Optimization of LLMs

Workflow Architecture

1. The implementation of Particle Swarm Optimization (PSO) for weight extraction was initiated but could not be successfully completed.

   • PSO Implementation Documentation.

2. Following a bottom-up approach, started working on the quantization of the model.

   • Quantized HuggingFace models, using Quanto Library.
   • Implemented 8-bit Modality Agnostic Quantizer.
   • Read Quantization Documentation for more.

3. Implemented a character-wise GPT model to gain a deeper understanding of the GPT structure.

   • Read GPT Implementation Documentation for more.
   • Quantized the implemented GPT model.

4. Working on Finetuning the LLMs.

   • Fine-tuning Documentation.
5. Starting up with Federated Learning
6. Deployment of Models.
   • Deployment Documentation.

Future Work

• Learn and implement federated learning.
• Explore the implementation of a 2-bit quantizer for edge devices.
• Quantize and fine-tune the LLaMA model.

PSO Implementation ->