Merge branch 'main' into export-D103724159

Author: anviatmeta

.ci/scripts/setup-macos.sh

@@ -116,7 +116,6 @@ setup_macos_env_variables
 # buck2 atm
 install_buck
 brew install libomp
-install_pip_dependencies
 
 # TODO(huydhn): Unlike our self-hosted runner, GitHub runner doesn't have access
 # to our infra, so compiler caching needs to be setup differently using GitHub
@@ -125,10 +124,17 @@ if [[ -z "${GITHUB_RUNNER:-}" ]]; then
   install_sccache
 fi
 
+# Install pinned torch before requirements-ci.txt so torchsr's transitive
+# torch dep is satisfied by the existing install and pip does not pull a
+# separate copy from PyPI. sccache is initialized above so source-build
+# cache misses still hit the cache.
 print_cmake_info
 install_pytorch_and_domains
-# We build PyTorch from source here instead of using nightly. This allows CI to test against
-# the pinned commit from PyTorch
+
+install_pip_dependencies
+
+# install_executorch's --use-pt-pinned-commit skips re-installing torch since
+# install_pytorch_and_domains already installed the pinned build above.
 if [[ "$EDITABLE" == "true" ]]; then
   install_executorch --use-pt-pinned-commit --editable
 else

.ci/scripts/test_lora.sh

@@ -159,7 +159,8 @@ Okay, so I need to calculate 15% of 80."
 EXPECTED_QUANT_LORA_PREFIX="
 <|im_start|>user Calculate 15% of 80?<|im_end|><|im_start|>assistant
 To calculate 15% of 80, we can multiply 80 by 15/100.
-So, 15% of 80 is equal to (80 * 15) / 100 = 1200 / 100 = 12.
+80 * 15/100 = 12.
+So, 15% of 80 is 12.
 #### 12
 The answer is: 12<|im_end|>"
 EXPECTED_QUANT_LORA_ALTERNATE_PREFIX="

.ci/scripts/utils.sh

@@ -105,7 +105,11 @@ install_pytorch_and_domains() {
   fi
   local python_version=$(python -c 'import platform; v=platform.python_version_tuple(); print(f"{v[0]}{v[1]}")')
   local torch_release=$(cat version.txt)
-  local torch_short_hash=${TORCH_VERSION:0:7}
+  # Download key must match the upload key below (basename of dist/*.whl,
+  # which always carries setup.py's resolved +gitHASH). Branch-ref pins
+  # like `release/2.11` would otherwise produce `+gitrelease` here and
+  # never hit the cache.
+  local torch_short_hash=$(git rev-parse --short=7 HEAD)
   local torch_wheel_path="cached_artifacts/pytorch/executorch/pytorch_wheels/${system_name}/${python_version}"
   local torch_wheel_name="torch-${torch_release}%2Bgit${torch_short_hash}-cp${python_version}-cp${python_version}-${platform:-}.whl"
 
@@ -123,10 +127,38 @@ install_pytorch_and_domains() {
   if [[ "${torch_wheel_not_found}" == "1" ]]; then
     echo "No cached wheel found, continue with building PyTorch at ${TORCH_VERSION}"
 
+    # Install PyTorch's own build-time deps so the source build does not
+    # silently inherit them from whatever else happens to be in the env
+    # (e.g. executorch's requirements-ci.txt).
+    pip install -r requirements-build.txt
     git submodule update --init --recursive
     USE_DISTRIBUTED=1 python setup.py bdist_wheel
     pip install "$(echo dist/*.whl)"
 
+    # Invariant: the basename setup.py just produced must match the cache
+    # URL we'd reconstruct on the next run. If they diverge (someone edits
+    # torch_wheel_name above, or PyTorch renames its wheels), the cache
+    # will silently miss and every macOS run will fall back to a ~30-min
+    # source build. Fail loudly so the regression is caught immediately.
+    shopt -s nullglob
+    local built_wheels=(dist/*.whl)
+    shopt -u nullglob
+    if [[ ${#built_wheels[@]} -ne 1 ]]; then
+      echo "ERROR: expected exactly 1 wheel in dist/, found ${#built_wheels[@]}" >&2
+      exit 1
+    fi
+    local built_wheel_name
+    built_wheel_name=$(basename "${built_wheels[0]}")
+    local expected_wheel_name="${torch_wheel_name//\%2B/+}"
+    if [[ "${built_wheel_name}" != "${expected_wheel_name}" ]]; then
+      echo "ERROR: built torch wheel name does not match cache URL key:" >&2
+      echo "  built:    ${built_wheel_name}" >&2
+      echo "  expected: ${expected_wheel_name}" >&2
+      echo "Fix torch_wheel_name construction in install_pytorch_and_domains" >&2
+      echo "in .ci/scripts/utils.sh" >&2
+      exit 1
+    fi
+
     # Only AWS runners have access to S3
     if command -v aws && [[ -z "${GITHUB_RUNNER:-}" ]]; then
       for wheel_path in dist/*.whl; do

.github/scripts/docathon-label-sync.py

@@ -9,7 +9,7 @@ def main() -> None:
     token = os.environ.get("GITHUB_TOKEN")
 
     repo_owner = "pytorch"
-    repo_name = "pytorch"
+    repo_name = "executorch"
     pull_request_number = int(sys.argv[1])
 
     g = Github(token)

.github/workflows/cuda.yml

@@ -148,6 +148,10 @@ jobs:
         # Run Qwen 3.5 MoE tests (quantize roundtrip + TurboQuant KV cache + sampler)
         python -m pytest examples/models/qwen3_5_moe/test_quantize_roundtrip.py examples/models/qwen3_5_moe/test_turboquant.py examples/models/qwen3_5_moe/test_sampler.py -v -o "addopts="
 
+        # Run Gemma 4 31B tests (quant unit tests + pipeline integration tests)
+        pip install gguf
+        python -m pytest examples/models/gemma4_31b/quant/tests/ examples/models/gemma4_31b/tests/ -v -o "addopts="
+
   export-model-cuda-artifact:
     name: export-model-cuda-artifact
     # Skip this job if the pull request is from a fork (HuggingFace secrets are not available)

CONTRIBUTING.md

@@ -321,6 +321,25 @@ CI is run automatically on all pull requests. However, if you want to run tests
 - The `test/run_oss_cpp_tests.sh` script will build and run C++ tests locally
 - Running `pytest` from the root directory will run Python tests locally. Make sure to run this after finishing [Dev Install](#dev-install).
 
+To build C++ tests manually with CMake, run the following from the repository root:
+
+```bash
+cmake . -Bcmake-out -DCMAKE_INSTALL_PREFIX=cmake-out -DEXECUTORCH_BUILD_TESTS=ON
+cmake --build cmake-out -j9 --target install
+```
+
+You can then use `ctest` to list or run individual C++ tests directly:
+
+```bash
+ctest --test-dir cmake-out -N
+ctest --test-dir cmake-out -R <test_name_regex> --output-on-failure
+```
+
+This workflow is useful when you want to rerun one test, attach a debugger to a
+test binary under `cmake-out`, or keep a build directory around for quick rebuild
+cycles. Add the same `-DEXECUTORCH_BUILD_*` options used by
+`test/run_oss_cpp_tests.sh` when the test needs optional kernels or extensions.
+
 ### Writing Tests
 To help keep code quality high, ExecuTorch uses a combination of unit tests and
 end-to-end (e2e) tests. If you add a new feature or fix a bug, please add tests

Makefile

@@ -91,7 +91,7 @@
 #
 # ==============================================================================
 
-.PHONY: voxtral-cuda voxtral-cpu voxtral-metal voxtral-mlx voxtral_realtime-cuda voxtral_realtime-cpu voxtral_realtime-metal voxtral_realtime-mlx voxtral_tts-cpu voxtral_tts-cuda whisper-cuda whisper-cuda-debug whisper-cpu whisper-metal parakeet-cuda parakeet-cuda-debug parakeet-cpu parakeet-metal parakeet-mlx parakeet-vulkan dinov2-cuda dinov2-cuda-debug sortformer-cuda sortformer-cpu silero-vad-cpu llama-cuda llama-cuda-debug llama-cpu llava-cpu gemma3-cuda gemma3-cpu qwen3_5_moe-cuda qwen3_5_moe-metal clean help
+.PHONY: voxtral-cuda voxtral-cpu voxtral-metal voxtral-mlx voxtral_realtime-cuda voxtral_realtime-cpu voxtral_realtime-metal voxtral_realtime-mlx voxtral_tts-cpu voxtral_tts-cuda whisper-cuda whisper-cuda-debug whisper-cpu whisper-metal parakeet-cuda parakeet-cuda-debug parakeet-cpu parakeet-metal parakeet-mlx parakeet-vulkan dinov2-cuda dinov2-cuda-debug sortformer-cuda sortformer-cpu silero-vad-cpu llama-cuda llama-cuda-debug llama-cpu llava-cpu gemma3-cuda gemma3-cpu gemma4_31b-cuda qwen3_5_moe-cuda qwen3_5_moe-metal clean help
 
 help:
 	@echo "This Makefile adds targets to build runners for various models on various backends. Run using \`make <target>\`. Available targets:"
@@ -126,6 +126,7 @@ help:
 	@echo "  llava-cpu           - Build Llava runner with CPU backend"
 	@echo "  gemma3-cuda         - Build Gemma3 runner with CUDA backend"
 	@echo "  gemma3-cpu          - Build Gemma3 runner with CPU backend"
+	@echo "  gemma4_31b-cuda     - Build Gemma 4 31B runner with CUDA backend"
 	@echo "  qwen3_5_moe-cuda    - Build Qwen3.5 MoE runner with CUDA backend"
 	@echo "  qwen3_5_moe-metal   - Build Qwen3.5 MoE runner with Metal backend"
 	@echo "  clean               - Clean build artifacts"
@@ -425,6 +426,15 @@ qwen3_5_moe-cuda:
 	@echo "✓ Build complete!"
 	@echo "  Binary: cmake-out/examples/models/qwen3_5_moe/qwen3_5_moe_runner"
 
+gemma4_31b-cuda:
+	@echo "==> Building and installing ExecuTorch with CUDA..."
+	cmake --workflow --preset llm-release-cuda
+	@echo "==> Building Gemma 4 31B runner with CUDA..."
+	cd examples/models/gemma4_31b && cmake --workflow --preset gemma4-31b-cuda
+	@echo ""
+	@echo "✓ Build complete!"
+	@echo "  Binary: cmake-out/examples/models/gemma4_31b/gemma4_31b_runner"
+
 qwen3_5_moe-metal:
 	@echo "==> Building and installing ExecuTorch with Metal..."
 	cmake --workflow --preset llm-release-metal

backends/aoti/slim/cuda/test/targets.bzl

@@ -1,8 +1,8 @@
-load("@fbcode_macros//build_defs:cpp_unittest.bzl", "cpp_unittest")
+load("@fbcode_macros//build_defs:gpu_cpp_unittest.bzl", "gpu_cpp_unittest")
 load("@fbcode_macros//build_defs/lib:re_test_utils.bzl", "re_test_utils")
 
 def cuda_slim_cpp_unittest(name):
-    cpp_unittest(
+    gpu_cpp_unittest(
         name = "test_" + name,
         srcs = [
             "test_" + name + ".cpp",
@@ -16,6 +16,7 @@ def cuda_slim_cpp_unittest(name):
         external_deps = [
             ("cuda", None, "cuda-lazy"),
         ],
+        hip_compatible = False,
         keep_gpu_sections = True,
         remote_execution = re_test_utils.remote_execution(
             platform = "gpu-remote-execution",

backends/arm/README.md

@@ -249,10 +249,15 @@ Some tests, with `u55`, `u85` and `vgf` in the name require external dependencie
   ```
 
 In addition, some model tests in the Arm backend require third-party libraries or packages.
-To run these tests, you need to install the required dependencies by running the script `examples/arm/setup.sh` with the flag `--setup-test-dependency`.
+To run these tests, install the required dependencies directly:
 
-Please note that installing model test dependencies is a standalone process. When using the `--setup-test-dependency` flag,
-the script will install only the necessary dependencies for model tests, skipping all other setup procedures.
+```
+bash backends/arm/scripts/install_models_for_test.sh
+```
+
+Installing model test dependencies is a standalone process. The script installs
+only the dependencies needed for model tests, skipping all other setup
+procedures.
 
 ## Using git hooks
 

backends/arm/_passes/__init__.py

@@ -140,6 +140,9 @@
 from .remove_getitem_pass import RemoveGetItemPass  # noqa
 from .remove_graph_asserts_pass import RemoveGraphAssertsPass  # noqa
 from .remove_noop_pass import RemoveNoopPass  # noqa
+from .remove_permutes_around_elementwise_tosa_ops import (  # noqa
+    RemovePermutesAroundElementwiseTosaOps,
+)
 from .replace_scalar_with_tensor_pass import (  # noqa
     ReplaceScalarWithTensorByProfilePass,
 )