Merge branch 'main' into gh/manuelcandales/128/orig

Author: manuelcandales

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-5616fa4a68718ead203314a3467f7dd9547153ae
+9b498d3bb28b8e3411ce464dd2755c5b96d92c8f

.ci/docker/common/install_conda.sh

@@ -13,7 +13,7 @@ source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 install_miniconda() {
   BASE_URL="https://repo.anaconda.com/miniconda"
   CONDA_FILE="Miniconda3-py${PYTHON_VERSION//./}_${MINICONDA_VERSION}-Linux-x86_64.sh"
-  if [[ $(uname -m) == "aarch64" ]]; then 
+  if [[ $(uname -m) == "aarch64" ]]; then
     CONDA_FILE="Miniconda3-py${PYTHON_VERSION//./}_${MINICONDA_VERSION}-Linux-aarch64.sh"
   fi
 
@@ -71,4 +71,8 @@ fix_conda_ubuntu_libstdcxx() {
 install_miniconda
 install_python
 install_pip_dependencies
-fix_conda_ubuntu_libstdcxx
+# Hack breaks the job on aarch64 but is still necessary everywhere
+# else.
+if [ "$(uname -m)" != "aarch64" ]; then
+    fix_conda_ubuntu_libstdcxx
+fi

.ci/docker/conda-env-ci.txt

@@ -1,4 +1,4 @@
-cmake=3.26.4
+cmake=3.31.2
 ninja=1.10.2
 libuv
 llvm-openmp

.ci/scripts/check_c10_sync.sh

@@ -12,3 +12,4 @@ pushd pytorch
 git checkout "$pytorch_pin"
 popd
 "$(dirname "${BASH_SOURCE[0]}")"/compare_dirs.sh runtime/core/portable_type/c10/c10 pytorch/c10
+"$(dirname "${BASH_SOURCE[0]}")"/compare_dirs.sh runtime/core/portable_type/c10/torch/standalone pytorch/torch/standalone

CMakeLists.txt

@@ -705,7 +705,7 @@ if(EXECUTORCH_BUILD_EXECUTOR_RUNNER)
   endif()
 
   add_executable(executor_runner ${_executor_runner__srcs})
-  if(CMAKE_BUILD_TYPE STREQUAL "Release")
+  if(NOT CMAKE_BUILD_TYPE STREQUAL "Debug")
     if(APPLE)
       target_link_options(executor_runner PRIVATE "LINKER:-dead_strip")
     else()

backends/arm/_passes/decompose_sqrt_pass.py

@@ -4,6 +4,8 @@
 # LICENSE file in the root directory of this source tree.
 
 # pyre-unsafe
+from typing import Tuple, Union
+
 import torch
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass
@@ -15,7 +17,7 @@
 )
 
 
-def get_sqrt_decomposition(op) -> tuple:
+def get_sqrt_decomposition(op) -> Union[Tuple, torch._ops.OpOverload]:
     # TODO : "MLETORCH-863 : Replace current sqrt -> pow.Tensor_Scalar workaround with pow.Tensor_Tensor"
     if op in edge_sqrt_ops:
         return exir_ops.edge.aten.pow.Tensor_Scalar

backends/arm/_passes/replace_scalar_with_tensor_pass.py

@@ -6,7 +6,7 @@
 # pyre-unsafe
 
 
-from typing import Dict
+from typing import Dict, Union
 
 import torch
 from executorch.backends.transforms.replace_scalar_with_tensor import (
@@ -18,7 +18,10 @@
 
 
 # Operators that are included for both TOSA profiles
-_common_ops: Dict[EdgeOpOverload, EdgeOpOverload] = {
+_common_ops: Dict[
+    Union[EdgeOpOverload, torch._ops.OpOverload],
+    Union[EdgeOpOverload, torch._ops.OpOverload],
+] = {
     exir_ops.edge.aten.add.Scalar: exir_ops.edge.aten.add.Tensor,
     exir_ops.edge.aten.sub.Scalar: exir_ops.edge.aten.sub.Tensor,
     exir_ops.edge.aten.mul.Scalar: exir_ops.edge.aten.mul.Tensor,

backends/arm/operators/op_index_tensor.py

@@ -189,11 +189,16 @@ def define_node(
             if i == 0:
                 gather_index_name = reshaped_idxs.name
             else:
+                add_idxs = tosa_graph.addIntermediate(
+                    reshaped_idxs.shape,
+                    reshaped_idxs.dtype,
+                )
                 tosa_graph.addOperator(
                     ts.TosaOp.Op().ADD,
                     [gather_index_name, reshaped_idxs.name],
-                    [gather_index_name],
+                    [add_idxs.name],
                 )
+                gather_index_name = add_idxs.name
 
         gather_vals_shape = [N, K, C]
         reshaped_input = tosa_graph.addIntermediate(gather_vals_shape, values.dtype)
@@ -314,11 +319,16 @@ def define_node(
             if i == 0:
                 gather_index_name = reshaped_idxs.name
             else:
+                add_idxs = tosa_graph.addIntermediate(
+                    reshaped_idxs.shape,
+                    reshaped_idxs.dtype,
+                )
                 tosa_graph.addOperator(
                     ts.TosaOp.Op().ADD,
                     [gather_index_name, reshaped_idxs.name],
-                    [gather_index_name],
+                    [add_idxs.name],
                 )
+                gather_index_name = add_idxs.name
 
         gather_vals_shape = [N, K, C]
         reshaped_input = tosa_graph.addIntermediate(gather_vals_shape, values.dtype)

backends/arm/test/test_model.py

@@ -64,7 +64,7 @@ def get_args():
     parser.add_argument(
         "--timeout",
         required=False,
-        default=60 * 10,
+        default=60 * 20,
         help="Timeout in seconds used when running the model",
     )
     args = parser.parse_args()
@@ -165,11 +165,6 @@ def build_ethosu_runtime(
     extra_flags: str,
     elf_build_path: str,
 ):
-
-    extra_build_flag = ""
-    if extra_flags:
-        extra_build_flag = f"--extra_build_flags={extra_flags}"
-
     run_external_cmd(
         [
             "bash",
@@ -182,7 +177,7 @@ def build_ethosu_runtime(
             "--build_type=Release",
             f"--system_config={system_config}",
             f"--memory_mode={memory_mode}",
-            extra_build_flag,
+            f"--extra_build_flags=-DET_DUMP_OUTPUT=OFF {extra_flags}",
             f"--output={elf_build_path}",
         ]
     )

backends/cadence/aot/memory_constraints.py

@@ -11,7 +11,7 @@
 import typing
 from collections import defaultdict
 from dataclasses import dataclass
-from typing import cast, DefaultDict, Iterable, Optional, Sequence
+from typing import Callable, cast, DefaultDict, Iterable, Optional, Sequence, TypeAlias
 
 import torch
 import torch.fx
@@ -573,23 +573,34 @@ def compute_slice_and_select_loc_constraints(
         graph_module.recompile()
 
 
+ConstraintsGenPass: TypeAlias = Callable[
+    [MemConstraints],
+    Callable[[torch.fx.GraphModule], Optional[PassResult]],
+]
+
+
 # The class to generate all the constraints that will be passed on to the memory
 # planning algorithm.
 class GenerateMemConstraints:
     def __init__(
         self,
         mem_constraints: MemConstraints,
-        additional_constraint_gen_passes: list | None = None,
+        additional_constraint_gen_passes: Sequence[ConstraintsGenPass] | None = None,
     ) -> None:
-        self.mem_constraints = mem_constraints
-        self.additional_constraint_gen_passes = additional_constraint_gen_passes or []
+        self.mem_constraints: MemConstraints = mem_constraints
+        self.additional_constraint_gen_passes: Sequence[ConstraintsGenPass] = (
+            additional_constraint_gen_passes or []
+        )
 
     def __call__(self, graph_module: torch.fx.GraphModule) -> PassResult:
-        constraint_gen_passes: list = [
-            GenerateMemoryViewConstraints,
-            GenerateSliceAndSelectNopConstraints,
-            GenerateCatNopConstraints,
-        ] + self.additional_constraint_gen_passes
+        constraint_gen_passes: Sequence[ConstraintsGenPass] = cast(
+            list[ConstraintsGenPass],
+            [
+                GenerateMemoryViewConstraints,
+                GenerateSliceAndSelectNopConstraints,
+                GenerateCatNopConstraints,
+            ],
+        ) + list(self.additional_constraint_gen_passes)
         # Create a filter using the opt level in mem_constraints, and filter
         # the relevant passes.
         pass_filter = create_cadence_pass_filter(self.mem_constraints.opt_level)
@@ -602,6 +613,7 @@ def __call__(self, graph_module: torch.fx.GraphModule) -> PassResult:
                         typing.Callable[[torch.fx.GraphModule], Optional[PassResult]],
                     ]
                 ],
+                # pyre-ignore[6]: Incompatible parameter type.
                 list(filter(pass_filter, constraint_gen_passes)),
             )
         ]