sgl-project/sglang · error · ValueError

Unsupported topk_mode {topk_mode}

Error message

Unsupported topk_mode {topk_mode}

What it means

FlashVDMVolumeDecoding selects a cross-attention processor strategy based on topk_mode: 'mean' uses the standard FlashVDM processor, anything other than 'mean'/'merge' has no implementation, so __init__ rejects it.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/vaes/hunyuan3d_vae.py:806

                batch_queries = repeat(queries, "p c -> b p c", b=batch_size)
                logits = geo_decoder(
                    queries=batch_queries.to(latents.dtype), latents=latents
                )
                batch_logits.append(logits)
            grid_logits = torch.cat(batch_logits, dim=1)
            next_logits[nidx] = grid_logits[0, ..., 0]
            grid_logits = next_logits.unsqueeze(0)
        grid_logits[grid_logits == -10000.0] = float("nan")

        return grid_logits


class FlashVDMVolumeDecoding:
    """Flash VDM volume decoder with adaptive KV selection."""

    def __init__(self, topk_mode="mean"):
        if topk_mode not in ["mean", "merge"]:
            raise ValueError(f"Unsupported topk_mode {topk_mode}")

        if topk_mode == "mean":
            self.processor = FlashVDMCrossAttentionProcessor()
        else:
            self.processor = FlashVDMTopMCrossAttentionProcessor()

    @torch.no_grad()
    def __call__(
        self,
        latents: torch.FloatTensor,
        geo_decoder: CrossAttentionDecoder,
        bounds: Union[Tuple[float], List[float], float] = 1.01,
        num_chunks: int = 10000,
        mc_level: float = 0.0,
        octree_resolution: int = None,
        min_resolution: int = 63,
        mini_grid_num: int = 4,
        enable_pbar: bool = True,

View on GitHub (pinned to 0132848349)

Solutions

1. Use topk_mode='mean' (FlashVDMCrossAttentionProcessor) or 'merge' (FlashVDMTopMCrossAttentionProcessor)
2. Strip/normalize the config string (whitespace, casing) before passing it
3. Upgrade if you need an additional mode added in a newer version

Example fix

# before
dec = FlashVDMVolumeDecoding(topk_mode="max")
# after
dec = FlashVDMVolumeDecoding(topk_mode="mean")

mode = str(topk_mode).strip().lower()
if mode not in ("mean", "merge"):
    mode = "mean"
dec = FlashVDMVolumeDecoding(topk_mode=mode)

def is_valid_topk_mode(m: str) -> bool:
    return str(m).strip().lower() in ("mean", "merge")

When it happens

Trigger: Constructing FlashVDMVolumeDecoding(topk_mode='max'), 'topk', 'sum', or any string other than 'mean'/'merge'; commonly from a config key typed incorrectly or from a newer spec introducing an unimplemented mode.

Common situations: YAML/config typo (e.g. 'mean '** with trailing space, or wrong case 'Mean'); porting settings from another FlashVDM implementation that supports more modes; version lag where the mode exists upstream but not here.

Related errors

• Unsupported mc_algo {mc_algo}, available: {list(SurfaceExtra
• Invalid threshold_type for topk: {threshold_type}. Choose 'q
• Invalid threshold_type: {threshold_type}. Choose 'query_head
• Unsupported scale {scale}. Choose from {list(mapping.keys())
• Either spatial_upsample or temporal_upsample must be True

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/9ff606cd71bf7649. Report an issue: GitHub.