Model Import

Prerequisites

Before importing, ensure the following conditions are met:

• OSS Bucket preparation: Create an OSS bucket and add a tag to it.

  Note

  • The supported OSS bucket storage classes exclude Archive, Cold Archive, and Deep Cold Archive. Buckets with server-side encryption enabled are supported. Private buckets are also supported.

  • Do not store model files in the root directory of the OSS bucket. Select an existing subdirectory or create a new one for Alibaba Cloud Model Studio to access.

  • You can import model files of any size. After import, they use the free storage space provided by Alibaba Cloud Model Studio.

• Model file preparation: Your model files must meet the import requirements and limitations. Place the model folder directly in the OSS bucket. The system detects it automatically.

Supported Base Models

You can currently import LoRA fine-tuned versions of the following base models:

Procedure

Follow these steps to import a LoRA model from OSS into Model Studio:

1. On the My Models page, click the Import Model button.

2. On the Import Model page, enter the following information:

   • Model name: Enter a display name for the model (up to 50 characters).

   • Base model: Select the base model corresponding to your LoRA model.

   • Training method: Select the training method used for the model. Options depend on the selected base model and appear automatically after selection.

   • Import source: Only "Import from OSS" is supported. This option is selected by default.

   • Bucket: Select the OSS bucket that stores your model files.

   • Model directory: Browse and select the directory in the chosen bucket that contains your model checkpoint.

   • Model encryption: To protect your data, the platform enables OSS server-side encryption (SSE-OSS) using fully managed keys. The encryption algorithm is AES256.

3. After verifying the information, click OK to submit the import request. To cancel, click Cancel to return to the list page. The system validates the model file format and integrity. If validation passes, the import starts. After completion, view the imported model on the My Models page and deploy, fine-tune incrementally, or delete it.

   Imported model states include the following: Creating (importing), Created (import succeeded, ready to deploy), Creation Failed (import failed), and Expired (model file changed). The list page also shows model name/ID, base model, source, supported deployment methods, and creation time. Use the search box at the top to filter models by name.

Import Requirements and Limitations

Before importing a LoRA model, ensure the following requirements are met:

• Required files: The OSS bucket must contain the following files:

  • adapter_model.safetensors: The LoRA adapter weight file stored in SafeTensors format.

  • adapter_config.json: The LoRA adapter configuration file containing key parameters such as rank and alpha.

• Rank parameter restriction: The rank value must be 8, 16, 32, or 64. All LoRA layers in the same model must use the same rank value.

• Models with modified vocabulary: Models that added new tokens or modified the original vocabulary during training cannot be imported. The system requires the vocabulary to exactly match that of the base model.

• Models with modified chat templates: Models that modified the chat_template during training cannot be imported. The system supports only the default chat_template of the corresponding open-source base model.

  The chat_template configuration is usually located in one of the following places:

  • The chat_template field in the model’s config.json file.

    

  • The chat_template field in the tokenizer_config.json file.

    

• Vision-language models with unfrozen VIT: For VL (Vision-Language) models, the Vision Transformer (VIT) part must be frozen. Models whose LoRA adapters include visual-related weights (indicating unfrozen VIT) cannot be imported.

  Run the following code to check:

  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)

  Method: Check whether the adapter_model.safetensors file contains visual-related weight parameters. If the file contains parameter keys that start with visual (for example, visual.encoder.layer.0...), the ViT component is not frozen, and the model cannot be imported.

  

  

FAQ