Model Import

​

Model import overview

​

Authorize OSS before first import

{
    "Action": [
        "ram:CreateServiceLinkedRole"
    ],
    "Resource": "*",
    "Effect": "Allow",
    "Condition": {
        "StringEquals": {
            "ram:ServiceName": "datahub.sfm.aliyuncs.com"
        }
    }
}

​

Upgrade from the legacy authorization method

​

OSS product not activated

​

Prepare LoRA model files

​

Required files and directory structure

​

Training parameter constraints

• Rank values: The rank must be one of 8, 16, 32, or 64, and all LoRA layers of the same model must use the same rank value; otherwise the model cannot be imported.
• Vocabulary must not be modified: Models that added new tokens or modified the original vocabulary during training cannot be imported; the vocabulary must exactly match that of the base model.
• Chat template must not be modified: Models that modified the chat_template during training cannot be imported; the chat_template must match the base model's default configuration. The chat_template is located in the chat_template field of config.json or tokenizer_config.json.
  
  
• Vision-language models must freeze VIT: Vision-language models must freeze the Vision Transformer part. If the LoRA adapter contains visual-related weight parameters (i.e., VIT is not frozen), the model cannot be imported.

from safetensors import safe_open
import argparse

def print_safetensor_structure(file_path):
    print(f"Loading safetensor file: {file_path}")
    print("="*80)

    with safe_open(file_path, framework="pt") as f:
        keys = f.keys()
        print(f"Found {len(keys)} tensors in the file:\n")

        for key in sorted(keys):
            tensor = f.get_tensor(key)
            shape = tuple(tensor.shape)
            dtype = str(tensor.dtype)
            device = tensor.device if hasattr(tensor, 'device') else 'cpu'

            lora_tag = " [LoRA]" if "lora_A" in key or "lora_B" in key else ""

            print(f"[{dtype:>14}] {shape} | {key} {lora_tag}")

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Print structure of a .safetensors LoRA adapter.")
    parser.add_argument("filepath", type=str, help="Path to the .safetensors file")
    args = parser.parse_args()

    print_safetensor_structure(args.filepath)

​

Pre-submission automatic validation

​

Import a LoRA model

1. On the My Models page, click the "Import Model" button in the upper-right corner to open the import model creation page.
2. Fill in the model information per the table below. The base model must match the base used for LoRA training, the bucket must have the bailian-datahub-access=read tag, and after selecting the model directory the system automatically validates the files.
3. After verifying the information, click OK to submit. The system automatically validates the file format and integrity, and begins importing once validation passes; to cancel, click Cancel to return to the list page. After import, view the status under Manage imported models.

​

Manage imported models

​

Status and transitions

​

Actions column availability

• Deploy: Available only in the Created state; clicking it opens the deployment creation page. For deployment operations, see Deployment and O&M. In other states, or when no deployment method is available, the button is disabled.
• Incremental training: Models imported from OSS do not support incremental training; the button is disabled.
• Delete: Models in the Creating state cannot be deleted; quantized models must be deleted from the model compression page.

​

FAQ