optimization_tuner.py

#   Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed 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.

import os
import sys
import copy
import shlex
import pathlib
import time
import shutil
import pickle
import json
import logging
import subprocess
import traceback

import paddle
from paddle.fluid import program_guard
from paddle.fluid.backward import append_backward
from paddle.distributed.passes import new_pass, PassContext
from paddle.distributed.utils import get_logger

from paddle.distributed.auto_parallel.dist_context import DistributedContext, get_default_distributed_context
from paddle.distributed.auto_parallel.completion import Completer
from paddle.distributed.auto_parallel.reshard import Resharder
from paddle.distributed.auto_parallel.partitioner import Partitioner
from paddle.distributed.auto_parallel.process_group import clear_all_process_groups, get_all_process_groups
from paddle.distributed.auto_parallel.utils import debug_program
from paddle.distributed.auto_parallel.utils import make_data_unshard, set_grad_var_shape

from .config import TuningConfig
from .algorithms import new_algorithm
from .trial import TrialStatus


def _get_new_params_grads(target_program, ref_program, ref_params_grads):
    ref_block = ref_program.global_block()
    target_block = target_program.global_block()
    target_params_grads = []

    for p, g in ref_params_grads:
        # NOTE grad var might not be generated
        assert ref_block.has_var(p.name)
        assert target_block.has_var(p.name)
        new_p = target_block.var(p.name)
        if g:
            new_g = target_block.var(g.name)
        else:
            new_g = None

        target_params_grads.append((new_p, new_g))

    return target_params_grads


def _get_new_loss(target_program, ref_program, loss):
    ref_block = ref_program.global_block()
    target_block = target_program.global_block()
    assert ref_block.has_var(loss.name)

    return target_block.var(loss.name)


def parse_process_groups():
    group_map = {}
    all_process_groups = get_all_process_groups()
    for process_group in all_process_groups:
        group_map[process_group.id] = process_group.ranks
    return group_map


def get_metric(results):
    assert isinstance(
        results,
        dict), "results should be type of dictionary, but got {}.".format(
            type(results))
    if 'Throughtput' in results and isinstance(results['Throughtput'], float):
        return float(results['Throughtput'])
    else:
        return -1.0


def parse_results(results):
    if results['Throughtput'] > 0:
        return "Throughtput: {} step / s.".format(results['Throughtput'])
    et = results.get("ErrorType", None)
    if et == "ResourceExhaustedError":
        return "Fail with OOM"
    else:
        return "Fail with UNKWON ERROR"


# TODO only dependent on dist context
# all env need to be start a new pass are member of dist context
def _copy_context(ref_dist_context):

    clear_all_process_groups()

    new_dist_context = DistributedContext()
    new_dist_context._serial_main_program = ref_dist_context.serial_main_program.clone(
        for_test=False)
    new_dist_context._serial_startup_program = ref_dist_context.serial_startup_program.clone(
        for_test=False)

    # mapping variable into new dist context
    if getattr(ref_dist_context, '_params_grads', None):
        new_dist_context._params_grads = _get_new_params_grads(
            new_dist_context.serial_main_program,
            ref_dist_context.serial_main_program,
            ref_dist_context._params_grads)
    new_dist_context._serial_loss = _get_new_loss(
        new_dist_context.serial_main_program,
        ref_dist_context.serial_main_program, ref_dist_context.serial_loss)

    for key, var_list in ref_dist_context._serial_feed_vars.items():
        new_var_list = []
        for var in var_list:
            block_idx = var.block.idx
            var_name = var.name
            var = new_dist_context._serial_main_program.blocks[
                block_idx]._var_recursive(var_name)
            new_var_list.append(var)
        new_dist_context._serial_feed_vars[key] = new_var_list

    for key, var_list in ref_dist_context._serial_fetch_vars.items():
        new_var_list = []
        for var in var_list:
            block_idx = var.block.idx
            var_name = var.name
            var = new_dist_context._serial_main_program.blocks[
                block_idx]._var_recursive(var_name)
            new_var_list.append(var)
        new_dist_context._serial_fetch_vars[key] = new_var_list

    # copy information in forward and backward
    new_dist_context._serial_optimizer = copy.deepcopy(
        ref_dist_context.serial_optimizer)
    new_dist_context._dist_tensors_for_program = copy.deepcopy(
        ref_dist_context._dist_tensors_for_program)
    new_dist_context._dist_ops_for_program = copy.deepcopy(
        ref_dist_context._dist_ops_for_program)
    for pm in ref_dist_context.process_meshes:
        new_dist_context.add_process_mesh(pm)
    new_dist_context._dist_op_context = copy.deepcopy(
        ref_dist_context._dist_op_context)
    new_dist_context._block_state = copy.deepcopy(ref_dist_context.block_state)

    return new_dist_context


class OptimizationTuner:
    """
    OptimizationTuner is used to manage the tuning procedure of hyper-parameters (configs) 
    of Optimization Pass in AutoParallel.
    """

    def __init__(
        self,
        user_configs,
        dist_context,
        dataset,
        inputs_spec,
        labels_spec,
        batch_size,
        rank,
    ):

        self._config = TuningConfig(user_configs, dist_context._strategy)
        # should not modify dist context from calling function
        self._baseline_dist_context = _copy_context(dist_context)
        self._baseline_completer = Completer(self._baseline_dist_context)

        self._rank = rank
        self._inputs_spec = inputs_spec
        self._labels_spec = labels_spec
        self._dataset = dataset
        self._batch_size = batch_size

        self._finished_trials = []
        self._best_metric = None
        self._best_iter = float("-inf")

        self._logger = get_logger(logging.INFO)

        self._build_programs_without_optimization()
        self._select_tuning_algorithm()

    @property
    def project_dir(self):
        dirname = self._config.project_dir
        if not os.path.exists(dirname):
            if self.rank == 0:
                pathlib.Path(dirname).mkdir(parents=True, exist_ok=True)
        return dirname

    @property
    def rank(self):
        return self._rank

    @property
    def device_id(self):
        return paddle.distributed.ParallelEnv().device_id

    # TODO Generate compelet program with all parts like forward, backward, update
    # as well as parallelism transformation.
    def _build_programs_without_optimization(self):

        serial_main_program = self._baseline_dist_context.serial_main_program
        serial_startup_program = self._baseline_dist_context.serial_startup_program
        serial_loss = self._baseline_dist_context.serial_loss

        with program_guard(serial_main_program, serial_startup_program):
            params_grads = append_backward(
                serial_loss,
                distop_context=self._baseline_dist_context.dist_op_context)

        self._baseline_completer.complete_backward_annotation(
            serial_main_program)
        self._baseline_dist_context.block_state.parse_backward_blocks(
            serial_main_program)
        self._baseline_dist_context._params_grads = params_grads

        if self._config.verbose:
            baseline_dir = os.path.join(self.project_dir, "baseline")
            if not os.path.exists(baseline_dir):
                pathlib.Path(baseline_dir).mkdir(parents=True, exist_ok=True)
            debug_program(self._baseline_dist_context._serial_main_program,
                          baseline_dir, "main")
            debug_program(self._baseline_dist_context._serial_startup_program,
                          baseline_dir, "startup")

    def _select_tuning_algorithm(self):

        selected_passes_set = self._config.tuning_passes_name
        algorithm_name = "_".join(sorted(selected_passes_set))
        self._algorithm = new_algorithm(algorithm_name, self._config)

    def _apply_optimization(self, trial):
        new_strategy = trial.space
        dist_context = _copy_context(self._baseline_dist_context)
        pass_context = PassContext()
        completer = Completer(dist_context)

        main_program = dist_context.serial_main_program
        startup_program = dist_context.serial_startup_program

        # applying optimization pass
        if new_strategy.amp:
            config = copy.deepcopy(new_strategy.amp_configs)
            config["dist_context"] = dist_context
            config["params_grads"] = dist_context._params_grads

            # TODO AMP Pass should not use loss var
            config["loss"] = dist_context.serial_loss
            config["input_data"] = self._baseline_dist_context.serial_feed_vars["inputs"] \
                + self._baseline_dist_context.serial_feed_vars["labels"]
            if config["use_pure_fp16"]:
                config["base_opt"] = dist_context.serial_optimizer
                auto_parallel_fp16_pass = new_pass("auto_parallel_fp16", config)
                auto_parallel_fp16_pass.apply([main_program], [startup_program],
                                              pass_context)
            else:
                auto_parallel_amp_pass = new_pass("auto_parallel_amp", config)
                auto_parallel_amp_pass.apply([main_program], [startup_program],
                                             pass_context)

        if new_strategy.recompute:
            config = copy.deepcopy(new_strategy.recompute_configs)
            config["dist_context"] = dist_context
            config["no_grad_set"] = None
            config["loss"] = dist_context.serial_loss
            auto_parallel_recompute_pass = new_pass("auto_parallel_recompute",
                                                    config)
            auto_parallel_recompute_pass.apply([main_program],
                                               [startup_program], pass_context)

        # Do logical partition
        partitioner = Partitioner(dist_context, self.rank)
        dist_main_prog, dist_startup_prog, dist_params_grads = partitioner.partition(
            main_program, startup_program, dist_context._params_grads)

        # Generate optimizer
        # FIXME should be remove from apply pass after pass support optimizers
        with program_guard(dist_main_prog, dist_startup_prog):
            optimizer_ops = dist_context.serial_optimizer.apply_gradients(
                dist_params_grads)
        completer.complete_update_annotation(dist_main_prog)

        # Do reshard process
        set_grad_var_shape(dist_main_prog, dist_context)
        resharder = Resharder(dist_main_prog, dist_startup_prog, self.rank,
                              dist_context, dist_params_grads)
        resharder.reshard()

        if new_strategy.sharding:
            config = copy.deepcopy(new_strategy.sharding_configs)
            config["dist_context"] = dist_context
            config["params_grads"] = dist_params_grads
            config["global_rank"] = self.rank
            auto_parallel_sharding_pass = new_pass("auto_parallel_sharding",
                                                   config)
            auto_parallel_sharding_pass.apply([dist_main_prog],
                                              [dist_startup_prog], pass_context)

        if new_strategy.gradient_merge:
            config = copy.deepcopy(new_strategy.gradient_merge_configs)
            config["dist_context"] = dist_context
            config["params_grads"] = dist_params_grads
            auto_parallel_gradient_merge_pass = new_pass(
                "auto_parallel_gradient_merge_pass", config)
            auto_parallel_gradient_merge_pass.apply([dist_main_prog],
                                                    [dist_startup_prog],
                                                    pass_context)
        trial.main_program, trial.startup_program = dist_main_prog, dist_startup_prog
        return trial

    def _get_profile_context(self, trial, result_path):

        profile_ctx = {}

        profile_ctx['distributed_env'] = copy.deepcopy(
            paddle.distributed.ParallelEnv())
        profile_ctx['group_map'] = parse_process_groups()
        profile_ctx[
            "loss_var_name"] = self._baseline_dist_context.serial_loss.name
        profile_ctx[
            "main_program_decs"] = trial.main_program.desc.serialize_to_string(
            )
        profile_ctx[
            "startup_program_decs"] = trial.startup_program.desc.serialize_to_string(
            )
        self._dataset.batch_size = self._batch_size
        self._dataset.input_names = self._get_input_names()

        profile_ctx["dataset"] = self._dataset
        profile_ctx["result_filename"] = result_path

        return profile_ctx

    def _get_input_names(self):
        input_names = []
        for input_spec in self._inputs_spec[:] + self._labels_spec[:]:
            input_names.append(input_spec.name)
        return input_names

    def _launch_profile(self, ctx_path, trial_dir):

        if os.environ.get("WITH_COVERAGE", "OFF") == "ON":
            coverage_args = ["-m", "coverage", "run", "--branch", "-p"]
        else:
            coverage_args = []

        profile_args = " ".join([
            "--rank",
            str(self.rank), "--device_id",
            str(self.device_id), "--ctx_filename", ctx_path,
            "--profile_start_step",
            str(self._config.profile_start_step), "--profile_end_step",
            str(self._config.profile_end_step)
        ])
        cmd_args = "-m paddle.distributed.auto_parallel.tuner.profiler" + " " + profile_args
        cmd = [sys.executable, "-u"] + coverage_args + shlex.split(cmd_args)

        parent_env = copy.copy(os.environ.copy())
        # env flags need for profile
        new_env = {
            "FLAGS_USE_STANDALONE_EXECUTOR": "False",
        }
        new_env.update(parent_env)

        # TODO if any rank hang or fail, kill all processes
        self._logger.debug("Executing cmd:\n{} .".format(" ".join(cmd)))
        # new_process = subprocess.Popen(cmd, env=new_env)
        with open(os.path.join(trial_dir, "stdout.log" + str(self.rank)),
                  "wb") as out, open(
                      os.path.join(trial_dir, "stderr.log" + str(self.rank)),
                      "wb") as err:
            result = subprocess.Popen(cmd, stdout=out, stderr=err, env=new_env)
            result.wait()
            out.flush()
            err.flush()
            os.fsync(out)
            os.fsync(err)

    def _profile_trial(self, trial):
        # Making working directory
        trial_dir = self._get_trial_dir(trial)
        if not os.path.exists(trial_dir):
            if self.rank == 0:
                pathlib.Path(trial_dir).mkdir(parents=True, exist_ok=True)
            else:
                while not os.path.exists(trial_dir):
                    pass
        ctx_filename = "profile_ctx." + str(self.rank)
        ctx_path = os.path.join(trial_dir, ctx_filename)
        result_path = os.path.join(trial_dir, "result.json")

        # Prepare Profile Context
        profile_ctx = self._get_profile_context(trial, result_path)
        with open(ctx_path, 'wb') as f:
            pickle.dump(profile_ctx, f, protocol=4)

        if self._config.verbose:
            debug_program(trial.main_program, trial_dir, "main_program")
            debug_program(trial.startup_program, trial_dir, "startup_program")

        # Run
        self._launch_profile(ctx_path, trial_dir)

        # Load results
        try:
            with open(result_path, 'r') as fp:
                results = json.load(fp)
            return results
        except FileNotFoundError:
            Error_results = {"Throughtput": -1, "ErrorType": 'FatalError'}
            return Error_results

    def _evaluate_trial(self, trial):

        self._logger.info("Trial {} evaluation start.".format(trial.name))
        self._apply_optimization(trial)

        if self._config.mode == "PROFILE":
            results = self._profile_trial(trial)

        elif self._config.mode == "COSTMODEL":
            raise NotImplementedError(
                "COSTMODEL mode for optimization tuning is not supported yet!")
        else:
            raise NotImplementedError("invalid evaluation mode: {}".format(
                self._config.mode))

        self._logger.info("Trial {} evaluation finish with {}.".format(
            trial.name, parse_results(results)))
        return results

    def _update(self, i, trial, results):
        self._finished_trials.append(trial)

        cur_mertic = get_metric(results)
        if self._best_metric == None or cur_mertic > self._best_metric:
            self._best_metric = cur_mertic
            self._best_iter = i

    def _get_trial_dir(self, trial):
        return os.path.join(self.project_dir, trial.name)

    def get_best_config(self):
        """
        Return the best optimization configuration found in the tuning.

        Returns:
            A object of fleet.DistributedStrategy with best configuration.       
        """
        assert self._best_iter >= 0, "The best configuration is not found yet !"
        best_trial = self._finished_trials[self._best_iter]
        return self._algorithm.get_config_from_trial(best_trial)

    def summary(self):
        """
        Display tuning result summary.
        """
        # TODO summary with the trial_name with metric_of_trial
        best_trial = self._finished_trials[self._best_iter]
        summary_ = """
Tuning Result Summary
Run total {} trials with {} min.
The best trial is: [{}], whose configuration is following: 
        """.format(len(self._finished_trials),
                   (time.time() - self._tuning_start_time) / 60,
                   best_trial.name)
        summary_ += "\n" + best_trial.summary() + "\n"\

        self._logger.info(summary_)
        with open(os.path.join(self.project_dir, "summary.txt"), "w+") as fw:
            for line in summary_.split("\n"):
                fw.write(line + "\n")

        full_strategy = self.get_best_config()
        full_strategy.save_to_prototxt(
            os.path.join(self.project_dir, "tuned_dist_strategy.prototxt"))

    def clear(self):
        """
        Clear the temporary file generated in tuning procedure.
        """
        # TODO clear up zombie process created by tuning
        if not self._config.verbose:
            for trial in self._finished_trials:
                trial_dir = self._get_trial_dir(trial)
                shutil.rmtree(trial_dir, ignore_errors=True)

    def tune(self):
        """
        Performs the search for best hyperparameter configuations 
        for the selected optimization pass(es). 
        """

        # step1: collect model info which might be used for
        # pruning the search space of the algorithm
        self._tuning_start_time = time.time()
        self._algorithm.collect_model_info(
            self._baseline_dist_context.serial_main_program,
            self._baseline_dist_context.serial_startup_program)

        # main search loop
        i = 0
        while i < self._config.max_num_trial:
            # step2: create a new trial
            trial = self._algorithm.next_trial()

            if trial.status == TrialStatus.STOPPED:
                break

            # step3: evaluate the trial
            results = self._evaluate_trial(trial)

            # step4: update the algorithm with last result,
            # which could be used by algorithm to pruning the
            # remaining search space.
            self._algorithm.update(results)
            self._update(i, trial, results)

            # early stop
            i += 1
            if self._config.early_stop and self._config.early_stop <= i - self._best_iter:
                self._logger.info(
                    "Early stop the Tuning since there is no better trial found within [{}] trials"
                    .format(self._config.early_stop))
                break

        # step5: summary the best config and return
        self.summary()

        self.clear()