Profile a model on a neural accelerator with MLIA for VS Code
Introduction
Prepare your model and VS Code workspace
Prepare the Android workload
Capture NX profiling data
Analyze measured performance in VS Code
Next Steps
Profile a model on a neural accelerator with MLIA for VS Code
Prepare the capture helper
Continue in the same Bash terminal. Save the
capture helper
as capture_nx.py in $NX_WORKSPACE.
The helper runs the scenario once. It captures the first dispatch with statistics mode 1. Then, it retrieves and validates the capture, and restores the previous GPU debug settings. Keep the device connected until the helper finishes.
To run the capture steps manually, follow the
Arm ML Inference Advisor (MLIA) for Visual Studio Code capture instructions
. After retrieving the capture, set CAPTURE_ROOT to the extracted directory containing capture.json. Then, continue with
Analyze measured performance in VS Code
using your capture and matching VGF model.
Run and retrieve one capture
Run the helper with the workspace and Android Arm64 layer from the Prepare the Android workload page:
cd "$NX_WORKSPACE"
python3.12 capture_nx.py --workspace "$NX_WORKSPACE" --layer "$LAYER_LIB"
On success, the helper reports Complete mode-1 capture. It saves logs in the printed Run directory beneath runs/.
After a successful run, set the analysis input from the saved capture path:
export CAPTURE_ROOT="$(cat "$NX_WORKSPACE/last-capture.txt")"
printf 'Profiling data: %s\n' "$CAPTURE_ROOT"
CAPTURE_ROOT is the directory containing capture.json and the captured statistics. Select this directory as Profiling data in MLIA.
For capture failures, see the MLIA capture troubleshooting guide . If device restoration fails, reconnect the same device and run the recovery command printed by the helper.
What you’ve accomplished and what’s next
You’ve captured one dispatch.
Next, you’ll open the capture and its matching VGF in MLIA for Visual Studio Code.