sgl-project/sglang · error · ValueError

SP DMD renoise requires packed video `batch.raw_latent_shape

Error message

SP DMD renoise requires packed video `batch.raw_latent_shape`.

What it means

In JoyEcho's SP (sequence-parallel) DMD renoise path, when video latents are sharded the sampler reconstructs noise on the full (unsharded) layout using batch.raw_latent_shape. It requires a 3-tuple (frames, height/latent, width/latent) packed video shape; anything else means the batch metadata is missing or malformed.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/joy_echo/denoising.py:207

            sigma_t = sigma_t.reshape(-1, *[1] * (original.ndim - 1))
        elif sigma_t.ndim == 2:
            sigma_t = sigma_t.reshape(*sigma_t.shape, *[1] * (original.ndim - 2))
        return (1.0 - sigma_t) * original + sigma_t * noise

    def _sample_sp_consistent_noise(
        self,
        local_reference: torch.Tensor,
        batch: Req,
        server_args: ServerArgs,
        *,
        shard_video: bool,
        shard_audio: bool,
    ) -> torch.Tensor:
        """Sample renoise on the global latent layout, then shard for SP."""
        if shard_video:
            raw_shape = batch.raw_latent_shape
            if not (isinstance(raw_shape, tuple) and len(raw_shape) == 3):
                raise ValueError(
                    "SP DMD renoise requires packed video `batch.raw_latent_shape`."
                )
            full_reference = torch.empty(
                tuple(raw_shape),
                device=local_reference.device,
                dtype=local_reference.dtype,
            )
            full_noise = self._randn_like_with_batch_generators(full_reference, batch)
            sharded_noise, _ = server_args.pipeline_config.shard_latents_for_sp(
                batch, full_noise
            )
            return sharded_noise

        if shard_audio:
            orig_audio_len = batch.sp_audio_orig_num_frames
            if orig_audio_len <= 0:
                raise ValueError(
                    "SP DMD renoise requires `batch.sp_audio_orig_num_frames`."

View on GitHub (pinned to 0132848349)

Solutions

1. Ensure the upstream stage sets batch.raw_latent_shape to the packed 3-tuple (C*T? no: frames, h_lat, w_lat) of the full unsharded video latent
2. Disable sharding (shard_video=False) if running single-GPU
3. Upgrade/align the stage that constructs the batch so SP metadata is populated

Example fix

# before
batch.raw_latent_shape = None  # or tuple(latents.shape)  # 4-D
# after
batch.raw_latent_shape = (num_latent_frames, h_lat, w_lat)  # packed 3-tuple

if shard_video and not (isinstance(getattr(batch, 'raw_latent_shape', None), tuple) and len(batch.raw_latent_shape) == 3):
    raise/shard_video = False  # or populate the field upstream

def has_packed_video_shape(b) -> bool:
    s = getattr(b, 'raw_latent_shape', None)
    return isinstance(s, tuple) and len(s) == 3

When it happens

Trigger: Running DMD denoising with shard_video=True but batch.raw_latent_shape unset, None, a list, or not length 3 (e.g. a 4-D tensor shape was stored instead of the packed 3-D latent shape).

Common situations: Building Req batches manually without the SP metadata; a pipeline change that stopped populating raw_latent_shape when sequence parallelism is enabled; enabling shard_video on a video path that never packed latents.

Related errors

• SP DMD renoise requires `batch.sp_audio_orig_num_frames`.
• MiniMax H3 Ulysses size must be positive.
• MiniMax H3 ring size must be positive.
• latent_window_size must be positive, got {latent_window_size
• v_cache must be provided

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/ba987a31cb18487c. Report an issue: GitHub.