Model aliases

Model aliases help you define a user-friendly name for a model or group of models, abstracting away the actual model name used by the model provider. LiteLLM provides the ability to use model aliases, which allow you to present a simplified or user-friendly model name to your end-users while invoking a different, more specific model name on the backend. This is useful when you want to abstract Bedrock model ids from clients or when dealing with multiple versions of models.

For example, you might display the model name claude-3 to the end-user while internally calling bedrock/anthropic.claude-3-5-sonnet-20240620-v1:0 on the backend.

In the configuration file proxy_config_model_alias.yaml, the model_name parameter (e.g., claude-3) is the user-facing name, while the litellm_params.model parameter contains the actual backend model id (e.g., bedrock/anthropic.claude-3-5-sonnet-20240620-v1:0) and any additional parameters needed for model configuration.

Sample configuration

model_list:
  - model_name: claude-3
    litellm_params:
      model: bedrock/anthropic.claude-3-5-sonnet-20240620-v1:0
      aws_region_name: $AWS_REGION

Steps to configure

Use the configuration file $BEDROCK_LITELLM_DIR/litellm/config/proxy_config_model_alias.yaml
To apply the new configuration, follow the steps outlined in Apply configuration changes.

Steps to test

Test the model alias by invoking the user-facing model name as shown

curl --location "https://${LITELLM_HOSTNAME}/chat/completions" \
  --header 'Content-Type: application/json' \
  --header 'Authorization: Bearer <your key>' \
  --data '{
    "model": "claude-3",
    "messages": [
      {
        "role": "user",
        "content": "what is Amazon S3"
      }
    ]
  }'