Comfy-Org/ComfyUI · critical · Exception

Block type {block_type} not supported

Error message

Block type {block_type} not supported

What it means

Raised while building the Stage-C UNet of Stable Cascade: the per-level block string in the model config contains a letter other than 'C' (ResBlock), 'A' (AttnBlock), 'F' (FeedForwardBlock) or 'T' (TimestepBlock). The get_block factory inside StageC.__init__ walks every character of each level's block spec and hits the final else-branch. It means the loaded config does not match the architecture this file implements.

Source

Thrown at comfy/ldm/cascade/stage_c.py:78

        self.clip_norm = operations.LayerNorm(c_cond, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)

        self.embedding = nn.Sequential(
            nn.PixelUnshuffle(patch_size),
            operations.Conv2d(c_in * (patch_size ** 2), c_hidden[0], kernel_size=1, dtype=dtype, device=device),
            LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6)
        )

        def get_block(block_type, c_hidden, nhead, c_skip=0, dropout=0, self_attn=True):
            if block_type == 'C':
                return ResBlock(c_hidden, c_skip, kernel_size=kernel_size, dropout=dropout, dtype=dtype, device=device, operations=operations)
            elif block_type == 'A':
                return AttnBlock(c_hidden, c_cond, nhead, self_attn=self_attn, dropout=dropout, dtype=dtype, device=device, operations=operations)
            elif block_type == 'F':
                return FeedForwardBlock(c_hidden, dropout=dropout, dtype=dtype, device=device, operations=operations)
            elif block_type == 'T':
                return TimestepBlock(c_hidden, c_r, conds=t_conds, dtype=dtype, device=device, operations=operations)
            else:
                raise Exception(f'Block type {block_type} not supported')

        # BLOCKS
        # -- down blocks
        self.down_blocks = nn.ModuleList()
        self.down_downscalers = nn.ModuleList()
        self.down_repeat_mappers = nn.ModuleList()
        for i in range(len(c_hidden)):
            if i > 0:
                self.down_downscalers.append(nn.Sequential(
                    LayerNorm2d_op(operations)(c_hidden[i - 1], elementwise_affine=False, eps=1e-6),
                    UpDownBlock2d(c_hidden[i - 1], c_hidden[i], mode='down', enabled=switch_level[i - 1], dtype=dtype, device=device, operations=operations)
                ))
            else:
                self.down_downscalers.append(nn.Identity())
            down_block = nn.ModuleList()
            for _ in range(blocks[0][i]):
                for block_type in level_config[i]:
                    block = get_block(block_type, c_hidden[i], nhead[i], dropout=dropout[i], self_attn=self_attn[i])

View on GitHub (pinned to 1c6d8d45b3)

Solutions

1. Check the level_config strings passed into StageC and correct any invalid characters to the supported set C/A/F/T (uppercase only)
2. If the config came from a checkpoint, verify the checkpoint is a standard Stable Cascade Stage C model and not a newer architecture variant
3. Update ComfyUI if the block letter comes from a newer upstream cascade variant that this version does not know

Example fix

# before
level_config = ["cva", "CA", "CA"]  # raises: Block type c not supported

# after
level_config = [["CVA"], ["CA"], ["CA"]][0]  # use uppercase C/A/F/T letters

VALID = {"C", "A", "F", "T"}
def validate_cascade_config(level_config):
    for level in level_config:
        bad = set(level) - VALID
        if bad:
            raise ValueError(f"Invalid StageC block letters {bad}; allowed: {VALID}")

def is_valid_stage_c_config(level_config) -> bool:
    return all(ch in {"C", "A", "F", "T"} for level in level_config for ch in level)

When it happens

Trigger: Instantiating comfy.ldm.cascade.stage_c.StageC (usually indirectly via StageC_Coder / cascade checkpoint loading) with kwargs whose 'level_config' entries (e.g. ["CVA","CA","CA"]) contain an unexpected character, lowercase letters, or a typo like 'c' instead of 'C'.

Common situations: Hand-editing a Stable Cascade config dict when creating a custom StageC model; loading a modified/custom cascade checkpoint whose embedded config uses new block types not supported by this ComfyUI version; passing a config from a newer upstream version of the cascade code.

Related errors

• Hidden size {params.hidden_size} must be divisible by num_he
• Got {params.axes_dim} but expected positional dim {pe_dim}
• Hidden size {params.hidden_size} must be divisible by num_he
• Unsupported nerf_final_head_type {params.nerf_final_head_typ
• provide num_res_blocks either as an int (globally constant)

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/3fc7d16e75d40fb0. Report an issue: GitHub.