Linux
The EdgeFirst Profiler ships prebuilt binaries for x86_64 and aarch64 Linux against the manylinux2014 baseline (glibc 2.17+). Both architectures get the same default backend, ONNX Runtime, plus optional accelerator backends that load at runtime from their vendor libraries.
Install
pip install edgefirst-profiler
curl -fsSL https://raw.githubusercontent.com/EdgeFirstAI/profiler-cli/main/install.sh | bash
The installer drops the binary into /usr/local/bin (run as root) or ~/.local/bin (otherwise). Pin a specific release with --version:
curl -fsSL https://raw.githubusercontent.com/EdgeFirstAI/profiler-cli/main/install.sh | bash -s -- --version 1.0.1
Confirm:
edgefirst-profiler --version
Default backend: ONNX Runtime
The default build profiles .onnx models on CPU. The profiler dlopens libonnxruntime.so at startup; install it once per system:
sudo apt install libonnxruntime # or pip install onnxruntime to grab the bundled .so
sudo dnf install onnxruntime
Download a matching release from microsoft/onnxruntime and place libonnxruntime.so.<version> somewhere on the loader path (e.g. /usr/local/lib) or set ORT_DYLIB_PATH.
Optional: NVIDIA CUDA (x86_64)
Pass --provider cuda to offload ONNX inference onto an NVIDIA GPU. CUDA and cuDNN libraries must be present on the system path. If they are missing, the profiler prints an actionable error naming the missing library and the exact pip install command — with the specific nvidia-…-cu12 package filled in — to resolve it.
Auto-selected depth on x86_64 Linux with --provider cuda: inference depth 4, preprocess depth 1. Each additional ORT inference slot overlaps GPU execution with host staging, peaking at depth 4; a single preprocess thread avoids CPU contention with CUDA kernel dispatch. Measured on an RTX 4060 with YOLOv8n fp16 at 640×640, this yields approximately 352 → 447 FPS (+27%) compared to the CPU default. The launch-time dialog shows these values under Auto when CUDA is selected.
CUDA zero-copy input path (OpenGL–CUDA interop): disabled by default on x86_64 Linux. On discrete GPUs the GL→CUDA synchronization competes with CUDA inference kernel dispatch, while the PCIe DMA path runs independently — measured at 347 FPS (CPU staging) vs 238 FPS (GL interop) on an RTX 4060. Set EDGEFIRST_ENABLE_CUDA_ZEROCOPY=1 to opt in; this is useful for benchmarking or integrated targets where the PCIe copy is the bottleneck.
Note
The previous environment variable EDGEFIRST_DISABLE_CUDA_ZEROCOPY no longer has any effect.
Optional: TensorFlow Lite
The TFLite backend handles .tflite models and the XNNPACK delegate for accelerated CPU inference, and requires a working TFLite environment on the host. The TFLite C library is also the gateway to the NXP Neutron and VSI NPU delegates on i.MX targets — those delegates are .so files passed to the profiler via --delegate. See the NXP i.MX 95 and NXP i.MX 8M Plus guides.
Preferred runtime on desktop
Outside embedded i.MX targets, ONNX Runtime with a platform-appropriate execution provider — CUDA, CoreML, or plain CPU — is the preferred runtime on desktop platforms. Reach for the TFLite backend when you are targeting the NXP Neutron/VSI NPU delegates.
Verifying the install
After install, sign in to Studio and confirm the credentials are saved:
edgefirst-profiler login
edgefirst-profiler # opens TUI on F1 Help
Then run a validation session — see Validation from Studio or Validation from the Profiler.
Validation needs a writable project
Creating a validation session requires write access to the Studio project — you cannot profile against the read-only public Sample Project directly. First copy its dataset into a project you own, then create the session there.
If libonnxruntime is not found when a session starts, the error message tells you exactly which library is missing and where it was looked for.
Uninstall
pip uninstall edgefirst-profiler
Delete the binary from wherever the installer placed it:
rm /usr/local/bin/edgefirst-profiler # or ~/.local/bin/edgefirst-profiler
The profiler stores cached downloads in ~/.cache/edgefirst-profiler/ (or $EDGEFIRST_CACHE). Remove the cache directory if you want a clean slate.