From b0ec99a27462159eeb8bb3919d4f00773ca22bef Mon Sep 17 00:00:00 2001
From: RJ Ascani <rja@meta.com>
Date: Fri, 30 Jan 2026 10:42:07 -0800
Subject: [PATCH 1/3] Add initial AGENTS.md

This PR adds an initial AGENTS.md file. It was generated by Claude
Code's /init.
---
 AGENTS.md | 327 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 327 insertions(+)
 create mode 100644 AGENTS.md

diff --git a/AGENTS.md b/AGENTS.md
new file mode 100644
index 00000000000..22927517579
--- /dev/null
+++ b/AGENTS.md
@@ -0,0 +1,327 @@
+# AGENTS.md
+
+This file provides guidance to coding agents, such as Claude Code (claude.ai/code) when working with code in this repository.
+
+## Overview
+
+ExecuTorch is PyTorch's unified solution for deploying AI models on-device (smartphones to microcontrollers). It uses ahead-of-time (AOT) compilation to prepare PyTorch models for edge deployment via export → compile → execute workflow, producing `.pte` (PyTorch ExecuTorch) files for on-device execution.
+
+**Production Status**: Powers billions of users at Meta across Instagram, WhatsApp, Quest 3, Ray-Ban Meta Smart Glasses.
+
+## Build System
+
+### Initial Setup
+
+```bash
+# Install dependencies
+./install_requirements.sh
+./install_executorch.sh
+
+# Configure CMake (one-time setup after cloning/pulling)
+rm -rf cmake-out
+mkdir cmake-out
+cd cmake-out
+cmake ..
+
+# Build (use core count + 1 for -j)
+cmake --build cmake-out -j9
+```
+
+### Common CMake Build Options
+
+Configure with `-D<OPTION>=ON/OFF`:
+- `EXECUTORCH_BUILD_XNNPACK` - XNNPACK backend (CPU optimization)
+- `EXECUTORCH_BUILD_COREML` - CoreML backend (Apple Neural Engine)
+- `EXECUTORCH_BUILD_MPS` - Metal Performance Shaders (Apple GPU)
+- `EXECUTORCH_BUILD_QNN` - Qualcomm backend
+- `EXECUTORCH_BUILD_VULKAN` - Vulkan backend (cross-platform GPU)
+- `EXECUTORCH_BUILD_KERNELS_OPTIMIZED` - Optimized kernels (vs portable)
+- `EXECUTORCH_BUILD_KERNELS_QUANTIZED` - Quantized kernel support
+- `EXECUTORCH_BUILD_EXTENSION_MODULE` - Module wrapper (simplified C++ API)
+- `EXECUTORCH_BUILD_EXTENSION_DATA_LOADER` - Data loader implementations
+- `EXECUTORCH_BUILD_PYBIND` - Python bindings
+- `EXECUTORCH_BUILD_TESTS` - Build tests
+- `EXECUTORCH_OPTIMIZE_SIZE` - Size optimization (-Os vs -O2)
+
+### Building Models
+
+Use Makefile targets for common model builds:
+
+```bash
+# Build model runners with specific backends
+make llama-cpu           # Llama with CPU
+make voxtral-cuda        # Voxtral with CUDA backend
+make whisper-metal       # Whisper with Metal backend (macOS)
+make gemma3-cuda         # Gemma3 with CUDA
+make help                # Show all available targets
+
+# Clean build artifacts
+make clean
+```
+
+Model binaries output to `cmake-out/examples/models/<model>/`
+
+## Testing
+
+### Python Tests
+
+```bash
+# Run all Python tests from repo root
+pytest
+
+# Test-specific paths are configured in pytest.ini
+# Tests cover: backends, codegen, devtools, examples, exir, kernels, runtime
+```
+
+### C++ Tests
+
+```bash
+# Build and run all C++ tests
+./test/run_oss_cpp_tests.sh
+
+# Run specific test directory
+./test/run_oss_cpp_tests.sh runtime/core/test/
+
+# Build size test (runtime + portable kernels)
+sh test/build_size_test.sh
+```
+
+### Single Test Execution
+
+For focused testing during development:
+```bash
+# Python: Run specific test file
+pytest path/to/test_file.py
+
+# Python: Run specific test function
+pytest path/to/test_file.py::test_function_name
+
+# C++: Build and run single directory (see run_oss_cpp_tests.sh)
+```
+
+## Code Architecture
+
+### Core Runtime (`runtime/`)
+
+**Minimal C++ runtime** - Loads and executes `.pte` programs. No operators/backends included.
+
+- `runtime/core/` - Core types: `Tensor`, `EValue`, `Error`, `Result`, `ArrayRef`
+  - `portable_type/` - Portable C++ types compatible with bare-metal (no stdlib dynamic allocation)
+  - `exec_aten/` - Dual-mode header: uses either ATen types or ExecuTorch portable types
+- `runtime/executor/` - `Program` and `Method` execution interfaces
+- `runtime/kernel/` - Kernel registration and management
+- `runtime/platform/` - Platform Abstraction Layer (PAL) for architecture-specific code
+- `runtime/backend/` - Backend delegate runtime APIs
+
+### Export and Lowering (`exir/`)
+
+**EXIR (EXport Intermediate Representation)** - AOT library for model capture and lowering.
+
+- `exir/capture/` - Program capture via `torch.export`
+- `exir/dialects/` - Op sets for different IR dialects (ATen, Edge, Backend)
+- `exir/passes/` - Compiler passes for optimization and transformation
+- `exir/backend/` - Backend delegate AOT APIs for partitioning and lowering
+- `exir/emit/` - Converts ExportedProgram to ExecuTorch execution instructions
+- `exir/_serialize/` - Serializes to final `.pte` artifact
+
+**Key Concepts**:
+- **ATen dialect**: Full PyTorch operator set
+- **Edge dialect**: Reduced operator set (Core ATen) for edge devices
+- **Backend dialect**: Device-specific lowered representation
+
+### Kernels (`kernels/`)
+
+Operator implementations registered with runtime:
+
+- `kernels/portable/` - **Reference implementations** of Core ATen ops (portable C++, no SIMD)
+  - `functions.yaml` - Defines portable kernel signatures and metadata
+- `kernels/optimized/` - **Optimized implementations** (SIMD, architecture-specific)
+- `kernels/quantized/` - Quantized kernel implementations
+- `kernels/prim_ops/` - Primitive ops for control flow and symbolic operations
+- `kernels/aten/` - ATen-compatible kernels (when building with PyTorch)
+
+**Selective Build**: Use YAML files to specify only needed operators, reducing binary size.
+
+### Backends (`backends/`)
+
+Hardware-specific delegate implementations. Each backend has:
+1. **Partitioner** - Splits graph into device-compatible subgraphs
+2. **Quantizer** - Backend-specific quantization (optional)
+3. **Runtime** - Executes lowered graph on target hardware
+
+**Available Backends**:
+- `backends/xnnpack/` - XNNPACK (optimized CPU kernels)
+- `backends/apple/coreml/` - CoreML (Apple Neural Engine)
+- `backends/apple/mps/` - MPS (Metal Performance Shaders)
+- `backends/qualcomm/` - Qualcomm (Hexagon DSP/NPU)
+- `backends/arm/` - ARM (Ethos-U NPU, VGF)
+- `backends/vulkan/` - Vulkan (cross-platform GPU)
+- `backends/mediatek/`, `backends/samsung/`, `backends/cadence/`, etc.
+
+### Extensions (`extension/`)
+
+Higher-level libraries built on runtime:
+
+- `extension/module/` - **Module wrapper** - Simplified C++ API (`Module::load()`, `Module::forward()`)
+- `extension/llm/` - LLM-specific runtime, quantization, and optimization
+  - `runner/` - Text/multimodal runner API for LLMs
+  - `tokenizers/` - Tokenizer implementations
+  - `custom_ops/` - Custom ops for LLM optimization
+- `extension/data_loader/` - FileDataLoader, BufferDataLoader implementations
+- `extension/tensor/` - TensorPtr and tensor creation utilities
+- `extension/pybindings/` - Python runtime API
+- `extension/runner_util/` - Helpers for C++ execution tools
+
+### Developer Tools (`devtools/`)
+
+- `devtools/etdump/` - **ETDump**: Runtime profiling/debugging data format
+- `devtools/etrecord/` - **ETRecord**: AOT debug artifact
+- `devtools/inspector/` - Python API to inspect ETDump and ETRecord
+- `devtools/bundled_program/` - Model validation tool with test I/O
+
+### Schema (`schema/`)
+
+FlatBuffer schema defining `.pte` file format (`program.fbs`).
+
+### Code Generation (`codegen/`)
+
+Tools to autogenerate kernel bindings between operators and runtime.
+
+## Key Workflows
+
+### Model Export to .pte
+
+```python
+import torch
+from executorch.exir import to_edge_transform_and_lower
+from executorch.backends.xnnpack.partition.xnnpack_partitioner import XnnpackPartitioner
+
+# 1. Export PyTorch model
+model = MyModel().eval()
+example_inputs = (torch.randn(1, 3, 224, 224),)
+exported_program = torch.export.export(model, example_inputs)
+
+# 2. Lower to backend (switch backends by changing partitioner)
+program = to_edge_transform_and_lower(
+    exported_program,
+    partitioner=[XnnpackPartitioner()]  # or CoreMLPartitioner(), QnnPartitioner()
+).to_executorch()
+
+# 3. Save .pte file
+with open("model.pte", "wb") as f:
+    f.write(program.buffer)
+```
+
+### C++ Runtime Execution
+
+```cpp
+#include <executorch/extension/module/module.h>
+#include <executorch/extension/tensor/tensor.h>
+
+// Using Module API (simplified)
+Module module("model.pte");
+auto input = make_tensor_ptr({1, 3, 224, 224}, input_data);
+auto outputs = module.forward(input);
+
+// Or using lower-level Program/Method API
+// See runtime/executor for Program and Method classes
+```
+
+### LLM Export and Execution
+
+```bash
+# Export LLM
+python -m executorch.extension.llm.export.export_llm \
+  --model llama3_2 \
+  --output llama.pte
+
+# C++ execution
+#include <executorch/extension/llm/runner/text_llm_runner.h>
+auto runner = create_llama_runner("llama.pte", "tiktoken.bin");
+runner->generate("Hello", config);
+```
+
+## Code Style
+
+### C++ Guidelines
+
+- **C++17 standard**
+- **Function names**: `lower_snake_case()` (PyTorch convention, not Google style)
+- **File names**: `lower_snake_case.cpp` (not `.cc` or `PascalCase.cpp`)
+- **Headers**: Use `#pragma once` (not include guards)
+- **Includes**: Use `<angle brackets>` always (not `"quotes"`)
+- **Documentation**: Doxygen syntax (`/** ... */` or `///`) with `@param`, `@retval`
+- **TODOs**: Reference issue numbers `TODO(#123): description` (not usernames)
+
+### Portability Restrictions
+
+ExecuTorch targets bare-metal systems. **Do not use**:
+- **Exceptions** - Not supported on many microcontrollers
+- **RTTI/dynamic_cast** - Adds binary bloat
+- **Dynamic allocation** - `std::vector`, `std::string`, `new`, `malloc()`
+- **Threading** - `std::thread`, `thread_local`, mutexes (except in optional libraries)
+- **Iostreams** - File I/O via iostream
+
+**Allowed stdlib**:
+- Static concepts: `std::move`, `std::forward`
+- Metaprogramming: `std::is_floating_point<>`, `std::enable_if<>`
+- Pure functions: `<cmath>`, `<cstring>` (avoid allocation functions like `strdup()`)
+- Placement `new` (with manual destruction)
+
+### Python Style
+
+Follow PyTorch core project conventions. Use `lintrunner`:
+
+```bash
+lintrunner init      # Setup
+lintrunner           # Check
+lintrunner -a        # Auto-fix
+```
+
+## Build Presets
+
+See `CMakePresets.json` for workflow presets:
+- `llm-release` - LLM build with optimized kernels
+- `llm-release-cuda` - LLM with CUDA backend
+- `llm-release-metal` - LLM with Metal backend
+
+Model-specific presets in `examples/models/<model>/CMakePresets.json`.
+
+## File Layout Constraints
+
+**IMPORTANT**: The repo directory must be named exactly `executorch` (not `executorch-1`, `et`, etc.) due to include path requirements (`#include <executorch/...>`). See CMakeLists.txt:290-298 for enforcement.
+
+## Backend Development
+
+When adding a backend delegate:
+1. Implement partitioner (split graph into subgraphs)
+2. Implement runtime delegate (execute lowered graph)
+3. Optional: Custom quantizer for backend-specific quantization
+4. Add CMake integration in `backends/<name>/CMakeLists.txt`
+5. Update backend list in root `CMakeLists.txt`
+
+See `docs/source/backend-delegates-integration.md` for detailed guide.
+
+## Memory and Binary Size
+
+- **Selective Build**: Specify operators via YAML to reduce binary size
+- **Memory Planning**: AOT memory allocation strategies
+- **Size Optimization**: Use `-DEXECUTORCH_OPTIMIZE_SIZE=ON` for `-Os`
+- **Strip Logging**: `-DEXECUTORCH_ENABLE_LOGGING=OFF` removes log strings
+- **Disable Verification**: `-DEXECUTORCH_ENABLE_PROGRAM_VERIFICATION=OFF` saves ~20KB
+
+## CI and Testing Requirements
+
+- CI runs automatically on PRs: https://hud.pytorch.org/hud/pytorch/executorch/main
+- All tests must pass before merge
+- Add tests for new features/bug fixes
+- C++ tests use GTest framework
+- Python tests use pytest
+
+## Additional Resources
+
+- Full documentation: https://pytorch.org/executorch
+- API reference: https://pytorch.org/executorch/main/api-section.html
+- Backend integration guide: docs/source/backend-delegates-integration.md
+- Portable C++ programming: docs/source/portable-cpp-programming.md
+- Contributing guide: CONTRIBUTING.md

From f4c02147c4c68c4dceb31b235cd8f7cb48bab199 Mon Sep 17 00:00:00 2001
From: RJ Ascani <rja@meta.com>
Date: Fri, 30 Jan 2026 11:28:12 -0800
Subject: [PATCH 2/3] Lint ignore hud link

---
 AGENTS.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/AGENTS.md b/AGENTS.md
index 22927517579..d30f4dbac0e 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -312,7 +312,7 @@ See `docs/source/backend-delegates-integration.md` for detailed guide.
 
 ## CI and Testing Requirements
 
-- CI runs automatically on PRs: https://hud.pytorch.org/hud/pytorch/executorch/main
+- CI runs automatically on PRs: https://hud.pytorch.org/hud/pytorch/executorch/main <!-- @lint-ignore -->
 - All tests must pass before merge
 - Add tests for new features/bug fixes
 - C++ tests use GTest framework

From c38b20f09f558e1d3fd35fc264d43de61011fa9e Mon Sep 17 00:00:00 2001
From: RJ Ascani <rja@meta.com>
Date: Fri, 30 Jan 2026 11:46:45 -0800
Subject: [PATCH 3/3] Defer environment and cmake instructions to the building
 from source doc

---
 AGENTS.md | 24 +-----------------------
 1 file changed, 1 insertion(+), 23 deletions(-)

diff --git a/AGENTS.md b/AGENTS.md
index d30f4dbac0e..df80b4e543c 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -12,20 +12,7 @@ ExecuTorch is PyTorch's unified solution for deploying AI models on-device (smar
 
 ### Initial Setup
 
-```bash
-# Install dependencies
-./install_requirements.sh
-./install_executorch.sh
-
-# Configure CMake (one-time setup after cloning/pulling)
-rm -rf cmake-out
-mkdir cmake-out
-cd cmake-out
-cmake ..
-
-# Build (use core count + 1 for -j)
-cmake --build cmake-out -j9
-```
+See `docs/source/using-executorch-building-from-source.md` for environment setup and installation.
 
 ### Common CMake Build Options
 
@@ -278,15 +265,6 @@ lintrunner           # Check
 lintrunner -a        # Auto-fix
 ```
 
-## Build Presets
-
-See `CMakePresets.json` for workflow presets:
-- `llm-release` - LLM build with optimized kernels
-- `llm-release-cuda` - LLM with CUDA backend
-- `llm-release-metal` - LLM with Metal backend
-
-Model-specific presets in `examples/models/<model>/CMakePresets.json`.
-
 ## File Layout Constraints
 
 **IMPORTANT**: The repo directory must be named exactly `executorch` (not `executorch-1`, `et`, etc.) due to include path requirements (`#include <executorch/...>`). See CMakeLists.txt:290-298 for enforcement.