sgl-project/sglang · error · TypeError

pairs must be a torch.Tensor

Error message

pairs must be a torch.Tensor

What it means

The 'dual_sigma_shift' postprocess fully validates its input before rebuilding both modality columns with the FlowMatchScheduler sigma transform. Its first check is a TypeError raised when pairs is not a torch.Tensor at all (list, tuple, numpy array, float, None).

Source

Thrown at python/sglang/multimodal_gen/runtime/models/schedulers/flow_match_pair.py:367

        if name == "dual_sigma_shift":
            visual_shift = float(kwargs.get("visual_shift", self.shift))
            audio_shift = float(kwargs.get("audio_shift", self.shift))
            visual_denoising_strength = float(
                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")

View on GitHub (pinned to 0132848349)

Solutions

1. Convert to a tensor before use: pairs = torch.as_tensor(pairs, dtype=torch.float32)
2. Prefer scheduler-native set_timesteps so pairs are always torch.Tensor [N, 2]
3. Add an isinstance assertion at the boundary of your custom loop
4. Return early / skip the postprocess when pairs is None instead of passing it through

Example fix

# before
post_pairs = np.linspace(1.0, 0.0, num_steps)  # ndarray

# after
post_pairs = torch.as_tensor(np.linspace(1.0, 0.0, num_steps), dtype=torch.float32)
post_pairs = torch.stack([post_pairs, post_pairs], dim=1)

if not isinstance(pairs, torch.Tensor):
    pairs = torch.as_tensor(pairs, dtype=torch.float32)

def is_pairs_tensor(p) -> TypeGuard[torch.Tensor]:
    return isinstance(p, torch.Tensor) and p.ndim == 2 and p.shape[1] == 2

try:
    out = scheduler.step(...)
except TypeError as e:
    if 'pairs must be a torch.Tensor' in str(e):
        pairs = torch.as_tensor(pairs).reshape(-1, 2)
        out = scheduler.step(...)  # retry with normalized input
    else:
        raise

When it happens

Trigger: set_pair_postprocess_by_name('dual_sigma_shift', ...) with a subsequent call where the pairs argument is a Python list, numpy.ndarray, or scalar — e.g. a schedule produced by numpy.linspace or deserialized from JSON/config.

Common situations: Config-driven pipelines that deserialize schedules from JSON/YAML into lists; numpy-based schedule generation not converted to torch; passing None when the scheduler cache is empty.

Understand the failure class

Background: "Wrong argument type", "must be a string", "expected Array or Prism::Scope": TypeError and ArgumentError when a library receives a value of the wrong type — this error's family across 28 libraries.

Related errors

• pairs length must be greater than 0
• 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/c78ef43238621381. Report an issue: GitHub.