Update base for Update on "make aoti_torch_empty_strided support creating incontiguous tensor"

This diff modifies the `aoti_torch_empty_strided` function to support the creation of incontiguous tensors.

To achieve it, this diff:
1. update the way to calculate the memory size by using both tensor size and the stride
2. skip stride check in ETensor by adding and checking cmake macro `USE_CUDA_BACKEND` when building with CUDA backend support.

we will soon bring the ETensor check back for every backend after migrating to use slimtensor.

Differential Revision: [D84938258](https://our.internmc.facebook.com/intern/diff/D84938258/)

[ghstack-poisoned]

Author: Gasoonjia

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-53a2908a10f414a2f85caa06703a26a40e873869
+e6f766c7d750d40603eee3f66c5915bac606b3ea

.ci/scripts/utils.sh

@@ -44,6 +44,44 @@ install_pip_dependencies() {
   popd || return
 }
 
+dedupe_macos_loader_path_rpaths() {
+  if [[ "$(uname)" != "Darwin" ]]; then
+    return
+  fi
+
+  local torch_lib_dir
+  pushd ..
+  torch_lib_dir=$(python -c "import importlib.util; print(importlib.util.find_spec('torch').submodule_search_locations[0])")/lib
+  popd
+  
+  if [[ -z "${torch_lib_dir}" || ! -d "${torch_lib_dir}" ]]; then
+    return
+  fi
+
+  local torch_libs=(
+    "libtorch_cpu.dylib"
+    "libtorch.dylib"
+    "libc10.dylib"
+  )
+
+  for lib_name in "${torch_libs[@]}"; do
+    local lib_path="${torch_lib_dir}/${lib_name}"
+    if [[ ! -f "${lib_path}" ]]; then
+      continue
+    fi
+
+    local removed=0
+    # Repeatedly remove the @loader_path rpath entries until none remain.
+    while install_name_tool -delete_rpath @loader_path "${lib_path}" 2>/dev/null; do
+      removed=1
+    done
+
+    if [[ "${removed}" == "1" ]]; then
+      install_name_tool -add_rpath @loader_path "${lib_path}" || true
+    fi
+  done
+}
+
 install_domains() {
   echo "Install torchvision and torchaudio"
   pip install --no-use-pep517 --user "git+https://github.com/pytorch/audio.git@${TORCHAUDIO_VERSION}"
@@ -101,6 +139,7 @@ install_pytorch_and_domains() {
     echo "Use cached wheel at ${cached_torch_wheel}"
   fi
 
+  dedupe_macos_loader_path_rpaths
   # Grab the pinned audio and vision commits from PyTorch
   TORCHAUDIO_VERSION=$(cat .github/ci_commit_pins/audio.txt)
   export TORCHAUDIO_VERSION

.github/workflows/pull.yml

@@ -351,6 +351,7 @@ jobs:
 
         # reinstall executorch
         bash ./install_executorch.sh --minimal
+        pip list
 
         # run python unittest
         python -m unittest examples.models.moshi.mimi.test_mimi

CMakeLists.txt

@@ -99,28 +99,6 @@ announce_configured_options(CCACHE_PROGRAM)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
-# Setup RPATH. See
-# https://gitlab.kitware.com/cmake/community/-/wikis/doc/cmake/RPATH-handling
-# Use separate rpaths during build and install phases
-set(CMAKE_SKIP_BUILD_RPATH OFF)
-# Don't use the install-rpath during the build phase
-set(CMAKE_BUILD_WITH_INSTALL_RPATH ON)
-# Automatically add all linked folders that are NOT in the build directory to
-# the rpath (per library?)
-#
-# TODO: Doesn't work for us right now because we are not installing .so's into
-# the correct locations. For example we have libcustom_ops_aot_lib.so depending
-# on _portable_lib.so, which was eventually put under
-# <site-packages>/executorch/extension/pybindings/ but this rpath is not
-# automatically added because at build time it seems `portable_lib` is being
-# built under the same directory, so no extra rpath is being added. To properly
-# fix this we need to install `portable_lib` into the correct path.
-set(CMAKE_INSTALL_RPATH_USE_LINK_PATH ON)
-# ------------------------------ OPTIONS -------------------------------------
-# WARNING: Please don't add example specific options in this CMakeLists.txt.
-# Instead please use `find_package(executorch REQUIRED)` in the example
-# directory and add a new executable in the example `CMakeLists.txt`.
-
 if(NOT EXECUTORCH_ENABLE_LOGGING)
   # Avoid pulling in the logging strings, which can be large. Note that this
   # will set the compiler flag for all targets in this directory, and for all
@@ -909,12 +887,13 @@ if(EXECUTORCH_BUILD_PYBIND)
 
   # Set RPATH to find PyTorch libraries relative to the installation location
   # This goes from executorch/extension/pybindings up to site-packages, then to
-  # torch/lib
+  # torch/lib. Don't do this to APPLE, as it will error out on the following
+  # error:
+  #
   if(APPLE)
-    set_target_properties(
-      portable_lib PROPERTIES BUILD_RPATH "@loader_path/../../../torch/lib"
-                              INSTALL_RPATH "@loader_path/../../../torch/lib"
-    )
+    # Skip setting @loader_path for APPLE, since it causes error like ld:
+    # duplicate LC_RPATH '@loader_path' in '<site-packages>/torch/lib/
+    # libtorch_cpu.dylib'
   else()
     set_target_properties(
       portable_lib PROPERTIES BUILD_RPATH "$ORIGIN/../../../torch/lib"

LICENSE

@@ -9,6 +9,7 @@ Copyright (c) 2023 Apple Inc.
 Copyright (c) 2024 MediaTek Inc.
 Copyright 2023 NXP
 Copyright (c) 2025 Samsung Electronics Co. LTD
+Copyright (c) Intel Corporation
 
 Redistribution and use in source and binary forms, with or without modification,
 are permitted provided that the following conditions are met:

backends/aoti/aoti_delegate_handle.h

@@ -71,6 +71,11 @@ using AOTInductorModelContainerGetNumConstantsFunc = AOTIRuntimeError (*)(
     AOTInductorModelContainerHandle container_handle,
     size_t* num_constants);
 
+// Update the model container with the constant tensors
+using AOTInductorModelUpdateConstantsFromBlobFunc = AOTIRuntimeError (*)(
+    AOTInductorModelContainerHandle container_handle,
+    const uint8_t* weight_blob_ptr);
+
 } // extern "C"
 
 // AOTI Delegate Handle structure
@@ -87,6 +92,7 @@ struct AOTIDelegateHandle {
   AOTInductorModelContainerGetNumInputsFunc get_num_inputs;
   AOTInductorModelContainerGetNumOutputsFunc get_num_outputs;
   AOTInductorModelContainerRunFunc run;
+  AOTInductorModelUpdateConstantsFromBlobFunc update_constants_from_blob;
 };
 
 } // namespace aoti

backends/apple/coreml/runtime/delegate/ETCoreMLAssetManager.h

@@ -99,6 +99,17 @@ NS_ASSUME_NONNULL_BEGIN
 - (NSUInteger)compact:(NSUInteger)sizeInBytes error:(NSError* __autoreleasing*)error;
 
 
+/// Executes a block with a unique temporary directory.
+///
+/// A new temporary subdirectory URL is created inside the receiver’s designated
+/// base directory. The directory is passed to the block, which can use it to
+/// perform temporary file operations. After the block finishes executing,
+/// the directory and its contents are removed.
+///
+/// @param block A block to execute. The block receives a unique URL.
+- (void)withTemporaryDirectory:(void (^)(NSURL* directoryURL))block;
+
+
 /// Purges the assets storage. The assets are moved to the trash directory and are asynchronously
 /// deleted.
 ///
@@ -117,6 +128,12 @@ NS_ASSUME_NONNULL_BEGIN
 /// contents are deleted asynchronously.
 @property (copy, readonly, nonatomic) NSURL* trashDirectoryURL;
 
+
+/// The staging directory URL, used to hold assets that are being prepared or processed
+/// before they are moved into their final location. The contents of this directory
+/// are temporary and may be cleared when no longer needed.
+@property (copy, readonly, nonatomic) NSURL* stagingDirectoryURL;
+
 /// The file manager.
 @property (strong, readonly, nonatomic) NSFileManager* fileManager;