GPT2-Quantization-HPC

Overview

This project implements a post-training quantization (PTQ) framework for GPT-2 to reduce inference latency and memory requirements on high-performance computing (HPC) systems. It includes:

• Post-Training Quantization: Reducing model weights to int8 precision with symmetric quantization.
• Custom CUDA Kernel: A tile-based int8 matrix multiplication kernel integrated with PyTorch for efficient computations.
• Performance Insights: Analysis of quantization's impact on memory usage and inference speed.

Key Features

1. Post-Training Quantization for GPT-2 using int8 weights and activations.
2. Custom CUDA kernel for int8 GEMM operations.
3. Integration with PyTorch for seamless execution.
4. Engineering insights on building and integrating custom HPC kernels.

Files and Structure

• src/: Contains source code.
  • __init__.py: Initialization for the module.
  • int8_matmul_kernel.cu: Custom CUDA kernel for int8 matrix multiplication.
  • int8_matmul.cpp: PyTorch extension for the custom kernel.
  • setup.py: Script to build and install the CUDA extension.
  • modelq.py: Quantized GPT-2 model.
  • llm.py: Language model utilities.
  • test_int8_mm.py: Script for testing the custom int8 kernel.
• docs/: Documentation and references.
  • article.pdf: Paper detailing the project and results.