sme-executorch-profiling

Build SME2-on and SME2-off ExecuTorch runner binaries for the profiling kit. Use when Codex needs macOS or Android executor_runner binaries, when executorch was refreshed, when CMake presets changed, when runners are missing, or before running SME2-on versus SME2-off profiling experiments.

Run SME2-on versus SME2-off profiling experiments for an exported ExecuTorch .pte model on macOS or Android. Use when Codex needs to create or edit pipeline JSON configs, execute timing or XNNPACK trace runs, compare runner variants, collect ETDump artifacts, rerun analysis-only mode, or validate profiling output directories.

Set up or repair the SME2 ExecuTorch profiling workspace. Use when Codex needs to create the repo-local Python virtual environment, clone or refresh the required executorch checkout, install ExecuTorch for profiling, handle setup failures, or validate that the workspace is ready before building runners, exporting models, or profiling.

Run or audit the end-to-end SME2 ExecuTorch profiling smoke test. Use when Codex needs to prove a fresh workspace works, validate setup/build/export/pipeline wiring with toy_cnn, check generated artifacts before a PR or demo, or diagnose which workflow stage is broken.

Analyze existing SME2 ExecuTorch profiling outputs. Use when Codex needs to regenerate ETDump CSV files, compute robust latency statistics, compare SME2-on versus SME2-off timeline CSVs, identify operator-category bottlenecks, inspect portable versus delegated behavior, or derive XNNPACK kernel evidence from xnntrace logs.

Export a registered PyTorch model to an ExecuTorch .pte artifact for SME2 profiling. Use when Codex needs to convert toy_cnn, EfficientSAM, EdgeTAM image encoder, or another registered model to portable ExecuTorch format with XNNPACK or portable backend, choose dtype or quantization settings, save graph metadata, or prepare artifacts for profiling.

Generate Markdown reports from SME2 ExecuTorch profiling outputs. Use when Codex needs a shareable performance report, macOS and Android combined report, latency and operator-category summary, SME2-on versus SME2-off speedup write-up, or report validation before publishing results.

Onboard the EdgeTAM image encoder as a model_profiling registry entry for ExecuTorch export and SME2 profiling. Use when Codex needs to clone or wire the EdgeTAM source, preserve third-party license notices, create an EagerModelBase wrapper, register edgetam_image_encoder, download or verify checkpoints, or prepare the image encoder for export.