sgl-project/sglang · error · ValueError

Invalid prompts type for score_prompts.

Error message

Invalid prompts type for score_prompts.

What it means

score_prompts only accepts prompts as a string, a list of strings, or a list of token-id lists (and similar list forms); after trying all supported shapes it falls through to this ValueError. Any other type (dict, tuple, nested irregular structure, None) is rejected.

Source

Thrown at python/sglang/srt/managers/tokenizer_manager_score_mixin.py:66

                items=prompts,  # type: ignore[arg-type]
                label_token_ids=label_token_ids,
                apply_softmax=apply_softmax,
                item_first=False,
                request=request,
            )

        # Tokenized prompts
        if isinstance(prompts, list) and (not prompts or isinstance(prompts[0], list)):
            return await self.score_request(
                query=[],
                items=prompts,
                label_token_ids=label_token_ids,
                apply_softmax=apply_softmax,
                item_first=False,
                request=request,
            )

        raise ValueError("Invalid prompts type for score_prompts.")

    def _build_multi_item_token_sequence(
        self, query: List[int], items: List[List[int]], delimiter_token_id: int
    ) -> Tuple[List[int], List[int]]:
        """
        Build a single token sequence for multi-item scoring.
        Format: query<delimiter>item1<delimiter>item2<delimiter>item3<delimiter>
        """
        combined_sequence = query[:]  # Start with query
        delimiter_indices = []

        for item in items:
            delimiter_indices.append(len(combined_sequence))
            combined_sequence.append(delimiter_token_id)  # Add delimiter
            combined_sequence.extend(item)  # Add item tokens

        # Add final delimiter after the last item for logprob extraction
        delimiter_indices.append(len(combined_sequence))

View on GitHub (pinned to 0132848349)

Solutions

1. Convert prompts to str, List[str], or List[List[int]] before calling score_prompts
2. If using numpy arrays, call .tolist() first
3. Guard the call with an isinstance check on the input shape

Example fix

# before
scores = engine.score_prompts(prompts=np.array(["a", "b"]))

# after
scores = engine.score_prompts(prompts=["a", "b"])
# or token ids
scores = engine.score_prompts(prompts=[[1,2,3]])

assert isinstance(prompts, (str, list)), type(prompts)
if isinstance(prompts, list):
    assert all(isinstance(p, (str, list)) for p in prompts)

def valid_score_prompts(p) -> bool:
    if isinstance(p, str): return True
    if isinstance(p, list):
        return all(isinstance(x, str) or (isinstance(x, list) and all(isinstance(t, int) for t in x)) for x in p)
    return False

try:
    scores = engine.score_prompts(prompts=prompts)
except ValueError as e:
    if "Invalid prompts type" in str(e):
        prompts = prompts.tolist() if hasattr(prompts, "tolist") else list(prompts)
        scores = engine.score_prompts(prompts=prompts)
    else:
        raise

When it happens

Trigger: Calling score_prompts(prompts=...) with a non-supported type such as a dict, tuple, numpy array, or None; or a heterogeneous list whose elements are neither strings nor int-lists.

Common situations: Passing tokenizer output objects (BatchEncoding) or numpy arrays directly instead of plain lists; refactoring code that previously called a different scoring API with different input shapes; None defaults leaking through.

Related errors

• Invalid combination of query/items types for score_request.
• num_token_non_padded must be a torch.Tensor
• actions must be a list[list[str]]
• rollout_noise_level must be a number, got {noise!r}
• {field_name} must be a JSON object

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/b8e90ca406d8d9f7. Report an issue: GitHub.