Comfy-Org/ComfyUI · critical · ValueError

Unknown block type: {self.block_type}

Error message

Unknown block type: {self.block_type}

What it means

DITBuildingBlock.forward dispatches on self.block_type with the same vocabulary as the constructor (self_attn/sa, cross_attn/ca, and the else covers mlp/ff plus anything else after the first two branches). If block_type was mutated after construction or the forward dispatch was extended in a fork without extending the constructor (or vice versa), forward raises 'Unknown block type'. Under normal flow the constructor check at blocks.py:647 fires first, so a runtime hit here indicates attribute mutation or inconsistent custom edits.

Source

Thrown at comfy/ldm/cosmos/blocks.py:720

                adaln_norm_state(self.norm_state, x, scale_1_1_1_B_D, shift_1_1_1_B_D),
            )
        elif self.block_type in ["full_attn", "fa"]:
            x = x + gate_1_1_1_B_D * self.block(
                adaln_norm_state(self.norm_state, x, scale_1_1_1_B_D, shift_1_1_1_B_D),
                context=None,
                rope_emb_L_1_1_D=rope_emb_L_1_1_D,
                transformer_options=transformer_options,
            )
        elif self.block_type in ["cross_attn", "ca"]:
            x = x + gate_1_1_1_B_D * self.block(
                adaln_norm_state(self.norm_state, x, scale_1_1_1_B_D, shift_1_1_1_B_D),
                context=crossattn_emb,
                crossattn_mask=crossattn_mask,
                rope_emb_L_1_1_D=rope_emb_L_1_1_D,
                transformer_options=transformer_options,
            )
        else:
            raise ValueError(f"Unknown block type: {self.block_type}")

        return x


class GeneralDITTransformerBlock(nn.Module):
    """
    A wrapper module that manages a sequence of DITBuildingBlocks to form a complete transformer layer.
    Each block in the sequence is specified by a block configuration string.

    Parameters:
        x_dim (int): Dimension of input features
        context_dim (int): Dimension of context features for cross-attention blocks
        num_heads (int): Number of attention heads
        block_config (str): String specifying block sequence (e.g. "ca-fa-mlp" for cross-attention,
                          full-attention, then MLP)
        mlp_ratio (float): MLP hidden dimension multiplier. Default: 4.0
        x_format (str): Input tensor format. Default: "BTHWD"
        use_adaln_lora (bool): Whether to use AdaLN-LoRA. Default: False

View on GitHub (pinned to 1c6d8d45b3)

Solutions

1. Do not reassign block_type after construction; rebuild the block with a supported type
2. If extending the block vocabulary in a fork, update both __init__ and forward branches together
3. Recreate the model from its config rather than mutating loaded modules

Example fix

# before
block.block_type = "window_attn"  # forward raises

# after
block = DITBuildingBlock("self_attn", x_dim, num_heads, operations=operations)

assert block.block_type in {"self_attn", "sa", "cross_attn", "ca", "full_attn", "fa", "mlp", "ff"}

def is_supported_block_type(block) -> bool:
    return block.block_type in {"self_attn", "sa", "cross_attn", "ca", "full_attn", "fa", "mlp", "ff"}

When it happens

Trigger: Setting block.block_type after construction (e.g. converting a cross_attn block to another type at runtime); running a partially-updated fork where __init__ accepts a type forward() does not handle.

Common situations: Custom nodes patching Cosmos block types for inference tricks; reusing serialized module configs with new block names against older forward code.

Related errors

• Unknown block type: {block_type}
• Scalar feature is not implemented yet.
• Unknown x_format {self.blocks[0].x_format}
• Input img and txt tensors must have 3 dimensions.
• Input img tensor must be in [B, C, H, W] format.

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/08173b8a31f0cf1b. Report an issue: GitHub.