Training Large Language Models: Efficient Methods from Pretraining to Deployment

Ever wondered how ChatGPT was trained? Or how companies can adapt massive language models on consumer hardware? This workshop dives into the cutting-edge techniques that make modern AI training efficient and accessible.

Prerequisites

• Understanding of transformer architecture
• Basic knowledge of deep learning training concepts
• Familiarity with PyTorch is helpful but not required

What You’ll Learn

• The Complete Training Pipeline:

• Pretraining objectives and strategies

• Efficient finetuning techniques

• Making models deployable through quantization

• Reinforcement Learning from Human Feedback (RLHF)

• Parameter-Efficient Methods:

• Low-Rank Adaptation (LoRA)

• QLoRA and quantization-aware training

• Prompt tuning and soft prompts

• The trade-offs between different approaches

• Training at Scale:

• Data Parallelism (DP)

• Tensor Parallelism (TP)

• Fully Sharded Data Parallelism (FSDP)

• ZeRO optimization stages

• Memory management techniques

• Practical Implementations:

• Setting up efficient training pipelines

• Best practices for different scales of models

• Common pitfalls and how to avoid them

By Workshop’s End

You’ll gain the ability to:

• Understand the full lifecycle of LLM training
• Choose appropriate training strategies for your compute budget
• Implement efficient finetuning techniques
• Scale training across multiple GPUs effectively

Ready to master modern LLM training techniques? Workshop Link