sgl-project/sglang · error · ValueError

dynamic_batch_seeds must be a list with one seed per prompt

Error message

dynamic_batch_seeds must be a list with one seed per prompt

What it means

_generate_seeds validates the optional dynamic_batch_seeds argument: when provided it must be a Python list whose length exactly equals the number of prompts in the request. Anything else (scalar, tuple, numpy array, wrong-length list) is rejected because per-prompt seed streams must map one-to-one onto prompts.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/input_validation.py:89

        """Generate deterministic per-output seeds.

        Batched requests pass one base seed per prompt through `extra`; each
        prompt expands to `num_outputs_per_prompt` consecutive seeds.
        """
        seed = batch.seed
        num_videos_per_prompt = batch.num_outputs_per_prompt

        assert seed is not None

        prompt_count = len(batch.prompt) if isinstance(batch.prompt, list) else 1
        dynamic_batch_seeds = batch.extra.get("dynamic_batch_seeds")

        if dynamic_batch_seeds is not None:
            if (
                not isinstance(dynamic_batch_seeds, list)
                or len(dynamic_batch_seeds) != prompt_count
            ):
                raise ValueError(
                    "dynamic_batch_seeds must be a list with one seed per prompt"
                )
            base_seeds = [int(item) for item in dynamic_batch_seeds]
            seeds = []
            for base_seed in base_seeds:
                seeds.extend([base_seed + i for i in range(num_videos_per_prompt)])
        elif isinstance(seed, list):
            if len(seed) != num_videos_per_prompt:
                raise ValueError(
                    f"seed list length must match num_outputs_per_prompt "
                    f"({num_videos_per_prompt}), got {len(seed)}"
                )
            seeds = [int(item) for item in seed]
        else:
            # Keep per-prompt seed streams deterministic and non-overlapping.
            base_seeds = [
                int(seed) + i * num_videos_per_prompt for i in range(prompt_count)
            ]

View on GitHub (pinned to 0132848349)

Solutions

1. Pass a plain Python list of ints with exactly len(prompts) entries, e.g. dynamic_batch_seeds=[42, 43] for two prompts
2. If you only have one seed, omit dynamic_batch_seeds and use the regular seed argument
3. Convert tensors/arrays via [int(s) for s in seeds] before calling

Example fix

// before
result = pipe(prompts=["a", "b"], dynamic_batch_seeds=[42])

// after
result = pipe(prompts=["a", "b"], dynamic_batch_seeds=[42, 43])

if dynamic_batch_seeds is not None:
    assert isinstance(dynamic_batch_seeds, list) and len(dynamic_batch_seeds) == len(prompts), \
        "dynamic_batch_seeds must have one entry per prompt"

def valid_dynamic_seeds(s, prompt_count: int) -> bool:
    return isinstance(s, list) and len(s) == prompt_count

When it happens

Trigger: Passing dynamic_batch_seeds as an int, a single-element list for a multi-prompt request, a tensor/array instead of a list, or a list built for a different prompt count than the current call.

Common situations: Porting code that used a single seed; batching multiple prompts but reusing a one-element seed list; passing numpy arrays or tensors from a data loader instead of plain lists; off-by-one when concatenating prompt batches.

Related errors

• seed list length must match num_outputs_per_prompt ({num_vid
• unsupported input for causal Conv3D cat/pad CUDA
• unsupported input for usp_merge_heads CUDA
• unsupported input for modulate_scale_shift CUDA
• unsupported input for LTX2 QKNorm split-RoPE CUDA

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/ca09dde3a86425ee. Report an issue: GitHub.