sgl-project/sglang · error · ValueError

text and input_ids cannot be provided at the same time

Error message

text and input_ids cannot be provided at the same time

What it means

A request may be seeded by exactly one of raw text or pre-tokenized input_ids, not both — tokenizing text that also carries input_ids is ambiguous. Raised at the start of normalize_batch_and_arguments.

Source

Thrown at python/sglang/srt/managers/io_struct.py:1186

    def _validate_rid_uniqueness(self):
        """Validate that request IDs within a batch are unique."""
        if isinstance(self.rid, list) and len(set(self.rid)) != len(self.rid):
            counts = Counter(self.rid)
            duplicates = [rid for rid, count in counts.items() if count > 1]
            raise ValueError(
                f"Duplicate request IDs detected within the request: {duplicates}"
            )

    def normalize_batch_and_arguments(self):
        # at least one of text, input_ids, or image should be provided
        if self.text is None and self.input_ids is None and self.image_data is None:
            raise ValueError(
                "At least one of text, input_ids, or image should be provided"
            )

        # text and input_ids cannot be provided at the same time
        if self.text is not None and self.input_ids is not None:
            raise ValueError("text and input_ids cannot be provided at the same time")

        # Derive the batch size
        self.batch_size = 0
        self.is_single = True

        # check the batch size of text
        if self.text is not None:
            if isinstance(self.text, list):
                self.batch_size += len(self.text)
                self.is_single = False
            else:
                self.batch_size += 1

        # check the batch size of input_ids
        if self.input_ids is not None:
            if isinstance(self.input_ids[0], list):
                self.batch_size += len(self.input_ids)
                self.is_single = False

View on GitHub (pinned to 0132848349)

Solutions

1. Drop one field: keep input_ids if you tokenized yourself, otherwise keep text
2. Set text=None explicitly when building from token ids

Example fix

// before
GenerateReqInput(text='hi', input_ids=[1,2,3])
// after
GenerateReqInput(input_ids=[1,2,3])

if input_ids is not None:
    text = None
assert (text is None) or (input_ids is None)

When it happens

Trigger: GenerateReqInput(text='hi', input_ids=[1,2,3]); chat clients that attach cached token ids while still forwarding the raw prompt string.

Common situations: Caching layers that add input_ids for reuse but forget to drop text; template rendering code that populates both fields 'for safety'.

Related errors

• Unknown req_type: {req_type!r} (expected 'generate' or 'embe
• Exactly one of 'prompt' or 'messages' must be provided.
• At least one of text, input_ids, or image should be provided
• token_ids_logprob must be a flat list of integers.
• Unknown serve backend {name!r}. Available values: {available

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/a5a966fb4ab6dcfd. Report an issue: GitHub.