macOS
The EdgeFirst Profiler ships a native arm64 binary for macOS 11 and later. macOS is supported for ONNX model development — the vendor accelerator backends (TFLite/Neutron/VSI delegates, Hailo, TensorRT, Ara240) are Linux-only and either fail to compile or require Linux-only runtime libraries.
The macOS-specific accelerator path is the CoreML execution provider through ONNX Runtime, which offloads supported subgraphs to the Apple Neural Engine or the Metal GPU.
Install
pip install edgefirst-profiler
curl -fsSL https://raw.githubusercontent.com/EdgeFirstAI/profiler-cli/main/install.sh | bash
The installer drops the binary into ~/.local/bin (default) or /usr/local/bin (run as root). If the install path is not on your PATH, the script prints the exact line to add to your shell profile.
Confirm:
edgefirst-profiler --version
ONNX Runtime — required
ONNX Runtime is the default and only fully-supported backend on macOS. Install it via Homebrew:
brew install onnxruntime
Homebrew installs libonnxruntime.dylib into /opt/homebrew/lib. That path is not on macOS's default dyld search path, but the profiler scans it explicitly — you do not need to set ORT_DYLIB_PATH or DYLD_LIBRARY_PATH.
If you maintain a non-Homebrew install, set ORT_DYLIB_PATH to the absolute path of your libonnxruntime.dylib.
CoreML execution provider
The bare --provider coreml flag is no longer accepted. Pass one of the three explicit compute-unit options instead (the old flag prints an error listing the replacements):
| Flag | Compute unit | Notes |
|---|---|---|
--provider coreml-cpu |
CoreML CPU kernels | Good baseline; no GPU or ANE scheduling overhead |
--provider coreml-gpu |
Metal Performance Shaders, CPU fallback | Best throughput for GPU-heavy models. Auto inference depth: 3, the only CoreML provider with a pinned platform default (measured knee — ~427 FPS at depth 2, ~476 FPS at depth 3, ~466 FPS at depth 4) |
--provider coreml-ane |
Apple Neural Engine, CPU fallback only | Uses the generic auto inference depth (2) — the ANE serializes device work internally, so deeper pipelines mainly buy preprocess overlap |
In the TUI: the launch-time dialog presents a navigable list of providers — CPU, CoreML CPU, CoreML GPU, CoreML ANE — instead of the previous fixed two-key prompt. Navigate with ↑/↓ and press Enter to confirm, or press the number key shown next to the row.
The --provider qnn-htp option also parses on macOS, but it is Android-only — on any non-Android host the run falls back to the CPU provider and says so.
The CoreML model is compiled fresh each run; this one-time compile happens during session setup and does not affect the measured inference timings.
CoreML coverage is partial — operators not supported by the selected compute unit fall back to CPU within the same graph. Studio's trace view shows exactly which ops ran where; the per-operation device-placement diagnostics are no longer printed to the terminal.
Verifying the install
edgefirst-profiler login # save Studio credentials
edgefirst-profiler # opens TUI on F1 Help
Then run a validation session — see Validation from Studio or Validation from the Profiler.
What is not supported on macOS
| Backend | Why not |
|---|---|
| TensorFlow Lite | libtensorflow-lite.so is Linux-only in the EdgeFirst distribution |
| NXP Neutron / VSI delegates | Delegates ship with NXP Linux BSPs only |
| Kinara Ara240 | ara2-proxy daemon is Linux-only |
| Hailo | HailoRT does not ship a macOS build |
| TensorRT | TensorRT is Jetson / Linux + NVIDIA GPU |
If you need to validate against these backends from a Mac, SSH into the target board and run the profiler there.