NEXT GENERATION TRIGGERS ML OPTIMIZATION: QUANTIZATION TECHNIQUES FOR SCALABLE INFERENCE

This project surveys state-of-the-art quantization techniques for machine learning models and evaluates how they can be applied to enable low-latency, high-throughput inference on resourceconstrained hardware such as FPGAs within CERN’s Next Generation Trigger (NGT) context. While these methods are generally platform-agnostic, our aim is to understand their trade-offs and integration within our machine learning platform. The work is structured around three core tasks: (1) researching and summarizing SOTA quantization techniques (e.g., symmetric/asymmetric, weight-only, dynamic range, mixed precision, and full-integer methods); (2) understanding, analyzing, and presenting these techniques with an emphasis on accuracy–compression–latency trade-offs relevant to NGT; and (3) developing a PoC showcasing the integration of such techniques into the NGT ML platform ngt.cern.ch, which demonstrates an automated Kubeflow workflow where users can upload a model, select a quantization strategy, and obtain a quantized version together with logs, compression plots, and structured metadata. The project establishes a foundation for enabling automated quantization workflows at CERN. Future work will focus on adding more quantization techniques, enhancing customization (e.g., calibration methods, per-layer precision, high-granularity quantization), and moving toward production-grade deployment within the NGT infrastructure. In doing so, this research bridges the gap between general-purpose quantization methods and their specialized application for FPGA-based real-time inference at CERN.