hpcai 0.1.6

Python SDK for HPC-AI cloud GPU fine-tuning

HPC-AI Python SDK

Overview

The HPC-AI Python SDK provides a powerful interface for distributed GPU training and fine-tuning on HPC-AI's cloud infrastructure.

Installation

we recommend using conda to install the SDK.

conda create -n hpcai python=3.12 -y
conda activate hpcai
git clone https://github.com/hpcaitech/HPC-AI-SDK
cd HPC-AI-SDK
pip install .

We only support installing from source currently, we will release official PIP package soon.

Quick Start

from hpcai import ServiceClient, TrainingClient

# Initialize the service client
client = ServiceClient(
    base_url="https://hpc-ai.com",
    api_key="your-api-key"
)

# Create a training client for LoRA fine-tuning
training_client = client.create_lora_training_client(
    base_model="Qwen/Qwen2.5-7B",
    rank=8,
    seed=42
)

Migration Guide

If you previously used another training SDK, update your code:

pip install .

Update your imports:

from hpcai import ServiceClient

Path Protocol

The SDK uses the hpcai:// protocol for model and checkpoint paths:

model_path = "hpcai://run-123/weights/checkpoint-001"

Note: Legacy path protocols are supported during the migration period for backward compatibility.

Environment Variables

Configure the SDK using these environment variables:

• HPCAI_API_KEY - Your API key
• HPCAI_BASE_URL - API endpoint (default: https://hpc-ai.com)
• HPCAI_TELEMETRY - Enable/disable telemetry (default: enabled)

Legacy environment variable names are supported for backward compatibility.

Features

• Distributed Training: Leverage HPC-AI's GPU cloud for efficient model training
• LoRA Fine-tuning: Memory-efficient fine-tuning with LoRA adapters
• Async Support: Full async/await support for concurrent operations
• Type Safety: Comprehensive type hints for better IDE support

Usage Example

A usage example for finetune "Qwen3-8B" model.

Documentation

API Reference

• ServiceClient API Reference - Main entry point for creating clients and querying server capabilities
• TrainingClient API Reference - Training operations including forward/backward passes and optimization
• RestClient API Reference - REST API operations for querying training runs and checkpoints

Third-Party Notice

This SDK provides interoperability with components based on the Tinker project (Apache License 2.0). Tinker is a trademark of its respective owner. This project is not affiliated with or endorsed by Thinking Machines Lab.

License

Licensed under the Apache License, Version 2.0. See LICENSE file for details.