​
Supported Frameworks

Below are the supported frameworks for models that can be uploaded into Inferless.
  • Pytorch
  • Tensorflow
Any other framework is not currently supported as of now. We will keep adding more frameworks as the support for them is added to our platform

​
File structure requirements(Pytorch)

In the case of a PyTorch model in Google Vertex AI,
  • The model file needs to have a model.pt file
  • By default, Google Vertex AI creates it as model.pth.
To make sure that the model is saved as a model.pt, please follow the below instructions.
  • For Vertex AI, typically when using the pytorch prebuilt container for training models we create a model archive file as follows :
mar_config = {
        "MODEL_NAME": model_display_name,
        "HANDLER": handler_path,
        "SERIALIZED_FILE": f"{model_artifacts_dir}/",
        "VERSION": model_version,
        "EXTRA_FILES": ",".join(extra_files),
        "EXPORT_PATH": f"{model_mar.path}/model-store",
    }    # generate model archive command
    archiver_cmd = (
        "torch-model-archiver --force "
        f"--model-name {mar_config['MODEL_NAME']} "
        f"--serialized-file {mar_config['SERIALIZED_FILE']} "
        f"--handler {mar_config['HANDLER']} "
        f"--version {mar_config['VERSION']}"
    )
  • Since Inferless runtime for pytorch supports model.pt format, the <pytorch-model-file> specified above should be of .pt format.
    • Eg: "SERIALIZED_FILE": f"{model_artifacts_dir}/ model.pt "
  • Please ensure to save your model in .pt format and in your custom_handler.py file do the necessary changes to load the model from .pt file.
  • You can now use this model file to import your file into Inferless.
You can view a example below on how we stored a file as per the steps above: <Video here>

​
**Tensorflow**

  • There is no need for any additional steps/file modifications. The Vertex AI link of the generated model can directly be used for import.