sgl-project/sglang · error · ValueError

Unsupported time_compression_ratio: {temporal_compression_ra

Error message

Unsupported time_compression_ratio: {temporal_compression_ratio}

What it means

The HunyuanVideoVAE encoder supports only temporal_compression_ratio values of 4 (implicit in the earlier branch) and 8 when deciding which down blocks add time downsampling. Any other ratio raises this ValueError during __init__.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/vaes/hunyuanvae.py:656

                raise ValueError(f"Unsupported down_block_type: {down_block_type}")

            input_channel = output_channel
            output_channel = block_out_channels[i]
            is_final_block = i == len(block_out_channels) - 1
            num_spatial_downsample_layers = int(np.log2(spatial_compression_ratio))
            num_time_downsample_layers = int(np.log2(temporal_compression_ratio))

            if temporal_compression_ratio == 4:
                add_spatial_downsample = bool(i < num_spatial_downsample_layers)
                add_time_downsample = bool(
                    i >= (len(block_out_channels) - 1 - num_time_downsample_layers)
                    and not is_final_block
                )
            elif temporal_compression_ratio == 8:
                add_spatial_downsample = bool(i < num_spatial_downsample_layers)
                add_time_downsample = bool(i < num_time_downsample_layers)
            else:
                raise ValueError(
                    f"Unsupported time_compression_ratio: {temporal_compression_ratio}"
                )

            downsample_stride_HW = (2, 2) if add_spatial_downsample else (1, 1)
            downsample_stride_T = (2,) if add_time_downsample else (1,)
            downsample_stride = tuple(downsample_stride_T + downsample_stride_HW)

            down_block = HunyuanVideoDownBlock3D(
                num_layers=layers_per_block,
                in_channels=input_channel,
                out_channels=output_channel,
                add_downsample=bool(add_spatial_downsample or add_time_downsample),
                resnet_eps=1e-6,
                resnet_act_fn=act_fn,
                resnet_groups=norm_num_groups,
                downsample_stride=downsample_stride,
                downsample_padding=0,
            )

View on GitHub (pinned to 0132848349)

Solutions

1. Set temporal_compression_ratio to 4 or 8 in the VAE config to match the checkpoint
2. Check the checkpoint's true temporal stride (compare latent length vs video frame count) and use that value, which must be 4 or 8
3. If you need another ratio, extend the branch chain in HunyuanVideoVAE.__init__ with the new add_time_downsample rule

Example fix

// before
"temporal_compression_ratio": 2

// after
"temporal_compression_ratio": 4

assert cfg["temporal_compression_ratio"] in (4, 8), 'temporal_compression_ratio must be 4 or 8'

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

When it happens

Trigger: Loading/constructing the VAE with temporal_compression_ratio set to anything other than 4 or 8 — e.g. 2, 16, or a float like 4.0 that fails the ==8 comparison path — in the config.

Common situations: Custom VAE checkpoints with different temporal stride (e.g. image-only or aggressively compressed variants); hand-edited configs; mismatch between a checkpoint's actual temporal stride and the config value copied from another model.

Related errors

• Unsupported time_compression_ratio: {time_compression_ratio}
• 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/5d7ce96ece3b96ac. Report an issue: GitHub.