Merge branch 'main' into feature/standalone-runner

Author: usamahz

.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

@@ -127,6 +127,10 @@ 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)"

.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)

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

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,
 )

backends/arm/_passes/arm_pass_manager.py

@@ -125,6 +125,7 @@
     RemoveGetItemPass,
     RemoveGraphAssertsPass,
     RemoveNoopPass,
+    RemovePermutesAroundElementwiseTosaOps,
     ReplaceInfAndLimitValuesPass,
     ReplaceScalarWithTensorByProfilePass,
     RewriteAvgPool2dPass,
@@ -164,9 +165,6 @@
     PostponePermuteOpBelowSqueezeOrUnsqueezeLikeView,
 )
 
-from executorch.backends.transforms.remove_permutes_around_elementwise_ops import (
-    RemovePermutesAroundElementwiseOps,
-)
 from executorch.exir import ExportedProgram
 from executorch.exir.pass_base import ExportPass
 from executorch.exir.pass_manager import PassManager
@@ -526,6 +524,7 @@ def _tosa_pipeline(
                 DecomposeSumPass(),
                 InsertTableOpsPass(exported_program),
                 RemoveNoopPass(),
+                InsertDataLayoutCastsPass(),
             ]
         )
 
@@ -538,7 +537,7 @@ def _tosa_pipeline(
                 RewriteMatmulPass(),
                 RewritePadPass(),
                 FuseViewCopyTransformPass(),
-                RemovePermutesAroundElementwiseOps(),
+                RemovePermutesAroundElementwiseTosaOps(),
                 PostponePermuteOpBelowSqueezeOrUnsqueezeLikeView(),
                 FuseCascadedTransposeOrPermuteOps(),
                 ConvertPermuteSingletonToViewPass(),
@@ -558,7 +557,6 @@ def _tosa_pipeline(
                 EnsureUniqueOutputNodesPass(),
                 RemoveNoopPass(),
                 InsertRescalePass(),
-                InsertDataLayoutCastsPass(),
             ]
         )
 

backends/arm/_passes/decompose_permute_for_u55_pass.py

@@ -14,7 +14,6 @@
 from executorch.backends.arm._passes.arm_pass import ArmPass
 from executorch.backends.arm._passes.rewrite_slice import RewriteSlicePass
 from executorch.backends.arm.arm_vela import vela_compile
-from executorch.backends.arm.ethosu.compile_spec import EthosUCompileSpec
 from executorch.backends.arm.tosa.mapping import map_dtype
 from executorch.backends.arm.tosa.specification import get_context_spec
 from executorch.exir.dialects._ops import ops as exir_ops
@@ -112,6 +111,9 @@ def _violates_exact_constraint(
         permutation and dtype to check wheter it is supported.
         """
 
+        # Lazy import to avoid circular dependency
+        from executorch.backends.arm.ethosu.compile_spec import EthosUCompileSpec
+
         if dtype not in (torch.int8, torch.bool, torch.int16):
             return True