sgl-project/sglang · error · ValueError

{name}[{index}] has {len(sequence_lengths)} entries; expecte

Error message

{name}[{index}] has {len(sequence_lengths)} entries; expected {self.prompt_batch_size} (per-prompt) or {self.sample_batch_size} (per-sample).

What it means

Raised by ConditionExpander._expand_sequence_lengths when a per-prompt/per-sample list-of-lists field (e.g. sequence lengths) has an inner list whose length matches neither the prompt batch size nor the sample batch size. The expander repeats each per-prompt entry `repeats` times to reach sample granularity; a mismatched length means the conditioning data is inconsistent with the current batch.

Source

Thrown at python/sglang/multimodal_gen/runtime/utils/condition_expansion.py:85

            return None
        repeats = self.sample_batch_size // self.prompt_batch_size
        expanded = []
        for index, sequence_lengths in enumerate(value):
            if (
                sequence_lengths is None
                or len(sequence_lengths) == self.sample_batch_size
            ):
                expanded.append(sequence_lengths)
            elif len(sequence_lengths) == self.prompt_batch_size:
                expanded.append(
                    [
                        sequence_length
                        for sequence_length in sequence_lengths
                        for _ in range(repeats)
                    ]
                )
            else:
                raise ValueError(
                    f"{name}[{index}] has {len(sequence_lengths)} entries; expected "
                    f"{self.prompt_batch_size} (per-prompt) or "
                    f"{self.sample_batch_size} (per-sample)."
                )
        return expanded

    def expand_field(self, batch, field_name: str) -> None:
        """Expand one field in place, dispatching from its value type."""
        value = getattr(batch, field_name)
        if value is None:
            return
        if isinstance(value, torch.Tensor) or (
            isinstance(value, list)
            and all(item is None or isinstance(item, torch.Tensor) for item in value)
        ):
            expanded = self._expand_tensors(value, field_name)
        elif isinstance(value, list) and all(
            item is None or isinstance(item, list) for item in value

View on GitHub (pinned to 0132848349)

Solutions

1. Check that each sequence-length list has exactly prompt_batch_size entries (or exactly sample_batch_size for already-expanded data) before calling expand_field
2. Rebuild the conditioning field after changing samples_per_prompt / batch composition
3. If the data is already per-sample, pass it as sample_batch_size-length lists so the per-sample branch applies

Example fix

# before
batch.cond_seq_lens = [[12]]  # 1 prompt but prompt_batch_size=2
# after
batch.cond_seq_lens = [[12], [12]]  # one entry per prompt

assert all(len(x) in (expander.prompt_batch_size, expander.sample_batch_size) for x in field), 'length mismatch'

When it happens

Trigger: Calling expand_conditioning_to_sample_batch / expand_field with a field that is a list of lists (sequence-length style) where len of the list at some index != prompt_batch_size and != sample_batch_size. Typically happens when cond_* inputs were built for a different number of prompts or when samples_per_prompt changed between construction and expansion.

Common situations: Mixing per-prompt tensors with per-sample sequence-length lists in one batch; changing samples_per_prompt or image/video counts after building the conditioning dict; off-by-one when padding prompts.

Related errors

• QwenImageEditPlus expects either one shared condition image
• {field_name} must be a tensor, list of tensors, list of sequ
• cos/sin shape does not cover image tokens and head_dim
• Unsupported image type: {type(image)}
• QwenImage RoPE text cache overflow before denoising: require

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/81362a1652f96efe. Report an issue: GitHub.