for detailed documentation."""
+    return await v1_score(_global_state.tokenizer_manager, raw_request)
+
+
 def _create_error_response(e):
     return ORJSONResponse(
         {"error": {"message": str(e)}}, status_code=HTTPStatus.BAD_REQUEST