fastmachinelearning/hls4ml

Automatically converts trained models from Keras, PyTorch, and other frameworks into synthesizable HLS code, supporting multiple vendor backends (Xilinx Vivado/Vitis, Intel, Catapult). Optimizes for sub-microsecond latency inference through techniques like quantization to binary/ternary precision, distributed arithmetic, and CNN/semantic segmentation acceleration. Originally developed for high-energy physics trigger systems but now deployed across quantum control, satellite monitoring, and biomedical signal processing applications.

1,849 stars. Actively maintained with 12 commits in the last 30 days.