Fine-Tuning
Major Advantages
Simplicity
- Simply import + hpcai + API Key to use
- Supports standard PyTorch syntax
- Low learning curve, existing code modifications < 10 lines
Flexible & Controllable
- Custom Loss Function, Handwritten Training Loop
- Supports LoRA and Full Fine-tuning
- Meets research-level fine-tuning needs
Colossal-AI Inside
- Customizable application data parallelism, tensor parallelism, and pipeline parallelism
- Increase throughput and reduce memory usage
- Run larger models with less money
Reliability
- Customizable handling of node failures, supports checkpoint export
- Supports breakpoint recovery
- Model weights belong to the user and can be downloaded and deployed at any time
Target Scenarios
Scale experiments effortlessly: template-based tuning is too restrictive for research, and custom distributed code is costly and fragile. The Fine-tuning SDK gives you full experimental freedom, letting you iterate locally and scale to large clusters without the engineering burden.
Fine-Tuning SDK Process
You Control (The Logic)
1. Dataset & Tokenizer Definitions
2. Hyperparameters (Learning Rate, Batch Size, Epochs)
3. Training Loop Construction (Step-by-step control)
4. Custom Algorithms
5. Evaluation Metrics
Install
(pip install hpcai)
[The Bridge: API_KEY]
We Handle (The Infrastructure)
1. Massive GPU Allocation & Orchestration
2. Environment Setup (CUDA, PyTorch, Dependencies)
3. Distributed Parallelism (Colossal-AI Acceleration)
4. Checkpointing & State Management
Models Support
Model training and inference rates. All prices are in USD per million tokens.
Base Model | Prefill | Sample | Train |
|---|
No information available
Frequently Asked Questions
Yes! New accounts automatically receive free credits ($5.00) to get started. This allows you to run initial experiments and test the SDK workflow without adding a credit card immediately.