API ReferenceServiceClient

ServiceClient

Primary entry point for the MinT API

Overview

ServiceClient is the main interface for accessing MinT’s functionality. All operations begin by creating a ServiceClient instance.

Key Methods

Server Information

get_server_capabilities()
get_server_capabilities_async()

Query supported features and capabilities of the MinT server.

Training Client Creation

create_lora_training_client(
    base_model,
    rank=32,
    train_mlp=True,
    train_attn=True,
    train_unembed=True,
    seed=None
)

Initialize a TrainingClient for LoRA fine-tuning.

Parameters:

  • base_model - Model identifier (e.g., “Qwen/Qwen3-4B-Instruct-2507”)
  • rank - LoRA rank dimension (default: 32)
  • train_mlp - Train MLP layers (default: True)
  • train_attn - Train attention layers (default: True)
  • train_unembed - Train unembedding layer (default: True)
  • seed - Random seed for reproducibility

Load from Checkpoint

create_training_client_from_state(path)
create_training_client_from_state_with_optimizer(path)
  • from_state - Load weights only
  • from_state_with_optimizer - Load weights and optimizer state

Sampling Client Creation

create_sampling_client(model_path=None, base_model=None)

Create a SamplingClient from either:

  • A saved checkpoint (model_path)
  • A base model (base_model)

REST Client

create_rest_client()

Access REST endpoints for checkpoint management and metadata operations.

Async Variants

All methods include async variants (e.g., create_lora_training_client_async()).

Model LineupTrainingClient