v0.1.0

v0.1.0 Release notes

Overview

v0.1.0 is the first public release of mlia-ethos-u, the ML Inference Advisor plugin package for Arm Ethos-U NPUs. It includes the Ethos-U target plugin, Vela compiler integration, Corstone simulation backends, and bundled target profiles.

This release brings together all development since the repository’s initial commit.

Supported targets:

• Ethos-U55: 128 and 256 MAC configurations.
• Ethos-U65: 256 and 512 MAC configurations.
• Ethos-U85: 128, 256, 512, 1024, and 2048 MAC configurations.

Features included in this release:

• Model compatibility analysis: Check LiteRT / TensorFlow Lite (.tflite) models for operator support and placement, with advice on unsupported operations and inefficient activation functions.
• Vela performance estimation: Analyze TFLite and TOSA (.tosa, .tosamlir) models, with metrics covering cycles, inference time, throughput, utilization, memory usage, and memory access costs where available.
• Corstone simulation: Run supported workloads through Corstone-300, Corstone-310, and Corstone-320, reporting model-wide NPU active, idle, and total cycles, utilization, and AXI traffic counters.
• PyTorch and ExecuTorch workflows: Analyze exported PyTorch (.pt2) models through the optional mlia-converters-pytorch plugin, using TOSA for Vela estimates or ExecuTorch AOT conversion for supported Corstone runs. Prepared ExecuTorch (.pte) workloads can also run directly on supported Corstone backends.
• Layer hotspot analysis: Identify expensive operators, low MAC utilization, and memory pressure where per-layer data is available.
• CLI and Python API integration: Access Ethos-U analysis through the MLIA CLI or programmatically, including structured results from run_advisor.
• Standardized reporting: Console and JSON output include backend attribution, result-level advice, warnings, and operator or layer breakdowns where available. Vela estimates and Corstone measurements retain distinct result scopes.
• Traceable metrics: Preserve TFLite source operator IDs and associate performance breakdowns with source operators where possible. Layer metrics include explicit aggregation policies for additive work and peak memory usage.
• Robust JSON output: Inference times use milliseconds; non-finite optional Vela metrics are omitted, and required metrics that cannot be derived are marked unavailable with a reason.
• Backend setup: Lazy loading supports MLIA’s first-use backend installation flow. Corstone-320 supports its bundled Python runtime, and backend resources are discovered across MLIA resource directories.
• Usability and documentation: Input validation, clearer backend failure reporting, Python logging integration, paging for long Vela operator listings, CLI and backend guides, troubleshooting documentation, and a Python API walkthrough notebook.

Requirements and support limits:

• Requires Python 3.10 or later and MLIA 0.12.2.
• Corstone backends support Linux on x86-64 and AArch64 and require acceptance of the applicable backend EULA.
• Direct .pt2 input requires mlia-converters-pytorch.
• ExecuTorch performance supports Ethos-U55 with Corstone-300 and Ethos-U85 with Corstone-320.
• Ethos-U65 .pte performance is unsupported. For Ethos-U65 PyTorch performance estimates, use the original .pt2 input with Vela.
• TOSA performance analysis uses Vela. Vela does not accept .pte inputs.
• Reported performance figures refer to the NPU only.

v0.1.0rc2

What's Changed

• ci: Require Black Duck before syncing main and release commits by @max-shvetsov
• ci: Pin public sync to the merged workflow commit by @max-shvetsov

Full Changelog: v0.1.0rc1...v0.1.0rc2

v0.1.0rc1

What's Changed

• fix: Ignore coverage.xml which is generated from tests now by @Mousius
• fix: Use same commitizen rules as mlia-core by @Mousius
• fix: Don't over-restrict package contents by @Mousius
• build: Add internal PR review steps by @Mousius
• fix: Search backend resources across MLIA resource directories by @feibi-allen
• fix: Do not reload plugins in plugin init by @Mousius
• feat: Add proof of concept for layer hotspot analysis by @feibi-allen
• fix: Remove not in tflite compatability check by @logan-scully-arm
• ci: Add codeowners file by @logan-scully-arm
• feat: Enable Vela backend to consume TOSA graphs as input by @IsabellaARM
• test: Improve test coverage and format in advisor by @logan-scully-arm
• tests: Refactor and fix Advice Generation tests by @logan-scully-arm
• fix: Change writing to stdout to use standard Python logging by @feibi-allen
• test: Improve coverage in config tests by @logan-scully-arm
• feat: Implement E2E tests by @Mousius
• test: Improve coverage in common reporters and operators by @logan-scully-arm
• test: Improve testing in performance and pattern analysis by @logan-scully-arm
• test: Implement improved testing in data analysis by @logan-scully-arm
• feat: Add Python API support for Ethos-U target plugin by @jedgar-arm
• docs: Add Jupyter notebook for API by @IsabellaARM
• fix: Allow Ethos-U advisors to override performance collector class by @jedgar-arm
• docs: Improve documentation by @IsabellaARM
• feat: Remove advice for legacy optimization by @IsabellaARM
• feat: Modernize license in toml by @IsabellaARM
• docs(openspec): add agent guidance and tool setup by @Mousius
• Executorch aot by @feibi-allen
• test: Add Ethos-U Python API e2e tests by @jedgar-arm
• docs: Update documentation for pt2 inputs by @IsabellaARM
• feat: Auto-install backends on a first run by @max-shvetsov
• feat: Add flow to accept pte files by @feibi-allen
• fix: Update CLI tests for latest mlia-core by @max-shvetsov
• ci: Sync releases to public repo by @Mousius
• Add Vela standardized performance metrics by @jedgar-arm
• Fix vela lazy loading tests by @feibi-allen
• docs: clarify bug reporting issue routing by @logan-scully-arm
• Add Ethos-U performance field refinements by @jedgar-arm
• fix(output): Use result-level advice in sidecar JSON by @jedgar-arm
• build: Pin mlia and add trademark notices by @max-shvetsov
• feat: Use new transformer interface by @logan-scully-arm
• Add Vela and Corstone performance JSON metrics by @jedgar-arm
• fix: Emit Vela inference time in milliseconds by @jedgar-arm
• feat: Validate user input by @feibi-allen
• docs: Documentation updates by @IsabellaARM
• test: Update tests to current mlia HEAD by @oluwatamilore-arm
• feat: Bump mlia-core version to 0.12.0<0.13 by @max-shvetsov
• feat: Add contribution guidelines by @max-shvetsov
• fix: Remove brittle Ethos-U E2E table assertions by @max-shvetsov
• feat: Introduce vault secrets by @oluwatamilore-arm
• ci: Read uv default index from Vault by @max-shvetsov
• feat: Use LiteRT instead of Tensorflow Lite by @IsabellaARM
• test: Remove tests that don't test mlia-ethos-u by @oluwatamilore-arm
• feat(corstone): Support bundled Python runtime by @Mousius
• chore: Fix mlia core version to 0.12.1 by @integratedintegral21
• feat: Publish versioned release documentation by @max-shvetsov
• ci: Codex PR review by @jslater8
• refactor(ethos-u): Adopt core standardized advisor output by @Rob-Hughes-Arm
• fix: Omit non-finite Vela metrics by @Rob-Hughes-Arm
• feat(metrics): Define layer aggregation policies by @Rob-Hughes-Arm
• fix: Measure coverage for installed plugin packages by @Rob-Hughes-Arm
• docs: Align Ethos-U guidance with current output by @Rob-Hughes-Arm
• fix(vela): Enable backend tests and preserve safe provenance by @Rob-Hughes-Arm
• fix(vela): Preserve TFLite source operator IDs by @Rob-Hughes-Arm
• refactor(ethos-u): Complete canonical result ownership by @Rob-Hughes-Arm
• fix: Fix output_message not referenced by @IsabellaARM
• ci: Publish documentation to GitHub Pages by @IsabellaARM
• fix: Rewrite relative links in METADATA by @integratedintegral21
• ci: Sync release notes to public repositories by @IsabellaARM
• docs: Update documentation by @IsabellaARM
• fix(performance): Separate Corstone and Vela result scopes by @Rob-Hughes-Arm
• ci: Add PyTorch E2E release coverage by @jedgar-arm
• feat: 0.1.0 release by @max-shvetsov
• ci: Gate tagged releases on checks scans and documentation by @max-shvetsov

New Contributors

• @feibi-allen made their first contribution
• @logan-scully-arm made their first contribution
• @IsabellaARM made their first contribution
• @jedgar-arm made their first contribution
• @max-shvetsov made their first contribution
• @oluwatamilore-arm made their first contribution
• @integratedintegral21 made their first contribution
• @jslater8 made their first contribution
• @Rob-Hughes-Arm made their first contribution

Full Changelog: v0.0.2...v0.1.0rc1