llm-training-workflows

What problem does it solve?

Fine-tune and train large language models efficiently by combining parameter-efficient fine-tuning, distributed training, and reinforcement-learning post-training to reduce compute, memory, and development time.

Core Features & Use Cases

• Parameter-Efficient Fine-Tuning (LoRA/QLoRA/PEFT) to train adapters instead of whole models.
• Distributed Training with PyTorch FSDP to scale across multiple GPUs and nodes.
• GRPO Reinforcement Learning with TRL for structured output and reward-based alignment.
• Multi-adapter serving and production inference patterns for scalable deployment.
• End-to-end training pipelines covering setup, training, evaluation, and deployment in a single workflow.

Quick Start

Install required libraries and begin a PEFT-based fine-tuning workflow on a multi-GPU setup.