[MateoLostanlen]

• NCNN is only fast on ARM. On x86 servers the engine used NCNN anyway and only logged a hint to switch to ONNX.
• Classifier now defaults to NCNN on ARM and ONNX elsewhere. An explicit format is still honored.
• Same model weights (v8.1.0), only the runtime changes. On an 8 vCPU Xeon, ONNX ran about 1.5x faster than NCNN in a quick benchmark.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 79.46%. Comparing base (f2dd3b4) to head (a8e7835).

Additional details and impacted files

@@           Coverage Diff            @@
##           develop     #429   +/-   ##
========================================
  Coverage    79.46%   79.46%           
========================================
  Files            6        6           
  Lines          852      852           
========================================
  Hits           677      677           
  Misses         175      175           

Flag	Coverage Δ
unittests	79.46% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[frgfm]

Interesting! Can you share more data of your benchmark to have a better informed decision?
I think we need peak RSS & latency on solobatch inference but on AMD64 with different config (your laptop + a typical VM we'd be running it on). If the 1.5x stays on both, it's a huge bonus. I just want to make sure we consider more likely potential bias here 👍

Excellent initiative either way! I can also open the opportunity to have different docker depending on the target architecture (NCNN installed when ARM, onnx when AMD64)

[MateoLostanlen]

Thanks @frgfm ! I measured the 1.5x while setting up the engine on the Argentinian VM, but honestly that's not really the point here. The code already logged that ONNX is recommended on non ARM hosts and kept NCNN anyway, so this just follows that advice.
Nothing changes on ARM, so all Pi deployments keep NCNN. Only x86 hosts (VMs, dev machines) switch to ONNX.
Agreed on per architecture images, I'd keep that for a follow up PR