sgl-project/sglang · error · TypeError

Unexpected SD2 mid block

Error message

Unexpected SD2 mid block: {type(mid_block).__name__}.

What it means

_replace_transformer_blocks walks the SD2 UNet expecting unet.mid_block to be a UNetMidBlock2DCrossAttn so it can patch its first attention block (layer 'mid_0_0'). If the mid block is any other type, this TypeError is raised before patching.

Solutions

  1. Load the Stable Diffusion 2.x UNet this pipeline expects
  2. Pin the diffusers version listed in the repo requirements
  3. If using a custom UNet, add support for its mid-block type before calling this function

Example fix

# before
_replace_transformer_blocks(sd1_unet, ...)  # mid block type differs

# after
unet = UNet2DConditionModel.from_pretrained("stabilityai/stable-diffusion-2-1", subfolder="unet")
_replace_transformer_blocks(unet, ...)
Defensive patterns

Strategy: type-guard

Validate before calling

from diffusers.models.unets.unet_2d_blocks import UNetMidBlock2DCrossAttn
assert isinstance(unet.mid_block, UNetMidBlock2DCrossAttn), type(unet.mid_block).__name__

Type guard

def is_sd2_mid_block(unet) -> bool:
    from diffusers.models.unets.unet_2d_blocks import UNetMidBlock2DCrossAttn
    return isinstance(unet.mid_block, UNetMidBlock2DCrossAttn)

Prevention

When it happens

Trigger: Calling _replace_transformer_blocks on a UNet whose mid_block is not UNetMidBlock2DCrossAttn — e.g. a UNet with a different mid-block config, a custom architecture, or a non-SD2 checkpoint.

Common situations: Loading a SD1/SDXL or custom UNet into the Hunyuan3D Paint pipeline; diffusers version differences in mid-block classes.

Related errors


AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/32818c411b7d2704. Report an issue: GitHub.

Appendix: source

Thrown at python/sglang/multimodal_gen/runtime/models/dits/hunyuan3d_paint.py:187

                f"Expected BasicTransformerBlock, got {type(transformer).__name__}."
            )
        model.transformer_blocks[0] = Hunyuan3DPaintTransformerBlock(
            transformer,
            layer_name,
            use_multiview_attention=use_multiview_attention,
            use_reference_attention=use_reference_attention,
            is_turbo=is_turbo,
        )

    for block_index, block in enumerate(unet.down_blocks):
        if not isinstance(block, CrossAttnDownBlock2D):
            continue
        for attention_index, attention in enumerate(block.attentions):
            replace(attention, f"down_{block_index}_{attention_index}_0")

    mid_block = unet.mid_block
    if not isinstance(mid_block, UNetMidBlock2DCrossAttn):
        raise TypeError(f"Unexpected SD2 mid block: {type(mid_block).__name__}.")
    replace(mid_block.attentions[0], "mid_0_0")

    for block_index, block in enumerate(unet.up_blocks):
        if not isinstance(block, CrossAttnUpBlock2D):
            continue
        for attention_index, attention in enumerate(block.attentions):
            replace(attention, f"up_{block_index}_{attention_index}_0")


@torch.no_grad()
def compute_voxel_grid_mask(
    position: torch.Tensor, grid_resolution: int = 8
) -> torch.Tensor:
    position = position.half()
    _, _, _, height, width = position.shape
    if height % grid_resolution != 0 or width % grid_resolution != 0:
        raise ValueError(
            f"Position map {height}x{width} is not divisible by {grid_resolution}."

View on GitHub (pinned to 0132848349)