hiyouga/LlamaFactory · error · ValueError

All 'dcp_path', 'hf_path', and 'config_path' are required.

Error message

All 'dcp_path', 'hf_path', and 'config_path' are required.

What it means

Config patching checks config.model_type == 'internlm3' and verifies the installed transformers is newer than 4.47.1 (patcher.py:413); InternLM3 architecture support only landed in transformers 4.47.1+, so older versions cannot instantiate the model class and the loader fails fast with RuntimeError instead of a confusing AutoModel error.

Source

Thrown at scripts/dcp2hf.py:46

import torch.distributed.checkpoint as dcp
import transformers
from transformers import AutoConfig


def convert(dcp_path: str, hf_path: str, config_path: str) -> None:
    """Convert DCP model weights to HF.

    Note: this script is used to convert a DCP checkpoint to HuggingFace model format,
    it will just convert the DCP checkpoint to a HuggingFace model format, for the tokenizer,
    you may need to copy from the original model.

    Args:
        dcp_path: DCP checkpoint directory.
        hf_path: Output path (directory) for HuggingFace model.
        config_path: Path to the HuggingFace model directory containing config.json.
    """
    if not dcp_path or not hf_path or not config_path:
        raise ValueError("All 'dcp_path', 'hf_path', and 'config_path' are required.")

    print(f"Loading config from {config_path}...")
    config = AutoConfig.from_pretrained(config_path)
    architectures = getattr(config, "architectures", [])
    if architectures:
        model_cls = getattr(transformers, architectures[0], transformers.AutoModelForCausalLM)
    else:
        model_cls = transformers.AutoModelForCausalLM

    print("Initializing model on CPU...")
    model = model_cls(config).to(torch.bfloat16)

    print(f"Loading DCP from {dcp_path}...")
    state_dict = model.state_dict()
    dcp.load(state_dict, checkpoint_id=dcp_path)
    model.load_state_dict(state_dict)

    print(f"Saving to HF format at {hf_path}...")

View on GitHub (pinned to f28afaf635)

Solutions

1. pip install -U 'transformers>=4.47.1'
2. If you must stay on old transformers, use a different model family; InternLM3 cannot be trained on <= 4.47.0

Example fix

# before
pip list | grep transformers  # 4.46.x
model_name_or_path: internlm/internlm3-8b  # -> RuntimeError

# after
pip install -U 'transformers>=4.47.1'

from transformers import AutoConfig
from llamafactory.extras.packages import is_transformers_version_greater_than
if AutoConfig.from_pretrained(model_path).model_type == 'internlm3':
    assert is_transformers_version_greater_than('4.47.1'), 'pip install -U \'transformers>=4.47.1\''

def internlm3_supported() -> bool:
    from llamafactory.extras.packages import is_transformers_version_greater_than
    return is_transformers_version_greater_than('4.47.1')

try:
    run_sft(train_args)
except RuntimeError as e:
    if 'InternLM3' in str(e):
        raise SystemExit('Upgrade: pip install -U \'transformers>=4.47.1\'') from e
    raise

When it happens

Trigger: model_name_or_path is an InternLM3 checkpoint (model_type internlm3) and the environment has transformers <= 4.47.0; is_transformers_version_greater_than('4.47.1') returns False during model load.

Common situations: Reusing an older pinned environment (e.g. created for a Qwen2 run) to train internlm/internlm3-8b; CI images with transformers pinned below 4.47.

Related errors

• Both 'hf_path' and 'dcp_path' are required.
• `num_layers` {num_layers} should be divisible by `num_expand
• Device not supported: {device_name}.
• Qwen2VL requires 3D position ids for mrope.
• Stage does not supported: {stage}.

AI-assisted analysis of hiyouga/LlamaFactory@f28afaf635 (2026-08-14). Data as JSON: /api/errors/aaf024ed9fb9fde8. Report an issue: GitHub.