sgl-project/sglang · error · ValueError

condition_frame_indexes={cond_indexes} exceeds the latent fr

Error message

condition_frame_indexes={cond_indexes} exceeds the latent frame count {num_latent_frames} for num_frames={batch.num_frames}

What it means

The stage computes the maximum condition_frame_index and compares it against the number of latent frames derived from batch.num_frames (after VAE temporal compression). Any conditioning frame index >= num_latent_frames is out of range for the latent tensor and is rejected.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/cosmos3.py:561

                pixel_input = batch.preprocessed_video.to(
                    device=device, dtype=vae_dtype
                )
                cond_indexes = Cosmos3ImagePreprocessStage._resolve_condition_indexes(
                    batch
                )
            else:
                pixel_input = batch.preprocessed_image.unsqueeze(2).to(
                    device=device, dtype=vae_dtype
                )
                cond_indexes = [0]

            with self.use_declared_component(component_name="vae", module=self.vae):
                with torch.no_grad():
                    cond_latent = self._vae_encode(pixel_input).to(dtype)

            max_idx = max(cond_indexes)
            if max_idx >= num_latent_frames:
                raise ValueError(
                    f"condition_frame_indexes={cond_indexes} exceeds the "
                    f"latent frame count {num_latent_frames} for "
                    f"num_frames={batch.num_frames}"
                )

            condition_latents = torch.zeros_like(noise)
            condition_mask = torch.zeros(
                batch_dim, 1, num_latent_frames, 1, 1, device=device, dtype=dtype
            )
            for idx in cond_indexes:
                src = min(idx, cond_latent.shape[2] - 1)
                condition_latents[:, :, idx, :, :] = cond_latent[:, :, src, :, :]
                condition_mask[:, :, idx, :, :] = 1.0

            latents = (
                condition_mask * condition_latents + (1.0 - condition_mask) * noise
            )
            batch.extra["condition_latents"] = condition_latents

View on GitHub (pinned to 0132848349)

Solutions

1. Reduce condition_frame_indexes so max index < num_latent_frames (typically (num_frames-1)//temporal_stride)
2. Increase num_frames so the clip yields enough latent frames to cover the requested condition indexes
3. Compute latent indexes from frame indexes with the VAE temporal compression factor before passing them

Example fix

# before
sp.condition_frame_indexes = [30]  # num_frames=57 -> ~8 latent frames

# after
sp.condition_frame_indexes = [0, 7]  # within latent frame count

num_latent = (batch.num_frames - 1) // temporal_stride + 1
assert max(cond_indexes) < num_latent

def valid_cond_indexes(idxs: list[int], num_latent_frames: int) -> bool:
    return bool(idxs) and max(idxs) < num_latent_frames

When it happens

Trigger: Setting condition_frame_indexes in sampling params to a value >= num_latent_frames, e.g. passing frame index 30 for a short clip where the VAE compresses num_frames down to only a few latent frames.

Common situations: Reusing condition_frame_indexes tuned for long clips on short num_frames requests; forgetting the (num_frames - 1) // temporal_compression reduction when mapping frame indices to latent indices.

Related errors

• Cosmos3 action input accepts one image field; use a list or
• Cosmos3 action prompt must be a string or non-empty list
• Cosmos3 batched prompts must tokenize to the same length bec
• sound generation was requested (sound_duration > 0) but the
• action_mode is set but the loaded Cosmos3 checkpoint has no

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/e0c6dd8efedc1afc. Report an issue: GitHub.