Comfy-Org/ComfyUI · critical · ValueError

Missing W4A8 group scale (weight_s_rel) for layer {layer_nam

Error message

Missing W4A8 group scale (weight_s_rel) for layer {layer_name}

What it means

Raised in comfy/ops.py for quant_format 'asym_w4a8_int8': an asymmetric int4-weight/int8-activation layout that requires a per-group fp8 scale stored under `weight_s_rel`. If that key is missing from the state dict the loader cannot reconstruct the weights and raises this error naming the layer.

Source

Thrown at comfy/ops.py:1209

            if scale is None:
                raise ValueError(f"Missing ConvRot W4A4 weight scale for layer {layer_name}")
            params_conf = layer_conf.get("params", {})
            if not isinstance(params_conf, dict):
                params_conf = {}
            scales = {
                "scale": scale,
                "convrot_groupsize": int(
                    layer_conf.get("convrot_groupsize", params_conf.get("convrot_groupsize", 256))
                ),
                "quant_group_size": 64,
                "linear_dtype": layer_conf.get("linear_dtype", params_conf.get("linear_dtype", "int4")),
            }
        elif module.quant_format == "asym_w4a8_int8":
            # int4 weight (packed int8 [N,K/2]) + fp8 per-group scale (weight_s_rel),
            # fp32 per-channel scale (weight_s_channel) + optional Lloyd-Max codebook.
            scale = pop_scale("weight_s_rel")
            if scale is None:
                raise ValueError(f"Missing W4A8 group scale (weight_s_rel) for layer {layer_name}")
            if scale.dtype == torch.uint8:
                scale = scale.view(torch.float8_e4m3fn)
            params_conf = layer_conf.get("params", {})
            if not isinstance(params_conf, dict):
                params_conf = {}
            scales = {
                "scale": scale,
                "s_channel": pop_scale("weight_s_channel"),
                "codebook": pop_scale("weight_codebook"),
                "group_size": int(layer_conf.get("group_size", params_conf.get("group_size", 16))),
                "convrot_groupsize": int(
                    layer_conf.get("convrot_groupsize", params_conf.get("convrot_groupsize", 256))
                ),
            }
        else:
            raise ValueError(f"Unsupported quantization format: {module.quant_format}")

        params = layout_cls.Params(**scales, orig_dtype=compute_dtype, orig_shape=module._orig_shape)

View on GitHub (pinned to 1c6d8d45b3)

Solutions

1. Verify the state dict contains `<prefix>.weight_s_rel` (uint8/fp8 view) for the layer named in the error.
2. Re-quantize the model with ComfyUI's asym_w4a8_int8 path so the group scale, optional `weight_s_channel` and `weight_codebook` are saved with the right names.
3. If the file came from elsewhere, rename the equivalent group-scale key to `weight_s_rel` (and view as float8_e4m3fn if stored as uint8).
4. Re-download the model in case scale shards are missing.

Example fix

# before: wrong scale key name
sd['layers.0.self_attn.q_proj.group_scale'] = group_scale

# after: expected key name
sd['layers.0.self_attn.q_proj.weight_s_rel'] = group_scale

prefix = failing_layer_prefix
assert (prefix + 'weight_s_rel') in sd, f'{prefix}weight_s_rel missing for asym_w4a8_int8 layer'

try:
    model_patcher = load_diffusion_model(path)
except ValueError as e:
    if 'weight_s_rel' in str(e):
        raise SystemExit('asym_w4a8_int8 checkpoint missing group scales; re-quantize or rename keys')
    raise

When it happens

Trigger: Loading an asym_w4a8_int8 checkpoint whose `weight_s_rel` group-scale tensor is missing; conversion tools that emit `weight_scale` instead of the expected `weight_s_rel` name; prefix mismatch between comfy_quant and the scale key.

Common situations: Quantized checkpoints produced by external tooling with different scale naming; partial downloads; older quantized files predating the weight_s_rel naming; manual state-dict surgery.

Related errors

• Missing NVFP4 scales for layer {layer_name}
• Missing INT8 weight scale for layer {layer_name}
• Missing ConvRot W4A4 weight scale for layer {layer_name}
• Unsupported quantization format: {module.quant_format}
• Unsupported dtype

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/e45a5621c125b45a. Report an issue: GitHub.