Alibaba Cloud Model Studio:Model import

Prerequisites

Before you import a model, ensure that you meet the following requirements:

• OSS bucket preparation: You have created an OSS bucket and added a tag to the target bucket.

  Note

  • The Archive, Cold Archive, and Deep Cold Archive storage classes are not supported for OSS buckets. Buckets with content encryption and private buckets are supported.

  • Alibaba Cloud Model Studio cannot access files in an OSS bucket's root directory. You must place your files in a subdirectory.

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

• Model file preparation: The model files must meet the import requirements and limitations. The model folder must be placed in a subdirectory of the OSS bucket, and the system will automatically detect it.

Supported base models

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

Procedure

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

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

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

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

   • Base model: Select the base model that corresponds to your LoRA model.

   • Import source: Currently, only **Import from OSS** is supported and is selected by default.

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

3. After confirming the information, click OK. The system automatically validates the model file format and integrity. After validation passes, the import process begins. When the import is complete, you can view the model on the My Models page and perform operations such as deployment and incremental training.

Import requirements and limitations

Before you import a LoRA model, ensure that it meets the following requirements:

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

  • adapter_model.safetensors: The LoRA adapter weights, stored in SafeTensors format.

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

• Rank parameter constraints: 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 a modified vocabulary: If you added new tokens or modified the original vocabulary during training, you cannot import the model. The vocabulary must be identical to that of the base model.

• Models with a modified chat template: If the chat_template configuration was modified during training, you cannot import the model. The chat_template must be identical to the default configuration of the corresponding open-source base model.

  The chat_template configuration is typically found in one of the following locations:

  • 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 an unfrozen VIT: For Vision-Language (VL) models, the Vision Transformer (VIT) component must be frozen. If the LoRA adapter contains visual-related weight parameters, indicating that the VIT was not frozen, you cannot import the model.

  You can run the following script to check this.

  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)

  How to check: Examine the adapter_model.safetensors file for any visual-related weight parameters. If the file contains any parameter keys that start with visual (for example, visual.encoder.layer.0...), it means the VIT component was not frozen, and the model cannot be imported.

  

  

FAQ