juda commented on code in PR #11911: URL: https://github.com/apache/tvm/pull/11911#discussion_r919996263
########## python/tvm/contrib/torch/optimize_torch.py: ########## @@ -0,0 +1,187 @@ +# pylint: disable=inconsistent-return-statements +#!/usr/bin/env python + +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +# pylint: disable=missing-module-docstring +# pylint: disable=missing-class-docstring +# pylint: disable=missing-function-docstring +""" +optimize_torch: aa function similar to `torch.jit.trace`, +which is used to optimize the `torch.nn.module` by TVM metaSchedule, +and returns a custom TorchScript operator +""" +import base64 +import contextlib +import tempfile +from typing import Dict, Optional, Tuple, Union + +import torch +import torch.utils.dlpack + +import tvm +from tvm import relay +from tvm._ffi import get_global_func, register_func +from tvm.ir.module import IRModule +from tvm.ir.transform import PassContext +from tvm.meta_schedule import TuneConfig, default_config +from tvm.meta_schedule.apply_history_best import ApplyHistoryBest +from tvm.meta_schedule.relay_integration import extract_task_from_relay +from tvm.meta_schedule.tune import tune_extracted_tasks +from tvm.meta_schedule.utils import autotvm_silencer +from tvm.runtime import vm +from tvm.runtime.module import Module +from tvm.runtime.ndarray import NDArray +from tvm.target.target import Target + + +# The python wrapper for GraphExecutorFactory +class GraphExecutorFactoryWrapper(torch.nn.Module): + def __init__(self, module: tvm.runtime.Module): + super().__init__() + self.inner_module = module + + def forward(self, *torch_inputs: Tuple[torch.Tensor]): + ret = self.inner_module.forward(torch_inputs) + if len(ret) == 1: + return ret[0] + return ret + + +def llvm_target(): + return "llvm -num-cores" + + +@register_func("script_torch.save_to_base64") +def save_to_base64(obj) -> bytes: + with tempfile.NamedTemporaryFile(suffix=".so") as tmpfile: + obj.export_library(tmpfile.name) + with open(tmpfile.name, "rb") as tfile: + return base64.b64encode(tfile.read()) + + +def tune_relay_auto( + mod: IRModule, + target: Union[str, Target], + config: TuneConfig, + work_dir: str, + backend: str = "graph", + params: Optional[Dict[str, NDArray]] = None, +) -> Union[Module, vm.Executable]: + """A wrapper of `tune_relay` but provide a default setting for the config. + + Parameters + ---------- + mod : IRModule + The module to tune. + target : Union[str, Target] + The target to tune for. + config : TuneConfig + The search strategy config. + params : Optional[Dict[str, tvm.runtime.NDArray]] + The associated parameters of the program + work_dir : Optional[str] + The working directory to save intermediate results. + backend : str = "graph" + The backend to use for relay compilation(graph / vm). + + Returns + ------- + lib : Union[Module, tvm.runtime.vm.Executable] + The built runtime module or vm Executable for the given relay workload. + """ + target = default_config.target(target) + extracted_tasks = extract_task_from_relay(mod, target, params) + if config is None: Review Comment: fixed ########## python/tvm/contrib/torch/optimize_torch.py: ########## @@ -0,0 +1,187 @@ +# pylint: disable=inconsistent-return-statements +#!/usr/bin/env python + +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +# pylint: disable=missing-module-docstring +# pylint: disable=missing-class-docstring +# pylint: disable=missing-function-docstring +""" +optimize_torch: aa function similar to `torch.jit.trace`, +which is used to optimize the `torch.nn.module` by TVM metaSchedule, +and returns a custom TorchScript operator +""" +import base64 +import contextlib +import tempfile +from typing import Dict, Optional, Tuple, Union + +import torch +import torch.utils.dlpack + +import tvm +from tvm import relay +from tvm._ffi import get_global_func, register_func +from tvm.ir.module import IRModule +from tvm.ir.transform import PassContext +from tvm.meta_schedule import TuneConfig, default_config +from tvm.meta_schedule.apply_history_best import ApplyHistoryBest +from tvm.meta_schedule.relay_integration import extract_task_from_relay +from tvm.meta_schedule.tune import tune_extracted_tasks +from tvm.meta_schedule.utils import autotvm_silencer +from tvm.runtime import vm +from tvm.runtime.module import Module +from tvm.runtime.ndarray import NDArray +from tvm.target.target import Target + + +# The python wrapper for GraphExecutorFactory +class GraphExecutorFactoryWrapper(torch.nn.Module): + def __init__(self, module: tvm.runtime.Module): + super().__init__() + self.inner_module = module + + def forward(self, *torch_inputs: Tuple[torch.Tensor]): + ret = self.inner_module.forward(torch_inputs) + if len(ret) == 1: + return ret[0] + return ret + + +def llvm_target(): + return "llvm -num-cores" + + +@register_func("script_torch.save_to_base64") +def save_to_base64(obj) -> bytes: + with tempfile.NamedTemporaryFile(suffix=".so") as tmpfile: + obj.export_library(tmpfile.name) + with open(tmpfile.name, "rb") as tfile: + return base64.b64encode(tfile.read()) + + +def tune_relay_auto( + mod: IRModule, + target: Union[str, Target], + config: TuneConfig, + work_dir: str, + backend: str = "graph", + params: Optional[Dict[str, NDArray]] = None, +) -> Union[Module, vm.Executable]: + """A wrapper of `tune_relay` but provide a default setting for the config. + + Parameters + ---------- + mod : IRModule + The module to tune. + target : Union[str, Target] + The target to tune for. + config : TuneConfig + The search strategy config. + params : Optional[Dict[str, tvm.runtime.NDArray]] + The associated parameters of the program + work_dir : Optional[str] + The working directory to save intermediate results. + backend : str = "graph" + The backend to use for relay compilation(graph / vm). + + Returns + ------- + lib : Union[Module, tvm.runtime.vm.Executable] + The built runtime module or vm Executable for the given relay workload. + """ + target = default_config.target(target) + extracted_tasks = extract_task_from_relay(mod, target, params) + if config is None: + config = TuneConfig( + num_trials_per_iter=16, + max_trials_global=16 * len(extracted_tasks), + ) + database = tune_extracted_tasks(extracted_tasks, config, work_dir) + relay_build = {"graph": relay.build, "vm": relay.vm.compile}[backend] + with target, autotvm_silencer(), ApplyHistoryBest(database): + with PassContext( + opt_level=3, + config={ + "relay.backend.use_meta_schedule": True, + "relay.backend.use_meta_schedule_dispatch": target.kind.name != "cuda", + }, + ): + return relay_build(mod, target=target, params=params) + + +def optimize_torch( + func, + example_inputs, + tuning_config=None, + target=None, + work_dir=None, +): + """Load PyTorch model that could be traced by TorchScript, then optimize it via MetaSchedule. + + Parameters + ---------- + func : callable or torch.nn.Module + A Python function or nn.Module that could run by TorchScript's trace. + (ie: torch.jit.trace(model, input)) + + example_inputs : tuple or torch.Tensor + Inputs to `torch.jit.trace`. + + tuning_config : tvm.meta_schedule.TuneConfig + The configuration for tuning by MetaSchedule. + If user doesn't set the config, the tuning will run with a default setting, + a number proportional to the tasks of the module. Review Comment: fixed -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
