sgl-project/sglang · error · ValueError

pairs length must be greater than 0

Error message

pairs length must be greater than 0

What it means

_dual_sigma_shift requires at least one pair row: pairs.shape[0] == 0 raises ValueError because num_steps = 0 would produce empty/degenerate linspace columns and divide-by-zero in the sigma transform. The check runs after the type/shape validations.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/schedulers/flow_match_pair.py:371

                kwargs.get("visual_denoising_strength", 1.0)
            )
            audio_denoising_strength = float(
                kwargs.get("audio_denoising_strength", 1.0)
            )
            visual_mu = kwargs.get(
                "visual_exponential_shift_mu", self.exponential_shift_mu
            )
            audio_mu = kwargs.get(
                "audio_exponential_shift_mu", self.exponential_shift_mu
            )

            def _dual_sigma_shift(pairs: torch.Tensor, *, source: str):
                if not isinstance(pairs, torch.Tensor):
                    raise TypeError("pairs must be a torch.Tensor")
                if pairs.ndim != 2 or pairs.shape[1] != 2:
                    raise ValueError("pairs must be a torch.Tensor of shape [N, 2]")
                if pairs.shape[0] == 0:
                    raise ValueError("pairs length must be greater than 0")
                if source not in ("timesteps", "sigmas"):
                    raise ValueError("source must be 'timesteps' or 'sigmas'")

                num_steps = pairs.shape[0]
                device = pairs.device
                dtype = pairs.dtype

                def _build_column(
                    shift_value: float, denoising_strength: float, mu_override
                ):
                    if shift_value <= 0:
                        raise ValueError("shift must be positive")
                    if denoising_strength <= 0:
                        raise ValueError("denoising_strength must be positive")

                    sigma_start = (
                        self.sigma_min
                        + (self.sigma_max - self.sigma_min) * denoising_strength

View on GitHub (pinned to 0132848349)

Solutions

1. Guard num_inference_steps >= 1 before calling set_timesteps / the scheduler
2. Check pairs.shape[0] > 0 in your loop and skip/return early for empty requests
3. Fix the config default that yields 0 steps (e.g. missing YAML key defaulting to 0)
4. Log the offending shape at the request boundary to catch upstream slicing bugs

Example fix

# before
steps = int(req.get("num_steps", 0))
scheduler.set_timesteps(steps)

# after
steps = int(req.get("num_steps", 50))
if steps < 1:
    raise ValueError("num_steps must be >= 1")
scheduler.set_timesteps(steps)

if not isinstance(pairs, torch.Tensor) or pairs.shape[0] == 0:
    raise ValueError("num_inference_steps must produce at least one pair")
# or: skip the request
if pairs.shape[0] == 0:
    return early_response(request)

try:
    scheduler.set_timesteps(n)
except ValueError as e:
    if 'length must be greater than 0' in str(e):
        n = max(1, n)  # or reject the request
    else:
        raise

When it happens

Trigger: set_pair_postprocess_by_name('dual_sigma_shift') then calling with an empty [0, 2] tensor — e.g. num_inference_steps=0 passed to set_timesteps, or an upstream filter/slice that removed all steps.

Common situations: num_inference_steps read from request config as 0 (missing field, bad default); slicing pairs with a boolean mask that selected nothing; edge case in batched serving where a request has zero scheduled steps.

Related errors

• pairs must be a torch.Tensor
• pairs must be a torch.Tensor of shape [N, 2]
• `dt_bias` must have {HV * K} elements (got {dt_bias.numel()}
• The batch size is expected to be 1 rather than {q.shape[0]}
• The number of initial states is expected to be equal to the

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/b03cefcda35a396b. Report an issue: GitHub.