|
| 1 | +# Copyright (c) Meta Platforms, Inc. and affiliates. |
| 2 | +# All rights reserved. |
| 3 | +# |
| 4 | +# This source code is licensed under the BSD-style license found in the |
| 5 | +# LICENSE file in the root directory of this source tree. |
| 6 | + |
| 7 | +import contextlib |
| 8 | +import os |
| 9 | +import tempfile |
| 10 | +import typing |
| 11 | + |
| 12 | +from typing import Any, Dict, final, List, Optional, Set |
| 13 | + |
| 14 | +import torch |
| 15 | +from executorch.exir._serialize._named_data_store import NamedDataStore |
| 16 | +from executorch.exir.backend.backend_details import ( |
| 17 | + BackendDetails, |
| 18 | + ExportedProgram, |
| 19 | + PreprocessResult, |
| 20 | +) |
| 21 | +from executorch.exir.backend.compile_spec_schema import CompileSpec |
| 22 | +from torch._inductor.codegen.cpp_wrapper_cpu import CppWrapperCpu |
| 23 | +from torch.export.passes import move_to_device_pass |
| 24 | + |
| 25 | + |
| 26 | +# exist fallback operators in et namespace; |
| 27 | +supported_fallback_kernels: Dict[str, Any] = {} |
| 28 | + |
| 29 | +# required fallback kernels but not supported |
| 30 | +missing_fallback_kernels: Set[str] = set() |
| 31 | + |
| 32 | + |
| 33 | +# context manager for non-fallback guarantee |
| 34 | +# it will raise exception when generating fallback kernels during aoti compile |
| 35 | +@contextlib.contextmanager |
| 36 | +def collect_unsupported_fallback_kernels(): |
| 37 | + original_generate_c_shim_extern_kernel_call = ( |
| 38 | + CppWrapperCpu.generate_c_shim_extern_kernel_call |
| 39 | + ) |
| 40 | + |
| 41 | + def generate_c_shim_extern_kernel_call_and_collect_unsupported_kernels( |
| 42 | + self, |
| 43 | + kernel: str, |
| 44 | + args: list[str], |
| 45 | + device: str, |
| 46 | + *, |
| 47 | + debug_args: Optional[list[str]] = None, |
| 48 | + ): |
| 49 | + if kernel not in supported_fallback_kernels: |
| 50 | + missing_fallback_kernels.add(kernel) |
| 51 | + |
| 52 | + original_generate_c_shim_extern_kernel_call( |
| 53 | + self, kernel, args, device, debug_args=debug_args |
| 54 | + ) |
| 55 | + |
| 56 | + CppWrapperCpu.generate_c_shim_extern_kernel_call = ( |
| 57 | + generate_c_shim_extern_kernel_call_and_collect_unsupported_kernels |
| 58 | + ) |
| 59 | + try: |
| 60 | + yield |
| 61 | + finally: |
| 62 | + CppWrapperCpu.generate_c_shim_extern_kernel_call = ( |
| 63 | + original_generate_c_shim_extern_kernel_call |
| 64 | + ) |
| 65 | + |
| 66 | + |
| 67 | +@final |
| 68 | +class CudaBackend(BackendDetails): |
| 69 | + """ |
| 70 | + CudaBackend is a backend that compiles a model to run on CUDA devices. It uses the AOTInductor compiler to generate |
| 71 | + optimized CUDA kernels for the model's operators with libtorch-free. The compiled model can be executed on CUDA devices |
| 72 | + using the Executorch runtime. |
| 73 | + """ |
| 74 | + |
| 75 | + @staticmethod |
| 76 | + def preprocess( |
| 77 | + edge_program: ExportedProgram, |
| 78 | + compile_specs: List[CompileSpec], |
| 79 | + ) -> PreprocessResult: |
| 80 | + # Move the edge_program from CPU to CUDA for aoti compile |
| 81 | + cuda_edge_program = move_to_device_pass(edge_program, "cuda") |
| 82 | + |
| 83 | + edge_program_module = cuda_edge_program.module() |
| 84 | + |
| 85 | + # Step 2: Grab all placeholders from the graph; last n should be user inputs |
| 86 | + user_input_names = cuda_edge_program.graph_signature.user_inputs |
| 87 | + user_input_placeholders = [] |
| 88 | + for node in cuda_edge_program.graph.nodes: |
| 89 | + if node.op == "placeholder" and node.name in user_input_names: |
| 90 | + user_input_placeholders.append(node.meta["val"]) |
| 91 | + |
| 92 | + # Step 3: Create pseudo user input using torch.randn and the generated input sizes |
| 93 | + faked_user_inputs = [] |
| 94 | + for placeholder in user_input_placeholders: |
| 95 | + if isinstance(placeholder, torch.Tensor): |
| 96 | + # Generate fake input with same shape and dtype, on CUDA |
| 97 | + fake_input = torch.randn( |
| 98 | + placeholder.shape, dtype=placeholder.dtype, device="cuda" |
| 99 | + ) |
| 100 | + faked_user_inputs.append(fake_input) |
| 101 | + |
| 102 | + faked_user_inputs = tuple(faked_user_inputs) |
| 103 | + |
| 104 | + # Create a temporary file path for the compiled shared library output |
| 105 | + output_path = tempfile.mktemp(suffix=".so", prefix="aoti_") |
| 106 | + |
| 107 | + options: dict[str, typing.Any] = { |
| 108 | + # Embed CUDA kernel binaries directly into the compiled shared object |
| 109 | + "aot_inductor.embed_kernel_binary": True, |
| 110 | + # Do not link against the full PyTorch/libtorch library |
| 111 | + "aot_inductor.link_libtorch": False, |
| 112 | + # Package model constants and other generated files directly in the shared object (.so) file |
| 113 | + "aot_inductor.package_constants_in_so": True, |
| 114 | + # Specify the output file path for the compiled shared object |
| 115 | + "aot_inductor.output_path": output_path, |
| 116 | + # Enable maximum automatic tuning for optimal performance |
| 117 | + "max_autotune": True, |
| 118 | + # Use TRITON for GEMM (General Matrix Multiply) operations tuning only to avoid using operators in libtorch |
| 119 | + "max_autotune_gemm_backends": "TRITON", |
| 120 | + # Use TRITON backend for convolution operations tuning only to avoid using operators in libtorch |
| 121 | + "max_autotune_conv_backends": "TRITON", |
| 122 | + } |
| 123 | + |
| 124 | + with collect_unsupported_fallback_kernels(): |
| 125 | + _ = torch._inductor.aot_compile(edge_program_module, faked_user_inputs, options=options) # type: ignore[arg-type] |
| 126 | + if len(missing_fallback_kernels) > 0: |
| 127 | + formatted_kernels = "\n - ".join(sorted(missing_fallback_kernels)) |
| 128 | + raise RuntimeError( |
| 129 | + f"Missing fallback kernels ({len(missing_fallback_kernels)} total):\n - {formatted_kernels}\n" |
| 130 | + "Please add them to the AOTI backend." |
| 131 | + ) |
| 132 | + |
| 133 | + with open(output_path, "rb") as f: |
| 134 | + so_data = f.read() |
| 135 | + |
| 136 | + named_data_store = NamedDataStore() |
| 137 | + named_data_store.add_named_data("so_blob", so_data, 1, "aoti_cuda_blob") |
| 138 | + |
| 139 | + # Clean up the temporary output file |
| 140 | + os.remove(output_path) |
| 141 | + |
| 142 | + return PreprocessResult( |
| 143 | + processed_bytes=b"", |
| 144 | + debug_handle_map={}, |
| 145 | + data_store_output=named_data_store.get_named_data_store_output(), |
| 146 | + ) |
0 commit comments