sgl-project/sglang · error · ValueError

Unsupported time_compression_ratio: {time_compression_ratio}

Error message

Unsupported time_compression_ratio: {time_compression_ratio}

What it means

The decoder counterpart of the encoder's temporal-ratio check: only time_compression_ratio of 4 or 8 are supported when computing which up blocks upsample along time. Other values raise this ValueError in HunyuanVideoVAE.__init__.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/vaes/hunyuanvae.py:786

        output_channel = reversed_block_out_channels[0]
        for i, up_block_type in enumerate(up_block_types):
            if up_block_type != "HunyuanVideoUpBlock3D":
                raise ValueError(f"Unsupported up_block_type: {up_block_type}")

            prev_output_channel = output_channel
            output_channel = reversed_block_out_channels[i]
            is_final_block = i == len(block_out_channels) - 1
            num_spatial_upsample_layers = int(np.log2(spatial_compression_ratio))
            num_time_upsample_layers = int(np.log2(time_compression_ratio))

            if time_compression_ratio == 4:
                add_spatial_upsample = bool(i < num_spatial_upsample_layers)
                add_time_upsample = bool(
                    i >= len(block_out_channels) - 1 - num_time_upsample_layers
                    and not is_final_block
                )
            else:
                raise ValueError(
                    f"Unsupported time_compression_ratio: {time_compression_ratio}"
                )

            upsample_scale_factor_HW = (2, 2) if add_spatial_upsample else (1, 1)
            upsample_scale_factor_T = (2,) if add_time_upsample else (1,)
            upsample_scale_factor = tuple(
                upsample_scale_factor_T + upsample_scale_factor_HW
            )

            up_block = HunyuanVideoUpBlock3D(
                num_layers=self.layers_per_block + 1,
                in_channels=prev_output_channel,
                out_channels=output_channel,
                add_upsample=bool(add_spatial_upsample or add_time_upsample),
                upsample_scale_factor=upsample_scale_factor,
                resnet_eps=1e-6,
                resnet_act_fn=act_fn,
                resnet_groups=norm_num_groups,

View on GitHub (pinned to 0132848349)

Solutions

1. Set time_compression_ratio to 4 or 8, matching the checkpoint's actual temporal upsampling factor
2. Keep time_compression_ratio and temporal_compression_ratio consistent within one config
3. Extend the branch chain in the decoder __init__ if a genuinely new ratio is needed

Example fix

// before
"time_compression_ratio": 2

// after
"time_compression_ratio": 4

assert cfg["time_compression_ratio"] in (4, 8), 'time_compression_ratio must be 4 or 8'
assert cfg.get("time_compression_ratio") == cfg.get("temporal_compression_ratio", cfg["time_compression_ratio"]), 'encoder/decoder temporal ratios should match'

def is_supported_time_ratio(r) -> bool:
    return r in (4, 8)

When it happens

Trigger: Constructing the VAE with time_compression_ratio set to a value other than 4 or 8 in the config (note the decoder config key is spelled time_compression_ratio, distinct from the encoder's temporal_compression_ratio).

Common situations: Config conversions where time_compression_ratio/temporal_compression_ratio got renamed or diverged between the two halves of the config; custom checkpoints with unusual temporal strides; inconsistent hand-edited configs.

Related errors

• Unsupported time_compression_ratio: {temporal_compression_ra
• Unsupported down_block_type: {down_block_type}
• Unsupported up_block_type: {up_block_type}
• Unsupported LTX-2.3 encoder block: {block_name}
• Unsupported latent_log_var: {latent_log_var}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/585d0bb634bfddf5. Report an issue: GitHub.