Update base for Update on "Use c10 version of half/bfloat16 in executorch"

Accomplished by importing relevant files from c10 into
executorch/runtime/core/portable_type/c10, and then using `using` in
the top-level ExecuTorch headers. This approach should keep the
ExecuTorch build hermetic for embedded use cases. In the future, we
should add a CI job to ensure the c10 files stay identical to the
PyTorch ones.

Differential Revision: [D66106969](https://our.internmc.facebook.com/intern/diff/D66106969/)

[ghstack-poisoned]

Author: swolchok

.github/workflows/android-perf.yml

@@ -98,6 +98,7 @@ jobs:
       - uses: actions/checkout@v3
 
       - name: Prepare the spec
+        id: prepare
         shell: bash
         env:
           BENCHMARK_CONFIG: ${{ toJSON(matrix) }}
@@ -111,7 +112,7 @@ jobs:
           # so let's just sed it
           sed -i -e 's,{{ model_path }},'"${MODEL_PATH}"',g' android-llm-device-farm-test-spec.yml.j2
 
-          BENCHMARK_CONFIG_ID="${{ matrix.model }}_${{ matrix.config }}"
+          BENCHMARK_CONFIG_ID=$(echo "${{ matrix.model }}_${{ matrix.config }}" | sed -e 's/[^A-Za-z0-9._-]/_/g')
           # The config for this benchmark runs, we save it in the test spec so that it can be fetched
           # later by the upload script
           sed -i -e 's,{{ benchmark_config_id }},'"${BENCHMARK_CONFIG_ID}"',g' android-llm-device-farm-test-spec.yml.j2
@@ -122,6 +123,7 @@ jobs:
 
           # Save the benchmark configs so that we can use it later in the dashboard
           echo "${BENCHMARK_CONFIG}" > "${BENCHMARK_CONFIG_ID}.json"
+          echo "benchmark-config-id=${BENCHMARK_CONFIG_ID}" >> $GITHUB_OUTPUT
 
       - name: Upload the spec
         uses: seemethere/upload-artifact-s3@v5
@@ -141,7 +143,7 @@ jobs:
             ${{ github.repository }}/${{ github.run_id }}/artifacts/benchmark-configs/
           retention-days: 1
           if-no-files-found: error
-          path: extension/benchmark/android/benchmark/${{ matrix.model }}_${{ matrix.config }}.json
+          path: extension/benchmark/android/benchmark/${{ steps.prepare.outputs.benchmark-config-id }}.json
 
   export-models:
     name: export-models

.github/workflows/apple-perf.yml

@@ -100,6 +100,7 @@ jobs:
       - uses: actions/checkout@v3
 
       - name: Prepare the spec
+        id: prepare
         shell: bash
         env:
           BENCHMARK_CONFIG: ${{ toJSON(matrix) }}
@@ -113,7 +114,7 @@ jobs:
           # so let's just sed it
           sed -i -e 's,{{ model_path }},'"${MODEL_PATH}"',g' default-ios-device-farm-appium-test-spec.yml.j2
 
-          BENCHMARK_CONFIG_ID="${{ matrix.model }}_${{ matrix.config }}"
+          BENCHMARK_CONFIG_ID=$(echo "${{ matrix.model }}_${{ matrix.config }}" | sed -e 's/[^A-Za-z0-9._-]/_/g')
           # The config for this benchmark runs, we save it in the test spec so that it can be fetched
           # later by the upload script
           sed -i -e 's,{{ benchmark_config_id }},'"${BENCHMARK_CONFIG_ID}"',g' default-ios-device-farm-appium-test-spec.yml.j2
@@ -124,6 +125,7 @@ jobs:
 
           # Save the benchmark configs so that we can use it later in the dashboard
           echo "${BENCHMARK_CONFIG}" > "${BENCHMARK_CONFIG_ID}.json"
+          echo "benchmark-config-id=${BENCHMARK_CONFIG_ID}" >> $GITHUB_OUTPUT
 
       - name: Upload the spec
         uses: seemethere/upload-artifact-s3@v5
@@ -143,7 +145,7 @@ jobs:
             ${{ github.repository }}/${{ github.run_id }}/artifacts/benchmark-configs/
           retention-days: 1
           if-no-files-found: error
-          path: extension/benchmark/apple/Benchmark/${{ matrix.model }}_${{ matrix.config }}.json
+          path: extension/benchmark/apple/Benchmark/${{ steps.prepare.outputs.benchmark-config-id }}.json
 
   export-models:
     name: export-models

.github/workflows/lint.yml

@@ -20,13 +20,19 @@ jobs:
     with:
       runner: linux.2xlarge
       docker-image: executorch-ubuntu-22.04-linter
+      submodules: 'true'
       fetch-depth: 0
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       timeout: 90
       script: |
         # The generic Linux job chooses to use base env, not the one setup by the image
         CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
         conda activate "${CONDA_ENV}"
+        
+        # For mypy linting, we need to first install executorch first so that
+        # it builds the python package information.
+        BUILD_TOOL="cmake"
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh "${BUILD_TOOL}"
 
         CACHE_DIRECTORY="/tmp/.lintbin"
         # Try to recover the cached binaries

.gitignore

@@ -1,5 +1,6 @@
 .hypothesis
 buck-out/
+.mypy_cache/
 buck2-bin/
 cmake-out*
 .DS_Store

.lintrunner.toml

@@ -285,3 +285,49 @@ command = [
     '--',
     '@{{PATHSFILE}}',
 ]
+
+[[linter]]
+code = 'MYPY'
+include_patterns = [
+    # TODO(https://github.com/pytorch/executorch/issues/7441): Gradually start enabling all folders.
+    # 'backends/**/*.py',
+    'build/**/*.py',
+    'codegen/**/*.py',
+    # 'devtools/**/*.py',
+    'docs/**/*.py',
+    # 'examples/**/*.py',
+    # 'exir/**/*.py',
+    # 'extension/**/*.py',
+    'kernels/**/*.py',
+    # 'profiler/**/*.py',
+    'runtime/**/*.py',
+    'scripts/**/*.py',
+    # 'test/**/*.py',
+    # 'util/**/*.py',
+    '*.py',
+]
+exclude_patterns = [
+    'third-party/**',
+    '**/third-party/**',
+    'scripts/check_binary_dependencies.py',
+]
+command = [
+    'python',
+    '-m',
+    'lintrunner_adapters',
+    'run',
+    'mypy_linter',
+    '--config=.mypy.ini',
+    '--show-disable',
+    '--',
+    '@{{PATHSFILE}}'
+]
+init_command = [
+    'python',
+    '-m',
+    'lintrunner_adapters',
+    'run',
+    'pip_init',
+    '--dry-run={{DRYRUN}}',
+    '--requirement=requirements-lintrunner.txt',
+]

.mypy.ini

@@ -0,0 +1,80 @@
+[mypy]
+allow_redefinition = True
+warn_unused_configs = True
+warn_redundant_casts = True
+show_error_codes = True
+show_column_numbers = True
+disallow_untyped_decorators = True
+follow_imports = normal
+local_partial_types = True
+enable_error_code = possibly-undefined
+warn_unused_ignores = False
+
+files =
+    backends,
+    codegen,
+    devtools
+    examples,
+    exir,
+    extension,
+    kernels,
+    profiler,
+    runtime,
+    scripts,
+    util
+
+mypy_path = executorch
+
+[mypy-executorch.codegen.*]
+follow_untyped_imports = True
+
+[mypy-executorch.devtools.*]
+follow_untyped_imports = True
+
+[mypy-executorch.exir.*]
+follow_untyped_imports = True
+
+[mypy-executorch.extension.*]
+follow_untyped_imports = True
+
+[mypy-executorch.kernels.*]
+follow_untyped_imports = True
+
+[mypy-executorch.runtime.*]
+follow_untyped_imports = True
+
+[mypy-requests.*]
+follow_untyped_imports = True
+
+[mypy-torchgen.*]
+follow_untyped_imports = True
+
+[mypy-buck_util]
+ignore_missing_imports = True
+
+[mypy-docutils.*]
+ignore_missing_imports = True
+
+[mypy-pandas]
+ignore_missing_imports = True
+
+[mypy-pytorch_sphinx_theme]
+ignore_missing_imports = True
+
+[mypy-ruamel]
+ignore_missing_imports = True
+
+[mypy-setuptools.*]
+ignore_missing_imports = True
+
+[mypy-sphinx.*]
+ignore_missing_imports = True
+
+[mypy-tomllib]
+ignore_missing_imports = True
+
+[mypy-yaml]
+ignore_missing_imports = True
+
+[mypy-zstd]
+ignore_missing_imports = True

backends/cadence/aot/compiler.py

@@ -18,17 +18,11 @@
 )
 from executorch.backends.cadence.aot.quantizer.fusion_pass import QuantFusion
 from executorch.backends.cadence.aot.quantizer.quantizer import CadenceQuantizer
-
-from executorch.backends.cadence.aot.replace_ops import ReplaceSafeSoftmaxWithSoftmax
 from executorch.backends.cadence.aot.utils import (
     get_default_memory_config,
     MemoryConfig,
-    model_gm_has_SDPA,
     model_is_quantized,
 )
-from executorch.backends.transforms.decompose_sdpa import (
-    DecomposeScaledDotProductAttention,
-)
 from executorch.devtools import generate_etrecord
 from executorch.exir import (
     EdgeCompileConfig,
@@ -91,16 +85,6 @@ def convert_pt2(
         .module()
     )
 
-    if model_gm_has_SDPA(model_gm):
-        # Decompose SDPA
-        DecomposeScaledDotProductAttention(False)(model_gm)
-
-        # Swap _safe_softmax with _softmax (see https://github.com/pytorch/pytorch/pull/133882
-        # for details).
-        result = ReplaceSafeSoftmaxWithSoftmax()(model_gm)
-        assert result is not None
-        model_gm = result.graph_module
-
     # Prepare
     prepared_model = prepare_pt2e(model_gm, quantizer)
 

backends/cadence/aot/compiler_utils.py

@@ -129,16 +129,16 @@ def get_transposed_dims(node: torch.fx.Node, dims: List[int]) -> List[int]:
 
 
 # Capture the effect of permute op on incoming dimension order
-def get_permuted_dims(node: torch.fx.Node, dims: Optional[List[int]]) -> List[int]:
+def get_permuted_dims(node: torch.fx.Node, dims: Optional[Sequence[int]]) -> List[int]:
     """
     Given a permute node, and the incoming dimension ordering of the input
     tensor to the permute node, return the net effect of permute op on the
     dimension order.
     """
     assert node.target == exir_ops.edge.aten.permute_copy.default
     # Permute each index of the dimension ordering (dims)
-    permute_dims = node.args[1]
-    assert isinstance(permute_dims, List)
+    # pyre-fixme[6]: This combined typecheck isn't supported yet.
+    permute_dims: List[int] = list(node.args[1])
     assert all(isinstance(x, int) for x in permute_dims)
     # If the dims is empty, we can simply return the permute order
     if not dims:

backends/cadence/aot/reorder_ops.py

@@ -438,9 +438,9 @@ def postpone_dequantize_op(self, graph_module: torch.fx.GraphModule) -> bool:
                         args=(user, *node.args[1:]),
                     )
                     dequant_node.meta = user.meta.copy()
-                    # Remove meta["debug_handle"] on new node. Reassign it at the
-                    # caller level by calling generate_missing_debug_handles
-                    dequant_node.meta.pop("debug_handle")
+                    # Remove meta["debug_handle"] on new node if it exists.
+                    # Reassign it at the caller level by calling generate_missing_debug_handles
+                    dequant_node.meta.pop("debug_handle", None)
                     user.replace_all_uses_with(dequant_node)
                     dequant_node.args = (user, *node.args[1:])
 

backends/cadence/aot/utils.py

@@ -235,14 +235,6 @@ def print_ops_info(
         )
 
 
-def model_gm_has_SDPA(model_gm: torch.fx.GraphModule) -> bool:
-    for node in model_gm.graph.nodes:
-        if node.op == "call_function":
-            if node.target == torch.ops.aten.scaled_dot_product_attention.default:
-                return True
-    return False
-
-
 def save_pte_program(
     prog: ExecutorchProgramManager, model_name: str, output_dir: str = ""
 ) -> None: