Merge branch 'main' into jz/update-android-readme

Author: kirklandsign

CMakeLists.txt

@@ -48,21 +48,33 @@ project(executorch)
 # MARK: - Start EXECUTORCH_H12025_BUILD_MIGRATION --------------------------------------------------
 
 include(${PROJECT_SOURCE_DIR}/tools/cmake/common/preset.cmake)
+include(${PROJECT_SOURCE_DIR}/tools/cmake/Utils.cmake)
+include(CMakeDependentOption)
+include(ExternalProject)
 
 if(NOT CMAKE_CXX_STANDARD)
   set(CMAKE_CXX_STANDARD 17)
 endif()
 announce_configured_options(CMAKE_CXX_STANDARD)
 
+if(NOT CMAKE_SYSTEM_PROCESSOR)
+  set(CMAKE_SYSTEM_PROCESSOR ${CMAKE_HOST_SYSTEM_PROCESSOR})
+endif()
+announce_configured_options(CMAKE_SYSTEM_PROCESSOR)
+
 if(NOT CMAKE_BUILD_TYPE)
   set(CMAKE_BUILD_TYPE Debug)
 endif()
 announce_configured_options(CMAKE_BUILD_TYPE)
 
+if(NOT PYTHON_EXECUTABLE)
+  resolve_python_executable()
+endif()
+announce_configured_options(PYTHON_EXECUTABLE)
+
 announce_configured_options(CMAKE_CXX_COMPILER_ID)
 announce_configured_options(CMAKE_TOOLCHAIN_FILE)
 announce_configured_options(BUCK2)
-announce_configured_options(PYTHON_EXECUTABLE)
 
 load_build_preset()
 include(${PROJECT_SOURCE_DIR}/tools/cmake/preset/default.cmake)
@@ -72,10 +84,6 @@ print_configured_options()
 
 # MARK: - End EXECUTORCH_H12025_BUILD_MIGRATION ----------------------------------------------------
 
-include(tools/cmake/Utils.cmake)
-include(CMakeDependentOption)
-include(ExternalProject)
-
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
 # Setup RPATH.
@@ -251,11 +259,6 @@ if(EXECUTORCH_BUILD_TESTS)
   include(CTest)
 endif()
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-message(STATUS "Using python executable '${PYTHON_EXECUTABLE}'")
-
 # TODO(dbort): Fix these warnings and remove this flag.
 set(_common_compile_options -Wno-deprecated-declarations -fPIC)
 

backends/apple/mps/CMakeLists.txt

@@ -18,10 +18,6 @@ endif()
 
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 set(_common_compile_options -Wno-deprecated-declarations)
 set(_common_include_directories ${EXECUTORCH_ROOT}/..)
 

backends/cadence/CMakeLists.txt

@@ -30,10 +30,6 @@ add_compile_definitions(C10_USING_CUSTOM_GENERATED_MACROS)
 if(EXECUTORCH_CADENCE_CPU_RUNNER)
   include(${EXECUTORCH_ROOT}/tools/cmake/Codegen.cmake)
 
-  if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-  endif()
-
   set(_common_compile_options -Wno-deprecated-declarations -fPIC)
 
   # Find prebuilt libraries. executorch package should contain portable_ops_lib,

backends/cadence/aot/pass_utils.py

@@ -157,3 +157,34 @@ def nodes_not_adjacent_in_gm(
         if node.next.target == succ_target:
             return False
     return True
+
+
+def get_arg(
+    node: torch.fx.Node,
+    arg_index: int,
+    kwarg_name: str,
+    *,
+    default: torch.fx.node.Argument = None,
+) -> torch.fx.node.Argument:
+    """
+    Get the arg at arg_index or kwarg with arg_name of the node. If neither is found
+    return default.
+    """
+    if arg_index < len(node.args):
+        return node.args[arg_index]
+    elif kwarg_name in node.kwargs:
+        return node.kwargs[kwarg_name]
+    else:
+        return default
+
+
+def set_arg(
+    node: torch.fx.Node, arg_index: int, kwarg_name: str, value: torch.fx.node.Argument
+) -> None:
+    """
+    Set the arg at arg_index if it exists, otherwise set the kwarg.
+    """
+    if arg_index < len(node.args):
+        node.update_arg(arg_index, value)
+    else:
+        node.update_kwarg(kwarg_name, value)

backends/cadence/aot/remove_ops.py

@@ -25,7 +25,9 @@
 import torch.fx
 from executorch.backends.cadence.aot.pass_utils import (
     CadencePassAttribute,
+    get_arg,
     register_cadence_pass,
+    set_arg,
 )
 
 from executorch.backends.cadence.aot.simplify_ops import SimplifySliceOpPass
@@ -37,7 +39,7 @@
 from executorch.exir.pass_manager import PassManager, PassType
 from executorch.exir.passes import dead_code_elimination_pass
 from executorch.exir.passes.spec_prop_pass import SpecPropPass
-from torch.fx.node import Argument
+from torch.fx.node import Argument, Node
 
 
 @register_cadence_pass(CadencePassAttribute(opt_level=0))
@@ -771,65 +773,52 @@ def remove_branched(
 
 
 class RemoveCatFromSliceCopyPass(ExportPass):
-    def _remove_unused_cat(  # noqa: C901
-        self, graph_module: torch.fx.GraphModule
-    ) -> None:
-        slice_copy_nodes = [
-            node
-            for node in graph_module.graph.nodes
-            if node.target == exir_ops.edge.aten.slice_copy.Tensor
-        ]
-        for slice_copy_node in slice_copy_nodes:
-            slice_dim, start_idx, end_idx, step = 0, 0, float("inf"), 1
-            input_node, *other_args = slice_copy_node.args
-            if len(other_args) >= 1:
-                slice_dim = other_args[0]
-            if len(other_args) >= 2:
-                start_idx = other_args[1]
-            if len(other_args) >= 3:
-                end_idx = other_args[2]
-            if len(other_args) >= 4:
-                step = other_args[3]
-            if step != 1:
-                continue
-            slice_copy_dtype = slice_copy_node.meta["val"].dtype
-            if input_node.target != exir_ops.edge.aten.cat.default:
-                continue
-            cat_dtype = input_node.meta["val"].dtype
-            if slice_copy_dtype != cat_dtype:
+    """
+    Simplifies cat->slice_copy chains where one of the cat inputs can be directly passed
+    to the slice_copy.
+    """
+
+    def _remove_unused_cat(self, graph_module: torch.fx.GraphModule) -> None:
+        for slice_copy_node in graph_module.graph.find_nodes(
+            op="call_function", target=exir_ops.edge.aten.slice_copy.Tensor
+        ):
+            cat_node = cast(Node, get_arg(slice_copy_node, 0, "input"))
+            slice_dim = cast(int, get_arg(slice_copy_node, 1, "dim", default=0))
+            start_idx = cast(int, get_arg(slice_copy_node, 2, "start", default=None))
+            end_idx = cast(int, get_arg(slice_copy_node, 3, "end", default=None))
+            step = cast(int, get_arg(slice_copy_node, 4, "step", default=1))
+
+            if cat_node.target != exir_ops.edge.aten.cat.default or step != 1:
                 continue
-            cat_dim = input_node.args[1:]
-            if len(cat_dim) == 0:
-                cat_dim = 0
+
+            # Make sure cat and slice happens on the same dimension.
+            cat_dim = cast(Node, get_arg(cat_node, 1, "dim", default=0))
             if cat_dim != slice_dim:
                 continue
-            cat_output_shape = input_node.meta["val"].shape
-            start_idx = (
-                cat_output_shape[cat_dim] + start_idx if start_idx < 0 else start_idx
-            )
-            end_idx = (
-                cat_output_shape[cat_dim]
-                if end_idx > cat_output_shape[cat_dim]
-                else end_idx
-            )
-            base_idx = 0
-            cat_input_to_keep = None
-            for cat_input_node in input_node.args[0]:
-                cat_input_dtype = cat_input_node.meta["val"].dtype
-                if slice_copy_dtype != cat_input_dtype:
-                    continue
+
+            # Canonicalize slice indices.
+            cat_output_shape = cat_node.meta["val"].shape
+            if start_idx is None:
+                start_idx = 0
+            elif start_idx < 0:
+                start_idx += cat_output_shape[cat_dim]
+            if end_idx is None or end_idx > cat_output_shape[cat_dim]:
+                end_idx = cat_output_shape[cat_dim]
+            elif end_idx < 0:
+                end_idx += cat_output_shape[cat_dim]
+
+            offset = 0
+            for cat_input_node in cast(List[Node], get_arg(cat_node, 0, "tensors")):
                 cat_input_shape = cat_input_node.meta["val"].shape
 
-                # check if the slice range overlaps with the cat range
-                if (
-                    base_idx <= start_idx
-                    and end_idx <= list(cat_input_shape)[cat_dim] + base_idx
-                ):
-                    cat_input_to_keep = cat_input_node
+                # Check if the slice range overlaps with the cat input range.
+                if offset <= start_idx and end_idx <= offset + cat_input_shape[cat_dim]:
+                    slice_copy_node.replace_input_with(cat_node, cat_input_node)
+                    set_arg(slice_copy_node, 2, "start", start_idx - offset)
+                    set_arg(slice_copy_node, 3, "end", end_idx - offset)
                     break
-                base_idx += list(cat_input_shape)[cat_dim]
-            if cat_input_to_keep is not None:
-                slice_copy_node.replace_input_with(input_node, cat_input_to_keep)
+
+                offset += cat_input_shape[cat_dim]
 
     def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
         self._remove_unused_cat(graph_module)

backends/cadence/aot/tests/test_remove_ops_passes.py

@@ -864,3 +864,30 @@ def forward(self, x, y):
 
         # Ensure both cat nodes were removed
         self.assertEqual(count_node(graph_module, exir_ops.edge.aten.cat.default), 0)
+
+    def test_remove_cat_from_slice_copy_second_input(self) -> None:
+        builder = GraphBuilder()
+        x = builder.placeholder("x", torch.randn(2, 4))
+        y = builder.placeholder("y", torch.randn(2, 4))
+        cat = builder.call_operator(
+            op=exir_ops.edge.aten.cat.default,
+            args=((x, y), 1),
+        )
+        slice_copy = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(cat, 1, 5, 7, 1),
+        )
+        builder.output([slice_copy])
+        graph_module = builder.get_graph_module()
+
+        inputs = (torch.randn(2, 4), torch.randn(2, 4))
+        expected_outputs = graph_module(*inputs)[0]
+
+        p = RemoveCatFromSliceCopyPass()
+        graph_module = cast(PassResult, p(graph_module)).graph_module
+
+        # Cat should be removed.
+        self.assertEqual(count_node(graph_module, exir_ops.edge.aten.cat.default), 0)
+
+        # Output should remain the same.
+        self.assertTrue(torch.equal(graph_module(*inputs)[0], expected_outputs))

backends/cadence/fusion_g3/operators/CMakeLists.txt

@@ -14,10 +14,6 @@ endif()
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 include(${EXECUTORCH_ROOT}/tools/cmake/Codegen.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 # ATen compliant ops that are needed to run this model.
 set(_aten_ops__srcs
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/activation_ops_util.cpp"

backends/cadence/hifi/operators/CMakeLists.txt

@@ -14,10 +14,6 @@ endif()
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 include(${EXECUTORCH_ROOT}/tools/cmake/Codegen.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 # ATen compliant ops that are needed to run this model.
 set(_aten_ops__srcs
     "${EXECUTORCH_ROOT}/backends/cadence/hifi/operators/op_add.cpp"

backends/cadence/reference/operators/CMakeLists.txt

@@ -14,10 +14,6 @@ endif()
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 include(${EXECUTORCH_ROOT}/tools/cmake/Codegen.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 # ATen compliant ops that are needed to run this model.
 set(_aten_ops__srcs
     "${CMAKE_CURRENT_SOURCE_DIR}/op_add.cpp"

backends/cortex_m/CMakeLists.txt

@@ -23,10 +23,6 @@ endif()
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 include(${EXECUTORCH_ROOT}/tools/cmake/Codegen.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 # Cortex-M ops kernel sources
 set(_cortex_m_kernels__srcs
     ${CMAKE_CURRENT_SOURCE_DIR}/ops/op_quantize_per_tensor.cpp

backends/vulkan/CMakeLists.txt

@@ -24,10 +24,6 @@ if(NOT RUNTIME_PATH)
   set(RUNTIME_PATH ${CMAKE_CURRENT_SOURCE_DIR}/runtime)
 endif()
 
-if(NOT PYTHON_EXECUTABLE)
-  set(PYTHON_EXECUTABLE python3)
-endif()
-
 # Include this file to access target_link_options_shared_lib This is required to
 # provide access to target_link_options_shared_lib which allows libraries to be
 # linked with the --whole-archive flag. This is required for libraries that

backends/vulkan/test/op_tests/CMakeLists.txt

@@ -45,10 +45,6 @@ find_library(LIB_TORCH torch HINTS ${TORCH_INSTALL_PREFIX}/lib)
 find_library(LIB_TORCH_CPU torch_cpu HINTS ${TORCH_INSTALL_PREFIX}/lib)
 find_library(LIB_C10 c10 HINTS ${TORCH_INSTALL_PREFIX}/lib)
 
-if(NOT PYTHON_EXECUTABLE)
-  set(PYTHON_EXECUTABLE python3)
-endif()
-
 # Third party include paths
 
 set(VULKAN_THIRD_PARTY_PATH ${CMAKE_CURRENT_SOURCE_DIR}/../../third-party)

backends/xnnpack/CMakeLists.txt

@@ -25,10 +25,6 @@ endif()
 
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 # NB: Enabling this will serialize execution of delegate instances Keeping this
 # OFF by default to maintain existing behavior, to be revisited.
 option(EXECUTORCH_XNNPACK_SHARED_WORKSPACE

configurations/CMakeLists.txt

@@ -18,10 +18,6 @@ endif()
 
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 set(_common_compile_options -Wno-deprecated-declarations)
 
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)

examples/models/llama/README.md

@@ -295,6 +295,7 @@ Note for Mac users: There's a known linking issue with Xcode 15.1. Refer to the
         -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
         -DEXECUTORCH_BUILD_XNNPACK=ON \
         -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
+        -DSUPPORT_REGEX_LOOKAHEAD=ON
         -Bcmake-out/examples/models/llama \
         examples/models/llama
 
@@ -353,6 +354,7 @@ cmake  -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
     -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
     -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
     -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \
+    -DSUPPORT_REGEX_LOOKAHEAD=ON
     -Bcmake-out-android/examples/models/llama \
     examples/models/llama
 

examples/models/qwen3/README.md

@@ -86,3 +86,6 @@ cmake-out/examples/models/llama/llama_main
 ```
 
 To run the model on an example iOS or Android app, see the Llama README's [Step 5: Build Mobile apps](../llama/README.md#step-5-build-mobile-apps) section.
+
+### FAQ
+For more help with exporting or running this model, feel free to ask in our [discord channel](https://lnkd.in/gWCM4ViK).

extension/flat_tensor/test/CMakeLists.txt

@@ -22,7 +22,7 @@ add_custom_command(
   OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/ModuleAddMulProgram.pte"
          "${CMAKE_CURRENT_BINARY_DIR}/ModuleAddMulProgram.ptd"
   COMMAND
-    python -m test.models.export_program --modules "ModuleAddMul"
+    ${PYTHON_EXECUTABLE} -m test.models.export_program --modules "ModuleAddMul"
     --external-constants --outdir "${CMAKE_CURRENT_BINARY_DIR}" 2> /dev/null
   WORKING_DIRECTORY ${EXECUTORCH_ROOT}
 )