together-ai-api-helper 0.1.0

A set of utils to make Together AI API easier to use. Implements the utils for the model fine-tuning and simplified endpoint fetching. Uses the official Together AI API and builds up on it (https://github.com/togethercomputer/together-python).

Together AI API Helper

A Python package that provides high-level utilities for working with Together AI's API, making endpoint management and fine-tuning operations simpler and more robust.

Features

🚀 Endpoint Management

• Automatic endpoint provisioning and configuration
• State monitoring and health checks
• Hardware configuration management

🎯 Fine-Tuning Operations

• Automated fine-tuning job lifecycle management
• Real-time job monitoring with event streaming
• Training file validation and upload
• Automatic Hugging Face repository creation

Installation

Install the package from PyPI:

uv add together-ai-api-helper

Requirements

• Python ≥ 3.13
• Together AI API key
• Hugging Face token (if you want to upload the fine-tuned models)

Setup

1. API tokens

Set your API keys as environment variables. (Note that the HF_TOKEN is optional and is only needed if you want to automatically upload the trained model to HF):

export TOGETHER_API_KEY="your-api-key-here"
export HF_TOKEN="your-hf-token-here"

Usage

Endpoint Management

from together_ai_api_helper.endpoints import EndpointClient, EndpointConfig, GenerateConfig

# Initialize the client
client = EndpointClient()

# Configure your endpoint
config = EndpointConfig(
    model_name="mistralai/Mistral-7B-Instruct-v0.1",
    display_name="My Mistral Endpoint",
    # Uses cheapest compatible hardware by default
)

# Get or create an endpoint
endpoint_name, endpoint_id = client.get_endpoint_for_model(config)
print(f"Endpoint ready: {endpoint_name} ({endpoint_id})")

Fine-Tuning

from together_ai_api_helper.training import TrainingClient, TrainingConfig

# Initialize the training client
client = TrainingClient()

# Configure your training job
config = TrainingConfig(
    model="mistralai/Mistral-7B-Instruct-v0.1",
    training_file="path/to/your/training_data.jsonl",
    suffix="my-custom-model",
    n_epochs=3,
    learning_rate=1e-5,
    **other_together_ai_params
)

# Train the model (this will monitor progress automatically)
model_name = client.start_job(config)

Logging

All operations are logged with detailed information. Logs include:

• Endpoint state changes
• Training progress and events
• Error messages and debugging information
• API interactions and timing

Log files are created automatically (can change the file paths):

• endpoints.log for endpoint operations
• training.log for fine-tuning operations

API Reference

EndpointClient Methods

• get_endpoint_for_model(config) - Get or create an endpoint
• list_active_endpoints() - List all active endpoints
• list_inactive_endpoints() - List all inactive endpoints
• restart_endpoint(endpoint_id) - Restart a stopped endpoint
• stop_endpoint(endpoint_id) - Stop a running endpoint
• delete_endpoint(endpoint_id) - Delete an endpoint

TrainingClient Methods

• start_job(config) - Start and monitor a fine-tuning job
• monitor_job(job_id) - Monitor an existing job
• cancel_job(job_id) - Cancel a running job

Contributing

This package builds on top of the official Together Python SDK.

For development use the dev extra dependencies:

uv sync --extra dev

Note that the linting and type checking will run automatically in pre-commit.

License

MIT License

Support

For issues and questions:

• Check the Together AI documentation
• Review the official Together Python SDK
• Open an issue in this repository