Add MLIA NX Learning Path by annietllnd · Pull Request #3876 · ArmDeveloperEcosystem/arm-learning-paths

annietllnd · 2026-10-08T22:32:14Z

No description provided.

jslater8 · 2026-10-09T10:54:20Z

I think this is a good start but it's quite hard to follow and understand. The commands used here generic, I feel like a specific example which uses a pre made scenario and also simplified bash commands instead of multi line confusing stuff.

That's great feedback. I think some of the multilines will be needed (like changing the GPU settings), but trying to focus on what's important and remove extracirrucular stuff :) Let's jump on a call to discuss in more detail if you have ideas on simplification after this iteration?

jslater8 · 2026-10-09T10:56:28Z

+  --target-profile neural-technology \
+  --performance \
+  --profiling-data "nx-demo/captures/$CAPTURE_NAME" \
+  --output-dir runs/baseline-measured


I think it would be good to add the model explorer command in this. Just to say something like this is best viewed with model explorer and then link to another learning path.

Rob-Hughes-Arm · 2026-10-12T11:51:09Z

+layout: learningpathall
+---
+
+## Verify ADB and build Scenario Runner


I think a short introductory paragraph could be beneficial before launching into the steps.

Great point!

Rob-Hughes-Arm · 2026-10-12T11:51:51Z

+<device-serial>  device
+```
+
+Your device should appear with the status `device`. If several devices are connected, set `ANDROID_SERIAL` to the intended serial. This workflow assumes that ADB can install APKs and configure GPU debug-layer settings on your device.


Might be worth clarifying the ANDROID_SERIAL refers to an environment variable

Rob-Hughes-Arm · 2026-10-12T11:55:03Z

+
+Follow the [Scenario Runner Android build instructions](https://github.com/arm/ai-ml-sdk-scenario-runner#building-with-the-script) to build a debuggable Android arm64 APK. This experimental build needs the Android NDK, Android SDK, Gradle, and public SDK source dependencies. Use VGF Library `release/v0.11.0` to match the host tools used here.
+
+Use the unmodified runner with the public Neural Statistics layer. The capture setup enables **Enable GPU debug layers**, selects the runner application, and activates `VK_LAYER_LGL_neural_statistics` through ADB.


This sentence seems a little out of place. Is it an instruction of what the user should do, or a description of the following steps?

Rob-Hughes-Arm · 2026-10-12T11:55:59Z

+
+Use the unmodified runner with the public Neural Statistics layer. The capture setup enables **Enable GPU debug layers**, selects the runner application, and activates `VK_LAYER_LGL_neural_statistics` through ADB.
+
+## Prepare the baseline scenario on the host


What does 'baseline' mean here? (Perhaps this is mentioned on previous pages of the LP, I'm afraid I've only read this one page)

Rob-Hughes-Arm · 2026-10-12T11:56:34Z

+
+## Prepare the baseline scenario on the host
+
+Run the remaining host commands from `mlia-demo` in the same Bash shell, and stop if any command fails. Create a separate environment for the public VGF tools and LiteRT CPU runtime. This keeps their packages separate from the tested MLIA installation:


" and stop if any command fails." < is this really necessary to say?

Rob-Hughes-Arm · 2026-10-12T12:20:16Z

+
+## Validate output and capture completion
+
+First, check the device's output against the CPU reference:


As mentioned above, I'm not sure this is necessary.

Rob-Hughes-Arm · 2026-10-12T12:21:41Z

+Captured dispatches: 1; warnings: []
+```
+
+The equality check is deliberately strict. If it fails, investigate conversion, bindings, preprocessing, and your driver's documented numerical behavior. Don't silently relax it to make the test pass; assess any accepted tolerance with representative data and application accuracy requirements. Review capture warnings before using the measurements.


This line seems related to the output tensor reference check, rather than the neural statistics capture. Therefore it should probably be moved further up (though as I mentioned above, ideally we could get rid of this entirely)

Rob-Hughes-Arm · 2026-10-12T12:23:07Z

+  --output-dir runs/baseline-measured
+```
+
+The output is similar to this abbreviated result. The `<measured ...>` placeholders are numeric cycle counts in your run; their values depend on your device and workload:


This LP deals with a specific workload, so saying it "depends on your workload" is confusing.

Showing some representative example values here may be quite useful to give readers confidence that their data is reasonable.

Rob-Hughes-Arm · 2026-10-12T12:23:39Z

+└────────────────┴────────────────────┴────────┘
+```
+
+`Status ok` confirms that MLIA analyzed the capture. `Mode measured` and the `neural-technology-profiling-data` producer distinguish these results from the compiler estimates. The command also prints per-entity breakdowns. Some metrics, such as inference time and throughput, can be unavailable; don't interpret a missing value as zero.


"don't interpret a missing value as zero." < not sure this is a useful phrase

Rob-Hughes-Arm · 2026-10-12T12:25:44Z

+
+The full results are saved in `runs/baseline-measured/mlia-output/mlia-output.json`. Supplying `--profiling-data` selects the measured backend automatically, so don't add `--backend nx-performance-estimator`. Use the complete schema-version-2 capture hierarchy, not the runner's timestamp JSON or a statistics file alone.
+
+This synthetic-input run validates execution and profiling, not model accuracy or representative performance. For benchmarks, follow the [SDK performance guidance](https://github.com/arm/ai-ml-sdk-for-vulkan/blob/main/docs/source/performance.rst) and collect repeated runs under controlled conditions.


This is a good point, but doesn't really give the user enough guidance on how to use the profiling workflow along with the linked ML SDK documentation.

IsabellaARM · 2026-10-12T15:25:21Z

+source mlia-venv/bin/activate
+python -m pip install --upgrade pip
+python -m pip install \
+  mlia==0.12.3 \


this can be avoided, mlia-neural-technolody will pull the correct dependency of mlia automatically

IsabellaARM · 2026-10-12T15:26:56Z

+Install the backend from the public wheel's bundled payload:
+
+```bash
+mlia backend install nx-performance-estimator


this is now optional, it will be installed at first run if not pre-install

IsabellaARM · 2026-10-12T15:28:38Z

+
+Keep the model in your working directory. No model binaries need to be added to the Learning Paths repository.
+
+## Convert the model before analysis


Why the manual step? if converter is installed that will be done automatically

Think we wanted this to just showcase you can start with other formats too - but makes sense to remove now to simplify

IsabellaARM · 2026-10-12T15:31:05Z

+Run a compatibility check on the converted model. Use a dedicated output directory to keep this report separate from performance results:
+
+```bash
+mlia check mlia-demo-models/baseline.tosamlir \


in theory we could skip the conversion to tosamlir and use the tflite here instead, the beauty of MLIA is actually avoid any manual conversion

annietllnd · 2026-10-12T21:50:36Z

Thanks all, all addressed and ready for round 2 :)

Rob-Hughes-Arm · 2026-10-13T09:31:04Z

+`Status ok` confirms that MLIA analyzed the capture. The full results are saved in `runs/baseline-measured/mlia-output/mlia-output.json`.
+
+
+## (Optional) Inspect the graph with Model Explorer


This section shows doesn't show how to actually view the performance stats in Model Explorer, which is missing the whole point! The user doesn't need to install model explorer separately, just passing --model-explorer open is all that's needed.

Great point - old ME habits die hard! Updated now

Add MLIA NX Learning Path

9246a50

jslater8 reviewed Oct 9, 2026

View reviewed changes

Rob-Hughes-Arm reviewed Oct 12, 2026

View reviewed changes

IsabellaARM reviewed Oct 12, 2026

View reviewed changes

Address MLIA NX feedback

5bbbeaf

Rob-Hughes-Arm reviewed Oct 13, 2026

View reviewed changes

Update MLIA NX LP

6f01f7a


		Follow the [Scenario Runner Android build instructions](https://github.com/arm/ai-ml-sdk-scenario-runner#building-with-the-script) to build a debuggable Android arm64 APK. This experimental build needs the Android NDK, Android SDK, Gradle, and public SDK source dependencies. Use VGF Library `release/v0.11.0` to match the host tools used here.

		Use the unmodified runner with the public Neural Statistics layer. The capture setup enables Enable GPU debug layers, selects the runner application, and activates `VK_LAYER_LGL_neural_statistics` through ADB.


		Use the unmodified runner with the public Neural Statistics layer. The capture setup enables Enable GPU debug layers, selects the runner application, and activates `VK_LAYER_LGL_neural_statistics` through ADB.

		## Prepare the baseline scenario on the host


		## Prepare the baseline scenario on the host

		Run the remaining host commands from `mlia-demo` in the same Bash shell, and stop if any command fails. Create a separate environment for the public VGF tools and LiteRT CPU runtime. This keeps their packages separate from the tested MLIA installation:


		## Validate output and capture completion

		First, check the device's output against the CPU reference:


		The full results are saved in `runs/baseline-measured/mlia-output/mlia-output.json`. Supplying `--profiling-data` selects the measured backend automatically, so don't add `--backend nx-performance-estimator`. Use the complete schema-version-2 capture hierarchy, not the runner's timestamp JSON or a statistics file alone.

		This synthetic-input run validates execution and profiling, not model accuracy or representative performance. For benchmarks, follow the [SDK performance guidance](https://github.com/arm/ai-ml-sdk-for-vulkan/blob/main/docs/source/performance.rst) and collect repeated runs under controlled conditions.


		Keep the model in your working directory. No model binaries need to be added to the Learning Paths repository.

		## Convert the model before analysis

		`Status ok` confirms that MLIA analyzed the capture. The full results are saved in `runs/baseline-measured/mlia-output/mlia-output.json`.


		## (Optional) Inspect the graph with Model Explorer

Conversation

annietllnd commented Oct 8, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

annietllnd commented Oct 12, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants