Metal backend: Add operator implementations

backends/aoti/aoti_model_container.cpp

@@ -25,6 +25,12 @@ AOTInductorModelContainerGetNumOutputsFunc
     AOTInductorModelContainerGetNumOutputs = nullptr;
 AOTInductorModelContainerRunFunc AOTInductorModelContainerRun = nullptr;
 
+// Global function pointers needed by Metal backend
+AOTInductorModelContainerGetInputNameFunc
+    AOTInductorModelContainerGetInputName = nullptr;
+AOTInductorModelContainerGetNumConstantsFunc
+    AOTInductorModelContainerGetNumConstants = nullptr;
+
 } // extern "C"
 
 } // namespace aoti

backends/aoti/aoti_model_container.h

@@ -70,6 +70,22 @@ extern AOTInductorModelContainerGetNumOutputsFunc
     AOTInductorModelContainerGetNumOutputs;
 extern AOTInductorModelContainerRunFunc AOTInductorModelContainerRun;
 
+// Function pointer types needed by Metal backend
+using AOTInductorModelContainerGetInputNameFunc = AOTIRuntimeError (*)(
+    AOTInductorModelContainerHandle container_handle,
+    size_t input_idx,
+    const char** input_name);
+
+using AOTInductorModelContainerGetNumConstantsFunc = AOTIRuntimeError (*)(
+    AOTInductorModelContainerHandle container_handle,
+    size_t* num_constants);
+
+// Global function pointers needed by Metal backend
+extern AOTInductorModelContainerGetInputNameFunc
+    AOTInductorModelContainerGetInputName;
+extern AOTInductorModelContainerGetNumConstantsFunc
+    AOTInductorModelContainerGetNumConstants;
+
 } // extern "C"
 
 // AOTI Delegate Handle structure

backends/aoti/common_shims.cpp

@@ -145,6 +145,11 @@ void cleanup_tensor_metadata() {
   internal::tensor_to_strides.clear();
 }
 
+// Needed by Metal backend
+size_t aoti_torch_dtype_element_size(int32_t dtype) {
+  return dtype_to_element_size(dtype);
+}
+
 } // extern "C"
 
 } // namespace aoti

backends/aoti/common_shims.h

@@ -68,6 +68,9 @@ void aoti_torch_grad_mode_set_enabled(bool enabled);
 // Cleanup functions for clearing global state
 void cleanup_tensor_metadata();
 
+// Needed by Metal backend
+size_t aoti_torch_dtype_element_size(int32_t dtype);
+
 } // extern "C"
 
 } // namespace aoti

backends/apple/metal/metal_backend.py

@@ -0,0 +1,173 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import contextlib
+import os
+import typing
+from enum import Enum
+
+from typing import Any, Dict, final, List, Optional, Set
+
+import torch
+from executorch.backends.apple.metal.replace_slice_copy_with_slice import (
+    ReplaceSliceCopyWithSlicePass,
+)
+from executorch.exir._serialize._named_data_store import NamedDataStore
+from executorch.exir._warnings import experimental
+from executorch.exir.backend.backend_details import (
+    BackendDetails,
+    ExportedProgram,
+    PreprocessResult,
+)
+from executorch.exir.backend.compile_spec_schema import CompileSpec
+from torch._inductor.codegen.cpp_wrapper_cpu import CppWrapperCpu
+from torch.export.passes import move_to_device_pass
+
+
+# exist fallback operators in et namespace;
+supported_fallback_kernels: Dict[str, Any] = {
+    "aoti_torch_mps_addmm_out": None,
+    "aoti_torch_mps_convolution": None,
+    "aoti_torch_mps_mm_out": None,
+    "at::_ops::_scaled_dot_product_attention_math_for_mps::call": None,
+}
+
+# required fallback kernels but not supported
+missing_fallback_kernels: Set[str] = set()
+
+
+class COMPILE_SPEC_KEYS(Enum):
+    METHOD_NAME = "method_name"
+
+
+# context manager for non-fallback guarantee
+# it will raise exception when generating fallback kernels during aoti compile
+@contextlib.contextmanager
+def collect_unsupported_fallback_kernels():
+    original_generate_c_shim_extern_kernel_call = (
+        CppWrapperCpu.generate_c_shim_extern_kernel_call
+    )
+
+    def generate_c_shim_extern_kernel_call_and_collect_unsupported_kernels(
+        self,
+        kernel: str,
+        args: list[str],
+        device: str,
+        *,
+        debug_args: Optional[list[str]] = None,
+        debug_handle: Optional[int] = None,
+    ):
+        if kernel not in supported_fallback_kernels:
+            missing_fallback_kernels.add(kernel)
+
+        original_generate_c_shim_extern_kernel_call(
+            self, kernel, args, device, debug_args=debug_args, debug_handle=debug_handle
+        )
+
+    CppWrapperCpu.generate_c_shim_extern_kernel_call = (
+        generate_c_shim_extern_kernel_call_and_collect_unsupported_kernels
+    )
+    try:
+        yield
+    finally:
+        CppWrapperCpu.generate_c_shim_extern_kernel_call = (
+            original_generate_c_shim_extern_kernel_call
+        )
+
+
+@final
+@experimental(
+    "This API and all of Metal backend related functionality are experimental."
+)
+class MetalBackend(BackendDetails):
+    @staticmethod
+    def preprocess(
+        edge_program: ExportedProgram,
+        compile_specs: List[CompileSpec],
+    ) -> PreprocessResult:
+        print("entering the lowerable parts in MetalBackend.preprocess....")
+        # Move the edge_program from CPU to MPS for aoti compile
+        mps_edge_program = move_to_device_pass(edge_program, "mps")
+
+        # replace slice_copy with slice
+        ReplaceSliceCopyWithSlicePass()(mps_edge_program.graph_module)
+
+        edge_program_module = mps_edge_program.module()
+
+        # Grab all input placeholders from the graph
+        user_input_names = mps_edge_program.graph_signature.user_inputs
+        user_input_placeholders = []
+        for node in mps_edge_program.graph.nodes:
+            if node.op == "placeholder" and node.name in user_input_names:
+                user_input_placeholders.append(node.meta["val"])
+
+        # Base options for all devices
+        options: dict[str, typing.Any] = {
+            # Do not link against the full PyTorch/libtorch library
+            "aot_inductor.link_libtorch": False,
+            # Package model constants and other generated files directly in the shared object (.so) file
+            "aot_inductor.package_constants_in_so": True,
+            # Enable maximum automatic tuning for optimal performance
+            "max_autotune": True,
+            # "aot_inductor.debug_compile": True,
+            # "aot_inductor.force_mmap_weights": False,
+        }
+
+        with collect_unsupported_fallback_kernels():
+            so_path = torch._inductor.aot_compile(edge_program_module, tuple(user_input_placeholders), options=options)  # type: ignore[arg-type]
+            if len(missing_fallback_kernels) > 0:
+                formatted_kernels = "\n  - ".join(sorted(missing_fallback_kernels))
+                raise RuntimeError(
+                    f"Missing fallback kernels ({len(missing_fallback_kernels)} total):\n  - {formatted_kernels}\n"
+                    "Please add them to the AOTI backend."
+                )
+
+        # pyre-ignorep[6]: Incompatible parameter type
+        with open(so_path, "rb") as f:
+            so_data = f.read()
+
+        named_data_store = NamedDataStore()
+        method_name = MetalBackend.method_name_from_compile_specs(compile_specs)
+        named_data_store.add_named_data(
+            method_name + "_so_blob", so_data, 1, "aoti_metal_blob"
+        )
+
+        # Clean up the generated so file; it has been packaged into the NamdeDataStore
+        # pyre-ignorep[6]: Incompatible parameter type
+        os.remove(so_path)
+
+        return PreprocessResult(
+            processed_bytes=b"",
+            debug_handle_map={},
+            data_store_output=named_data_store.get_named_data_store_output(),
+        )
+
+    @staticmethod
+    def generate_method_name_compile_spec(
+        method_name: str,
+    ) -> CompileSpec:
+        """
+        Returns the compile spec representing the model compute precision, for additional details
+        please refer to the documentation for ``coremltools.precision``.
+        """
+        return CompileSpec(
+            COMPILE_SPEC_KEYS.METHOD_NAME.value,
+            method_name.encode("utf-8"),
+        )
+
+    @staticmethod
+    def method_name_from_compile_specs(
+        compile_specs: List[CompileSpec],
+    ) -> str:
+        """
+        Returns the method name from the compile specs.
+        """
+        for spec in compile_specs:
+            if spec.key == COMPILE_SPEC_KEYS.METHOD_NAME.value:
+                return spec.value.decode("utf-8")
+        raise RuntimeError(
+            f"Could not find method name in compile specs: {compile_specs}"
+        )

backends/apple/metal/metal_partitioner.py

@@ -0,0 +1,77 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Callable, Dict, final, List, Optional, Tuple
+
+import torch
+from executorch.backends.apple.metal.metal_backend import MetalBackend  # usort: skip
+from executorch.exir._warnings import experimental
+from executorch.exir.backend.compile_spec_schema import CompileSpec
+from executorch.exir.backend.partitioner import (
+    DelegationSpec,
+    Partitioner,
+    PartitionResult,
+)
+from executorch.exir.backend.utils import tag_constant_data, tag_mutated_buffer
+from torch.export.exported_program import ExportedProgram
+
+
+@final
+@experimental(
+    "This API and all of Metal backend related functionality are experimental."
+)
+class MetalPartitioner(Partitioner):
+    """
+    Metal partitioner for AOTInductor backend integration.
+
+    This partitioner creates a single partition containing all operators from the input graph.
+    It skips core ATen decomposition, allowing the Metal backend to handle decomposition using
+    AOTInductor's MPS-specific decomposition table.
+
+    Only operators that cannot be handled by the aoti-mps library will be excluded from
+    the partition and fall back to ExecuTorch's default or custom handling.
+    """
+
+    def __init__(self, compile_spec: List[CompileSpec]) -> None:
+        self.delegation_spec = DelegationSpec(MetalBackend.__name__, compile_spec)
+
+    def partition(self, exported_program: ExportedProgram) -> PartitionResult:
+        """
+        Fully delegate the graph to AOTInductor by tagging all nodes as a single partition.
+        """
+
+        partition_tags: Dict[str, DelegationSpec] = {}
+        tag = "tag0"
+
+        for node in exported_program.graph.nodes:
+            if node.op != "call_function":
+                continue
+            node.meta["delegation_tag"] = tag
+
+        partition_tags[tag] = self.delegation_spec
+
+        tag_constant_data(exported_program)
+        tag_mutated_buffer(exported_program)
+
+        return PartitionResult(
+            tagged_exported_program=exported_program, partition_tags=partition_tags
+        )
+
+    def ops_to_not_decompose(
+        self, ep: ExportedProgram
+    ) -> Tuple[List[torch._ops.OpOverload], Optional[Callable[[torch.fx.Node], bool]]]:
+        """
+        Return a list of operations that should not be decomposed and let the AOT compiler handle them.
+        Currently we skip ATen decompositon for all ops, and let the Metal backend handle them.
+        """
+        do_not_decompose = set()
+
+        for node in ep.graph.nodes:
+            if node.op == "call_function" and isinstance(
+                node.target, torch._ops.OpOverload
+            ):
+                do_not_decompose.add(node.target)
+        return list(do_not_decompose), None