Speakers
Description
Machine Learning has become a key component of high-energy physics, particularly for real-time data processing in trigger systems and in view of the forthcoming HL-LHC upgrade. Such environments impose stringent constraints, requiring event processing under tight latency and memory requirements, thereby motivating the development of highly efficient inference solutions.
The ML4EP team at CERN is actively developing solutions to address these challenges. We present our recent work for efficient deployment of machine-learning models in triggers. We report aie4ml, a tool for porting models to next-generation AMD-FPGAs, and PQuantML, a library for hardware-aware model pruning and quantization.
For inference on CPUs/GPUs, we introduce SOFIE, a tool which translates trained models into optimized C++ code for heterogeneous architectures under performance constraints. By leveraging alpaka, SOFIE enables heterogeneous inference while eliminating host–device data-transfers. SOFIE integrates with PQuantML supporting inference on quantized models. It can also be used in offline workflows, like event reconstruction and data analysis. We demonstrate the integration of SOFIE into RooFit as a neural surrogate for likelihood evaluation, with automatic differentiation provided via CLAD, a source-to-source AD tool for C++.
For the various developments, we present example use cases, performance benchmarks, and discuss implications for future deployment and scalability.
| I read the instructions above | Yes |
|---|