sgl-project/sglang · error · ValueError

`sigmas` and `timesteps` should have the same length as num_

Error message

`sigmas` and `timesteps` should have the same length as num_inference_steps, if `num_inference_steps` is provided

What it means

If num_inference_steps is given explicitly, any provided sigmas or timesteps list must have exactly that length. This catches inconsistency between the declared step count and the custom schedule arrays.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/schedulers/scheduling_flow_match_euler_discrete.py:317

        """

        if self.config.use_dynamic_shifting and mu is None:
            raise ValueError(
                "`mu` must be passed when `use_dynamic_shifting` is set to be `True`"
            )

        if (
            sigmas is not None
            and timesteps is not None
            and len(sigmas) != len(timesteps)
        ):
            raise ValueError("`sigmas` and `timesteps` should have the same length")

        if num_inference_steps is not None:
            if (sigmas is not None and len(sigmas) != num_inference_steps) or (
                timesteps is not None and len(timesteps) != num_inference_steps
            ):
                raise ValueError(
                    "`sigmas` and `timesteps` should have the same length as num_inference_steps, if `num_inference_steps` is provided"
                )
        else:
            if sigmas is not None:
                num_inference_steps = len(sigmas)
            elif timesteps is not None:
                num_inference_steps = len(timesteps)
            else:
                raise ValueError(
                    "Either num_inference_steps, sigmas, or timesteps must be provided"
                )

        self.num_inference_steps = num_inference_steps

        # 1. Prepare default sigmas
        is_timesteps_provided = timesteps is not None

        timesteps_array: np.ndarray | None = None

View on GitHub (pinned to 0132848349)

Solutions

1. Match the list length to num_inference_steps exactly (decide terminal-sigma convention)
2. Omit num_inference_steps and let it be inferred from len(sigmas)/len(timesteps)
3. Regenerate the schedule for the target step count

Example fix

# before
scheduler.set_timesteps(num_inference_steps=28, sigmas=my_sigmas)  # len 30
# after
scheduler.set_timesteps(num_inference_steps=len(my_sigmas), sigmas=my_sigmas)

if num_inference_steps is not None:
    if sigmas is not None: assert len(sigmas) == num_inference_steps
    if timesteps is not None: assert len(timesteps) == num_inference_steps

When it happens

Trigger: set_timesteps(28, sigmas=[...30 values...]) or a timesteps list off by one (terminal sigma inclusion).

Common situations: Interpolating schedules at a different resolution/count and forgetting to update num_inference_steps; distillation presets with off-by-one lists.

Related errors

• `sigmas` and `timesteps` should have the same length
• kv-canary: {name} length must be {expected}, got {actual}
• pairs must be a torch.Tensor of shape [N, 2]
• pairs must be a torch.Tensor
• pairs length must be greater than 0

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/2159d183dba74e32. Report an issue: GitHub.