fastmachinelearning/hls4ml

Health v2 · maintenance only

Adoption not part of the score

2114 stars · 582 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

hls4ml is a Python package that converts machine learning models from Keras, PyTorch, and ONNX into high-level synthesis (C++) code for FPGA firmware. It targets ultra-low-latency, low-power neural network inference, originally for LHC trigger systems but now used across scientific and industrial domains.

Use cases

• deploy neural networks on FPGAs with microsecond latency
• convert Keras or PyTorch models to HLS firmware
• build low-latency trigger systems for particle physics detectors
• run quantized neural network inference on low-power hardware
• implement ML inference for satellite or biomedical signal processing
• compile ONNX models to FPGA bitstreams

When to choose

• you need ultra-low-latency (microsecond-scale) ML inference on FPGAs
• your model is small and can be compressed with quantization and pruning
• you use supported architectures like MLPs, CNNs, or RNNs with Keras, PyTorch, or ONNX
• you work with AMD/Xilinx, Intel, or Catapult HLS toolchains

When to avoid

• you need large models like transformers or graph networks, which are not yet stable
• you target Windows or macOS, which are not supported
• you don't have access to a vendor HLS toolchain
• you need GPU or CPU inference rather than FPGA deployment

Facets

library · maturity active

machine-learning compiler llm-inference embedded machine-learning embedded-systems hardware deep-learning gpu-computing python embedded fpga high-level-synthesis neural-network-inference low-latency quantization vivado vitis catapult keras pytorch onnx qkeras high-energy-physics linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="fastmachinelearning/hls4ml")