Update on "Add data path to runner"

^

Differential Revision: [D73617661](https://our.internmc.facebook.com/intern/diff/D73617661/)

[ghstack-poisoned]

Author: lucylq

CMakeLists.txt

@@ -410,20 +410,6 @@ endif()
 message(STATUS "executorch: Using sources file ${EXECUTORCH_SRCS_FILE}")
 include(${EXECUTORCH_SRCS_FILE})
 
-#
-# Modify default options when cross-compiling.
-#
-# The intent is for the EXECUTORCH_BUILD_HOST_TARGETS option to affect the
-# default ON/OFF values of host targets around the tree. This way, a user can
-# disable EXECUTORCH_BUILD_HOST_TARGETS to disable all host targets, and then
-# optionally re-enable some of those targets. Or they could leave
-# EXECUTORCH_BUILD_HOST_TARGETS enabled and then optionally disable any given
-# host target.
-#
-# We can then use various cross-compilation hints to set the default value of
-# EXECUTORCH_BUILD_HOST_TARGETS, which can still be overridden if desired.
-#
-
 # Detect if an iOS toolchain is set.
 if(CMAKE_TOOLCHAIN_FILE MATCHES ".*(iOS|ios\.toolchain)\.cmake$")
   set(CMAKE_TOOLCHAIN_IOS ON)
@@ -458,16 +444,6 @@ if(EXECUTORCH_USE_CPP_CODE_COVERAGE)
   endif()
 endif()
 
-# EXECUTORCH_BUILD_HOST_TARGETS: Option to control the building of host-only
-# tools like `flatc`, along with example executables like `executor_runner` and
-# libraries that it uses, like `gflags`. Disabling this can be helpful when
-# cross-compiling, but some required tools that would have been built need to be
-# provided directly.
-cmake_dependent_option(
-  EXECUTORCH_BUILD_HOST_TARGETS "Build host-only targets." ON
-  "NOT CMAKE_TOOLCHAIN_IOS" OFF
-)
-
 #
 # flatc: Flatbuffer commandline tool to generate .h files from .fbs files
 #
@@ -665,10 +641,7 @@ add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/configurations)
 #
 # gflags: Commandline flag host library.
 #
-cmake_dependent_option(
-  EXECUTORCH_BUILD_GFLAGS "Build the gflags library." ON
-  EXECUTORCH_BUILD_HOST_TARGETS OFF
-)
+option(EXECUTORCH_BUILD_GFLAGS "Build the gflags library." ON)
 if(EXECUTORCH_BUILD_GFLAGS)
   add_subdirectory(third-party/gflags)
 endif()
@@ -694,7 +667,7 @@ install(FILES tools/cmake/executorch-config.cmake DESTINATION lib/cmake/ExecuTor
 #
 cmake_dependent_option(
   EXECUTORCH_BUILD_EXECUTOR_RUNNER "Build the executor_runner executable" ON
-  EXECUTORCH_BUILD_HOST_TARGETS OFF
+  "NOT CMAKE_TOOLCHAIN_IOS" OFF
 )
 
 # Add googletest if any test targets should be built

backends/arm/quantizer/TARGETS

@@ -1,5 +1,15 @@
 load("@fbcode_macros//build_defs:python_library.bzl", "python_library")
 
+# Exposed through __init__.py
+python_library(
+    name = "quantization_config",
+    srcs = ["quantization_config.py"],
+    deps = [
+        "//caffe2:torch",
+    ],
+)
+
+# Exposed through __init__.py
 python_library(
     name = "arm_quantizer",
     srcs = ["arm_quantizer.py"],
@@ -22,17 +32,19 @@ python_library(
 )
 
 python_library(
-    name = "quantization_config",
-    srcs = ["quantization_config.py"],
+    name = "arm_quantizer_utils",
+    srcs = ["arm_quantizer_utils.py"],
     deps = [
-        "//caffe2:torch",
+        ":quantization_config",
     ],
 )
 
 python_library(
-    name = "arm_quantizer_utils",
-    srcs = ["arm_quantizer_utils.py"],
+    name = "lib",
+    srcs = ["__init__.py"],
     deps = [
+        ":arm_quantizer",
         ":quantization_config",
-    ],
+        ":arm_quantizer_utils",
+    ]
 )

backends/arm/quantizer/__init__.py

@@ -1,4 +1,15 @@
-# Copyright 2024 Arm Limited and/or its affiliates.
+# Copyright 2024-2025 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
+
+
+from .quantization_config import QuantizationConfig  # noqa  # usort: skip
+from .arm_quantizer import (  # noqa
+    EthosUQuantizer,
+    get_symmetric_quantization_config,
+    TOSAQuantizer,
+)
+
+# Used in tests
+from .arm_quantizer_utils import is_annotated  # noqa

backends/arm/quantizer/arm_quantizer.py

@@ -19,16 +19,11 @@
 import torch
 from executorch.backends.arm._passes import ArmPassManager
 
-from executorch.backends.arm.quantizer import arm_quantizer_utils
-from executorch.backends.arm.quantizer.arm_quantizer_utils import (  # type: ignore[attr-defined]
-    mark_node_as_annotated,
-)
-from executorch.backends.arm.quantizer.quantization_annotator import (  # type: ignore[import-not-found]
-    annotate_graph,
-)
-
-from executorch.backends.arm.quantizer.quantization_config import QuantizationConfig
+from executorch.backends.arm.quantizer import QuantizationConfig
 from executorch.backends.arm.tosa_specification import TosaSpecification
+
+from .arm_quantizer_utils import is_annotated, mark_node_as_annotated
+from .quantization_annotator import annotate_graph
 from executorch.backends.arm.arm_backend import (
     get_tosa_spec,
     is_ethosu,
@@ -337,7 +332,7 @@ def _annotate_io(
         quantization_config: QuantizationConfig,
     ):
         for node in model.graph.nodes:
-            if arm_quantizer_utils.is_annotated(node):
+            if is_annotated(node):
                 continue
             if node.op == "placeholder" and len(node.users) > 0:
                 _annotate_output_qspec(

backends/arm/quantizer/quantization_annotator.py

@@ -10,8 +10,7 @@
 
 import torch
 import torch.fx
-from executorch.backends.arm.quantizer import arm_quantizer_utils
-from executorch.backends.arm.quantizer.quantization_config import QuantizationConfig
+from executorch.backends.arm.quantizer import QuantizationConfig
 from executorch.backends.arm.tosa_utils import get_node_debug_info
 from torch.ao.quantization.quantizer import QuantizationSpecBase, SharedQuantizationSpec
 from torch.ao.quantization.quantizer.utils import (
@@ -20,6 +19,13 @@
 )
 from torch.fx import Node
 
+from .arm_quantizer_utils import (
+    is_annotated,
+    is_ok_for_quantization,
+    is_output_annotated,
+    mark_node_as_annotated,
+)
+
 logger = logging.getLogger(__name__)
 
 
@@ -69,7 +75,7 @@ def _is_ok_for_quantization(
     """
     # Check output
     if quant_properties.quant_output is not None:
-        if not arm_quantizer_utils.is_ok_for_quantization(node, gm):  # type: ignore[attr-defined]
+        if not is_ok_for_quantization(node, gm):  # type: ignore[attr-defined]
             logger.debug(
                 f"Could not quantize node due to output: "
                 f"{get_node_debug_info(node, gm)}"
@@ -87,7 +93,7 @@ def _is_ok_for_quantization(
 
         for n_arg in _as_list(node.args[quant_property.index]):
             assert isinstance(n_arg, Node)
-            if not arm_quantizer_utils.is_ok_for_quantization(n_arg, gm):  # type: ignore[attr-defined]
+            if not is_ok_for_quantization(n_arg, gm):  # type: ignore[attr-defined]
                 logger.debug(
                     f'could not quantize node due to input "{node}": '
                     f"{get_node_debug_info(node, gm)}"
@@ -99,7 +105,7 @@ def _is_ok_for_quantization(
 
 
 def _annotate_input(node: Node, quant_property: _QuantProperty):
-    assert not arm_quantizer_utils.is_annotated(node)
+    assert not is_annotated(node)
     if quant_property.optional and (
         quant_property.index >= len(node.args)
         or node.args[quant_property.index] is None
@@ -114,11 +120,11 @@ def _annotate_input(node: Node, quant_property: _QuantProperty):
         assert isinstance(n_arg, Node)
         _annotate_input_qspec_map(node, n_arg, qspec)
         if quant_property.mark_annotated:
-            arm_quantizer_utils.mark_node_as_annotated(n_arg)  # type: ignore[attr-defined]
+            mark_node_as_annotated(n_arg)  # type: ignore[attr-defined]
 
 
 def _annotate_output(node: Node, quant_property: _QuantProperty):
-    assert not arm_quantizer_utils.is_annotated(node)
+    assert not is_annotated(node)
     assert not quant_property.mark_annotated
     assert not quant_property.optional
     assert quant_property.index == 0, "Only one output annotation supported currently"
@@ -343,7 +349,7 @@ def any_or_hardtanh_min_zero(n: Node):
     elif node.target in _one_to_one_shared_input_or_input_act_qspec:
         input_qspec = (
             SharedQuantizationSpec(node.args[0])  # type: ignore[arg-type]
-            if arm_quantizer_utils.is_output_annotated(node.args[0])  # type: ignore
+            if is_output_annotated(node.args[0])  # type: ignore
             else input_act_qspec
         )
         quant_properties.quant_inputs = [_QuantProperty(0, input_qspec)]  # type: ignore[arg-type]
@@ -396,7 +402,7 @@ def any_or_hardtanh_min_zero(n: Node):
         if not isinstance(node.args[0], Node):
             return None
 
-        if not arm_quantizer_utils.is_output_annotated(node.args[0]):  # type: ignore[attr-defined]
+        if not is_output_annotated(node.args[0]):  # type: ignore[attr-defined]
             return None
 
         shared_qspec = SharedQuantizationSpec(node.args[0])
@@ -426,7 +432,7 @@ def annotate_graph(  # type: ignore[return]
         if node.op != "call_function":
             continue
 
-        if arm_quantizer_utils.is_annotated(node):
+        if is_annotated(node):
             continue
 
         if filter_fn is not None and not filter_fn(node):
@@ -442,7 +448,7 @@ def annotate_graph(  # type: ignore[return]
         if quant_properties.quant_output is not None:
             _annotate_output(node, quant_properties.quant_output)
 
-        arm_quantizer_utils.mark_node_as_annotated(node)  # type: ignore[attr-defined]
+        mark_node_as_annotated(node)  # type: ignore[attr-defined]
 
         # Quantization does not allow kwargs for some reason.
         # Remove from ops we know have and where we know it does not break anything.

backends/arm/test/TARGETS

@@ -42,7 +42,7 @@ python_library(
         ":common",
         "//executorch/backends/xnnpack/test/tester:tester",
         "//executorch/backends/arm:arm_partitioner",
-        "//executorch/backends/arm/quantizer:arm_quantizer",
+        "//executorch/backends/arm/quantizer:lib",
         "//executorch/backends/arm:tosa_mapping",
         "//executorch/devtools/backend_debug:delegation_info",
         "fbsource//third-party/pypi/tabulate:tabulate",

backends/arm/test/models/test_deit_tiny_arm.py

@@ -0,0 +1,44 @@
+# Copyright 2025 Arm Limited and/or its affiliates.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+
+from typing import Tuple
+
+import timm
+
+import torch
+
+from executorch.backends.arm.test.tester.test_pipeline import TosaPipelineMI
+
+from timm.data import IMAGENET_INCEPTION_MEAN, IMAGENET_INCEPTION_STD
+from torchvision import transforms
+
+logger = logging.getLogger(__name__)
+logger.setLevel(logging.INFO)
+
+
+deit_tiny = timm.models.deit.deit_tiny_patch16_224(pretrained=True)
+deit_tiny.eval()
+
+normalize = transforms.Normalize(
+    mean=IMAGENET_INCEPTION_MEAN, std=IMAGENET_INCEPTION_STD
+)
+model_inputs = (normalize(torch.rand((1, 3, 224, 224))),)
+
+input_t = Tuple[torch.Tensor]
+
+
+def test_deit_tiny_tosa_MI():
+    pipeline = TosaPipelineMI[input_t](
+        deit_tiny,
+        model_inputs,
+        aten_op=[],
+        exir_op=[],
+        use_to_edge_transform_and_lower=True,
+        atol=6.5,  # This needs to go down: MLETORCH-940
+        qtol=1,
+    )
+    pipeline.run()

backends/arm/test/ops/test_expand.py

@@ -15,7 +15,7 @@
 
 import torch
 
-from executorch.backends.arm.quantizer.arm_quantizer import (
+from executorch.backends.arm.quantizer import (
     EthosUQuantizer,
     get_symmetric_quantization_config,
     TOSAQuantizer,

backends/arm/test/ops/test_hardtanh.py

@@ -13,7 +13,7 @@
 
 import torch
 
-from executorch.backends.arm.quantizer.arm_quantizer import (
+from executorch.backends.arm.quantizer import (
     EthosUQuantizer,
     get_symmetric_quantization_config,
     TOSAQuantizer,

backends/arm/test/ops/test_max_pool.py

@@ -12,7 +12,7 @@
 import pytest
 
 import torch
-from executorch.backends.arm.quantizer.arm_quantizer import (
+from executorch.backends.arm.quantizer import (
     EthosUQuantizer,
     get_symmetric_quantization_config,
     TOSAQuantizer,

backends/arm/test/ops/test_permute.py

@@ -13,7 +13,7 @@
 
 import torch
 
-from executorch.backends.arm.quantizer.arm_quantizer import (
+from executorch.backends.arm.quantizer import (
     EthosUQuantizer,
     get_symmetric_quantization_config,
     TOSAQuantizer,

backends/arm/test/ops/test_relu.py

@@ -10,7 +10,7 @@
 from typing import Tuple
 
 import torch
-from executorch.backends.arm.quantizer.arm_quantizer import (
+from executorch.backends.arm.quantizer import (
     EthosUQuantizer,
     get_symmetric_quantization_config,
     TOSAQuantizer,

backends/arm/test/ops/test_repeat.py

@@ -13,7 +13,7 @@
 
 import torch
 
-from executorch.backends.arm.quantizer.arm_quantizer import (
+from executorch.backends.arm.quantizer import (
     EthosUQuantizer,
     get_symmetric_quantization_config,
     TOSAQuantizer,

backends/arm/test/ops/test_sigmoid_16bit.py

@@ -6,7 +6,7 @@
 import pytest
 
 import torch
-from executorch.backends.arm.quantizer.arm_quantizer import (
+from executorch.backends.arm.quantizer import (
     get_symmetric_quantization_config,
     TOSAQuantizer,
 )

backends/arm/test/ops/test_sigmoid_32bit.py

@@ -5,7 +5,7 @@
 
 import pytest
 import torch
-from executorch.backends.arm.quantizer.arm_quantizer import TOSAQuantizer
+from executorch.backends.arm.quantizer import TOSAQuantizer
 from executorch.backends.arm.quantizer.quantization_config import QuantizationConfig
 from executorch.backends.arm.test import common
 from executorch.backends.arm.test.tester.test_pipeline import (

backends/arm/test/ops/test_var.py

@@ -11,7 +11,7 @@
 import unittest
 
 import torch
-from executorch.backends.arm.quantizer.arm_quantizer import (
+from executorch.backends.arm.quantizer import (
     EthosUQuantizer,
     get_symmetric_quantization_config,
     TOSAQuantizer,

backends/arm/test/ops/test_where.py

@@ -9,7 +9,7 @@
 
 import torch
 
-from executorch.backends.arm.quantizer.arm_quantizer import (
+from executorch.backends.arm.quantizer import (
     EthosUQuantizer,
     get_symmetric_quantization_config,
     TOSAQuantizer,