Parameter Efficient Finetuning for LLM

This research addressed the challenge of determining optimal hyperparameters and training configurations for state-of-the-art (SoTA) Large Language Models when implementing Parameter Efficient Fine-tuning (PEFT) methods. The goal was to establish best practices for fine-tuning these models while maintaining computational efficiency and achieving superior performance across diverse use cases.

Enabled enterprise customers to fine-tune Large Language Models for domain-specific business applications while maintaining optimal cost efficiency and performance. This capability was integrated into AWS SageMaker Canvas, Amazon's managed machine learning service for no-code/low-code model development.

• Designed and executed large-scale experiments to understand the effects of fine-tuning LLMs using PEFT methods (LoRA, qLoRA), uncovering overall performance envelopes and investigating failure conditions.
• Conducted comprehensive analysis of the hyperparameter landscape for PEFT using multiple methodologies and performance profiling techniques.
• Analyzed performance trade-offs when applying PEFT to domain-specific datasets in resource-constrained environments.
• Experimented with multiple distributed training paradigms including Fully Sharded Data Parallel (FSDP), Distributed Data Parallel (DDP), and Data Parallel (DP).

• Large Language Models
• PEFT
• LoRA
• qLoRA
• AWS SageMaker Canvas
• Distributed Training