hiyouga/LlamaFactory · error · RuntimeError

Failed to export model: weight conversion reversal is not su

Error message

Failed to export model: weight conversion reversal is not supported for this model architecture (NotImplementedError in transformers.core_model_loading.reverse_op). This is a known issue with transformers>=5.0 for certain model types (e.g. Mistral/Ministral). Workarounds: (1) use transformers<5.0, or (2) report the issue to the transformers repository.

What it means

export_model wraps `model.save_pretrained` (src/llamafactory/train/tuner.py:222). On transformers>=5.0, saving certain architectures (e.g. Mistral/Ministral) can hit NotImplementedError in `transformers.core_model_loading.reverse_op`, which cannot invert a weight-conversion op back to the checkpoint format. LlamaFactory re-raises it as RuntimeError with the known-issue explanation and workarounds. The bug is upstream in transformers, not in your config.

Source

Thrown at src/llamafactory/train/tuner.py:222

        else:
            output_dtype = getattr(torch, model_args.infer_dtype)

        setattr(model.config, "torch_dtype", output_dtype)
        model = model.to(output_dtype)
        logger.info_rank0(f"Convert model dtype to: {output_dtype}.")

    # Prepare save arguments (safe_serialization removed in transformers v5.0.0)
    save_kwargs = {
        "save_directory": model_args.export_dir,
        "max_shard_size": f"{model_args.export_size}GB",
    }
    if not is_transformers_version_greater_than("5.0.0"):
        save_kwargs["safe_serialization"] = not model_args.export_legacy_format

    try:
        model.save_pretrained(**save_kwargs)
    except NotImplementedError as err:
        raise RuntimeError(
            "Failed to export model: weight conversion reversal is not supported for this model architecture "
            "(NotImplementedError in transformers.core_model_loading.reverse_op). "
            "This is a known issue with transformers>=5.0 for certain model types (e.g. Mistral/Ministral). "
            "Workarounds: (1) use transformers<5.0, or (2) report the issue to the transformers repository."
        ) from err

    if model_args.export_hub_model_id is not None:
        # Prepare push arguments (safe_serialization removed in transformers v5.0.0)
        push_kwargs = {
            "max_shard_size": f"{model_args.export_size}GB",
        }
        if not is_transformers_version_greater_than("5.0.0"):
            push_kwargs["safe_serialization"] = not model_args.export_legacy_format

        model.push_to_hub(
            model_args.export_hub_model_id,
            token=model_args.hf_hub_token,
            **push_kwargs,

View on GitHub (pinned to f28afaf635)

Solutions

1. Downgrade: `pip install "transformers<5.0.0"` (and align any dependent pins), then re-export.
2. Report/track the issue in the transformers repository and upgrade once fixed.
3. As a stopgap on 5.x, save via a different route (e.g. state_dict + safetensors manually) if you accept non-standard layout.

Example fix

# shell
pip install "transformers<5.0.0"
llamafactory-cli export mistral_export.yaml

from packaging.version import Version
import transformers
def transformers5_export_safe() -> bool:
    return Version(transformers.__version__) < Version("5.0.0")

try:
    export_model(args)
except RuntimeError as e:
    if "reverse_op" in str(e):
        # pin transformers<5.0.0 and retry once
        raise SystemExit("Re-run with transformers<5.0.0") from e
    raise

When it happens

Trigger: `llamafactory-cli export` with transformers>=5.0.0 installed and an affected architecture (Mistral, Ministral, and other models whose conversion path lacks a reverse op).

Common situations: Environments that auto-upgraded transformers to 5.x; exporting community Mistral fine-tunes that previously worked on 4.x.

Related errors

• Dict is not supported.
• Quantization dataset is necessary for exporting.
• vLLM only accepts a single adapter. Merge them first.
• Cannot find satisfying example, considering decrease `export
• AutoGPTQ only accepts 2/3/4/8-bit quantization.

AI-assisted analysis of hiyouga/LlamaFactory@f28afaf635 (2026-08-14). Data as JSON: /api/errors/3c1ce1b1617ede7e. Report an issue: GitHub.