sgl-project/sglang · error · ValueError

vec must be 1D

Error message

vec must be 1D

What it means

_make_pairs_from_vector duplicates a 1D vector (timesteps or sigmas) into two stacked columns for the paired scheduler; a 2D+ tensor cannot be unambiguously paired. It is called internally by _refresh_pair_cache on self.timesteps/self.sigmas.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/schedulers/flow_match_pair.py:446

                    if self.reverse_sigmas:
                        base = 1 - base

                    if source == "timesteps":
                        return base * self.num_train_timesteps
                    return base

                col0 = _build_column(visual_shift, visual_denoising_strength, visual_mu)
                col1 = _build_column(audio_shift, audio_denoising_strength, audio_mu)
                return torch.stack([col0, col1], dim=1)

            _dual_sigma_shift._requires_source = True
            self.set_pair_postprocess(_dual_sigma_shift)
            return
        raise ValueError(f"Unknown pair_postprocess name: {name}")

    def _make_pairs_from_vector(self, vec: torch.Tensor) -> torch.Tensor:
        if vec.ndim != 1:
            raise ValueError("vec must be 1D")
        return torch.stack([vec, vec], dim=1)

    def get_pairs(self, source: str = "timesteps") -> torch.Tensor:
        if source == "timesteps":
            if self.pair_timesteps is None:
                self._refresh_pair_cache()
            return self.pair_timesteps
        if source == "sigmas":
            if self.pair_sigmas is None:
                self._refresh_pair_cache()
            return self.pair_sigmas
        raise ValueError("source must be 'timesteps' or 'sigmas'")

    def timestep_to_sigma(self, timestep: torch.Tensor | float) -> torch.Tensor:
        """Return sigma for a scalar timestep via nearest neighbor lookup.

        Args:
            timestep: Scalar timestep value.

View on GitHub (pinned to 0132848349)

Solutions

1. Ensure scheduler.timesteps and scheduler.sigmas remain 1D tensors
2. Flatten or index the tensor before assigning it to the scheduler
3. If pairing columns is needed, use the pair postprocess mechanism rather than 2D inputs

Example fix

# before
sched.timesteps = timesteps_2d  # (B, T)
# after
sched.timesteps = timesteps_2d.flatten()  # or select sched.timesteps = timesteps_2d[0]

assert sched.timesteps is None or sched.timesteps.ndim == 1
assert sched.sigmas is None or sched.sigmas.ndim == 1

def is_1d(t: torch.Tensor) -> bool:
    return isinstance(t, torch.Tensor) and t.ndim == 1

When it happens

Trigger: Internal: timesteps or sigmas were set to a 2D tensor (e.g. via a custom postprocess or manual assignment) before _refresh_pair_cache runs (triggered by get_pairs or set_pair_postprocess).

Common situations: Assigning a batched schedule tensor directly to scheduler.timesteps; a custom set_timesteps override returning a 2D array.

Related errors

• pairs must be a torch.Tensor of shape [N, 2]
• [pred_noise_to_pred_video] Invalid timestep shape: {timestep
• `dt_bias` must have {HV * K} elements (got {dt_bias.numel()}
• `mixed_qkv` must be 2D (got ndim={mixed_qkv.ndim}).
• num_token_non_padded must be a single-element tensor, got sh

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/7994274b70b83e3e. Report an issue: GitHub.