Loops SDK reference

The Loops Python SDK runs LoRA training on Baseten from Python. For installation and an end-to-end walkthrough, see the Loops quickstart. This reference documents each class and type.

​

Clients

• ServiceClient: provision trainers and samplers, manage the session, and list checkpoints.
• TrainingClient: run forward and backward passes, optimizer steps, and publish weights.
• SamplingClient: generate completions from current or version-pinned weights.

​

Commonly used methods

• create_lora_training_client(): provision a trainer and get a TrainingClient.
• forward_backward() and optim_step(): run one training step.
• save_state(): save a checkpoint.
• save_weights_and_get_sampling_client(): publish weights and get a pinned SamplingClient.
• sample(): generate from the trained model.
• list_checkpoints(): list a run’s checkpoints.

​

Reference

• Types: training inputs, configuration, and result handles.
• Errors: the SDK exception types and when each is raised.
• Helpers: utility functions like attach_reference_logprobs and loops_log_kl_sample_train.