Support configuring precision and quantization in HuggingFaceClient

[yifanmai]

Support passing the torch_dtype, load_in_8bit and load_in_4bit parameters to HuggingFaceClient from the ModelDeployment configuration.

Example usage:

prod_env/model_deployments.yaml:

model_deployments:
  # Example model with precision set to bfloat16
  - name: yifanmai/gpt2-bfloat16
    tokenizer_name: "huggingface/gpt2"
    max_sequence_length: 1024
    window_service_spec:
      class_name: "helm.benchmark.window_services.huggingface_window_service.HuggingFaceWindowService"
      args:
        pretrained_model_name_or_path: gpt2
    client_spec:
      class_name: "helm.proxy.clients.huggingface_client.HuggingFaceClient"
      args:
        pretrained_model_name_or_path: gpt2
        torch_dtype: torch.bfloat16
  # Example model with 8-bit quantization
  - name: yifanmai/gpt2-8bit
    tokenizer_name: "huggingface/gpt2"
    max_sequence_length: 1024
    window_service_spec:
      class_name: "helm.benchmark.window_services.huggingface_window_service.HuggingFaceWindowService"
      args:
        pretrained_model_name_or_path: gpt2
    client_spec:
      class_name: "helm.proxy.clients.huggingface_client.HuggingFaceClient"
      args:
        pretrained_model_name_or_path: gpt2
        load_in_8bit: true

Also clean up revision - treat it as as another kwarg rather than treating it specially.

[yifanmai]

Also, eventually we might require trust_remote_code=True to be set explicitly in the same way.

[JosselinSomervilleRoberts]

LGTM

[JosselinSomervilleRoberts]

Should we add a comment here describing common kwargs that should be specified such as revision, precision, ... ?

[yifanmai]

I would prefer to do this in the docs instead.

[JosselinSomervilleRoberts]

I think this is now handled by #1903 and we can close this