invoke-ai/InvokeAI · error · ValueError

Unsupported lora format: {state_dict.keys()}

Error message

Unsupported lora format: {state_dict.keys()}

What it means

any_lora_layer_from_state_dict sniffs a state dict's keys to decide which LoRA layer class to build (LoRALayer, IA3Layer, NormLayer, FullLayer). If none of the known key signatures (e.g. 'lora_down.weight'/'lora_up.weight', 'on_input', 'w_norm') are present, it raises ValueError naming the actual keys found. This means the per-layer state dict is not in any LoRA format InvokeAI recognizes.

Source

Thrown at invokeai/backend/patches/layers/utils.py:35

    # https://github.com/KohakuBlueleaf/LyCORIS/tree/8ad8000efb79e2b879054da8c9356e6143591bad/lycoris/modules
    if "dora_scale" in state_dict:
        return DoRALayer.from_state_dict_values(state_dict)
    elif "lora_up.weight" in state_dict:
        # LoRA a.k.a LoCon
        return LoRALayer.from_state_dict_values(state_dict)
    elif "hada_w1_a" in state_dict:
        return LoHALayer.from_state_dict_values(state_dict)
    elif "lokr_w1" in state_dict or "lokr_w1_a" in state_dict:
        return LoKRLayer.from_state_dict_values(state_dict)
    elif "diff" in state_dict:
        # Full a.k.a Diff
        return FullLayer.from_state_dict_values(state_dict)
    elif "on_input" in state_dict:
        return IA3Layer.from_state_dict_values(state_dict)
    elif "w_norm" in state_dict:
        return NormLayer.from_state_dict_values(state_dict)
    else:
        raise ValueError(f"Unsupported lora format: {state_dict.keys()}")

View on GitHub (pinned to 0b6a024f2f)

Solutions

1. Inspect the printed keys in the message and compare them to the recognized signatures in invokeai/backend/patches/layers/utils.py; rename or convert keys to a supported format.
2. Use the correct converter function for the LoRA's source format (flux_control, flux_diffusers, flux_kohya, flux_onetrainer, flux_xlabs) instead of feeding raw keys.
3. Re-export the LoRA from the training tool in a standard (diffusers/PEFT or Kohya) format.

Example fix

// before
layer = any_lora_layer_from_state_dict({"down.weight": w, "up.weight": u})  # ValueError
// after
layer = any_lora_layer_from_state_dict({"lora_down.weight": w, "lora_up.weight": u})

KNOWN = {"lora_down.weight", "lora_A.weight", "on_input", "w_norm"}
if not (set(layer_sd) & KNOWN):
    raise ValueError(f"pre-check: layer not a supported LoRA layer: {sorted(layer_sd)}")

def is_supported_lora_layer(sd: dict) -> bool:
    keys = set(sd)
    return bool(keys & {"lora_down.weight", "lora_A.weight", "on_input", "w_norm"})

try:
    layer = any_lora_layer_from_state_dict(sd)
except ValueError as e:
    logger.error("Unsupported LoRA layer keys: %s", e)
    layer = None

When it happens

Trigger: Calling any_lora_layer_from_state_dict (directly or via lora_model_from_flux_* converters or _make_layer_patch) with a per-layer dict whose keys match no known pattern - e.g. missing 'lora_down.weight'/'lora_A.weight' prefixes, or an arbitrary dict passed in.

Common situations: Loading a LoRA trained/saved in an unsupported format or by a tool InvokeAI doesn't convert (wrong exporter, PEFT vs Kohya key naming mismatch), or hand-assembled state dicts with typos in key names.

Related errors

• LoRA model is in unsupported FLUX format
• Unknown lora: {lora_key}!
• Unknown lora: {lora_key}!
• LoRA '{lora_key}' is for {stored_config.base.value if stored
• LoRA '{lora_key}' has conflicting weights on the transformer

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/23cfac7671c48e35. Report an issue: GitHub.