External Backend and Writer Plugins

hls4ml can discover and load backend implementations from external Python packages. This enables specialised flows, such as the AMD AIE backend, to live in independent projects that version and iterate at their own cadence while reusing the core conversion infrastructure.

Discovery

Plugin packages advertise themselves through the hls4ml.backends Python entry point group. Each entry exposes a callable that receives register_backend and register_writer helpers and performs any setup that is required. hls4ml automatically scans for these entry points during hls4ml.backends import so third-party backends become available without additional user configuration.

In addition to entry points, modules listed in the HLS4ML_BACKEND_PLUGINS environment variable are imported and treated as registration callables. The variable accepts an os.pathsep separated list (: on Linux/macOS or ; on Windows):

export HLS4ML_BACKEND_PLUGINS=aie4ml.plugin:another_pkg.hls4ml_backend

Authoring a Plugin

A minimal plugin registers both a backend and an accompanying writer. The example below shows how the aie4ml package exposes its backend via pyproject.toml and a register function:

[project.entry-points."hls4ml.backends"]
AIE = "aie4ml.plugin:register"

# aie4ml/plugin.py
from aie4ml.aie_backend import AIEBackend
from aie4ml.writer import AIEWriter

def register(*, register_backend, register_writer):
    register_writer('AIE', AIEWriter)
    register_backend('AIE', AIEBackend)

When the plugin is installed, hls4ml.backends.get_available_backends() will report the new backend just like the built-in FPGA toolflows.

Packaging Data Files

Backends often rely on firmware templates or device description files. These assets should be packaged alongside the Python sources using the usual setuptools mechanisms (package-data or include-package-data) so they are available from the installed distribution.

For an end-to-end example see the companion aie4ml [https://github.com/dimdano/aie4ml] package that ships alongside this project as a standalone distribution; it encapsulates the existing AMD AIE backend as an installable plugin depending on hls4ml.