mindspore.train.callback._flops_collector 源代码

# Copyright 2024 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
"""FLops Utilization collector Callback class."""
from __future__ import absolute_import

import time
import os
import stat
import hashlib

from math import floor
from mindspore import _checkparam as Validator
from mindspore import log as logger
from mindspore.train.callback._callback import Callback
from mindspore.common.api import flops_collection

from mindspore.communication.management import (create_group, get_group_size,
                                                get_rank)
from mindspore.parallel._auto_parallel_context import auto_parallel_context
from mindspore.ops import operations as P
from mindspore.common import Tensor
from mindspore import context
import mindspore.nn as nn


class AllReduceNet(nn.Cell):
    """
    Used to accumulate flops in pipeline parallel.
    """
    def __init__(self, group_name):
        super(AllReduceNet, self).__init__()
        self.allreduce_sum = P.AllReduce(op=P.ReduceOp.SUM, group=group_name)
        self.add_flags(skip_auto_parallel_compile=True)

    def construct(self, x):
        return self.allreduce_sum(x)


[文档]class FlopsUtilizationCollector(Callback): """ The FlopsUtilizationCollector interface counts the model utilization information MFU and the hardware utilization information HFU. Currently, the API counts only the forward and backward flops of MatMul, BatchMatMul, FlashAttentionScore, and Conv2D operators. Only used in graph mode with static shape. Args: data_size (int): How many steps are the intervals between print information each time. computility (int): The peak flops of each compute card. Default: ``1`` . full_flops(bool): Whether to count the full model flops. If set full_flops to False, FlopsUtilizationCollector would count the shard model flops in each device. Default: ``True`` . Raises: TypeError: If data_size is not positive int. TypeError: If full_flops is not bool. AssertionError: If the training mode is not a static graph or not a static shape. Examples: >>> import numpy as np >>> import mindspore.dataset as ds >>> from mindspore import nn >>> from mindspore.train import Model, FlopsUtilizationCollector >>> from mindspore import context >>> context.set_context(mode=context.GRAPH_MODE) >>> data = {"x": np.float32(np.random.rand(64, 10)), "y": np.random.randint(0, 5, (64,))} >>> train_dataset = ds.NumpySlicesDataset(data=data).batch(32) >>> net = nn.Dense(10, 5) >>> crit = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean') >>> opt = nn.Momentum(net.trainable_params(), 0.01, 0.9) >>> flops_callback = FlopsUtilizationCollector(train_dataset.get_dataset_size(), computility=10e6) >>> model = Model(network=net, optimizer=opt, loss_fn=crit, metrics={"recall"}) >>> model.train(2, train_dataset, callbacks=[flops_callback]) Full model flops is 6400, Full hardware flops is 6400, Shard model flops is 6400, Shard hardware flops is 6400 Train per step time: 135.572 ms, mfu:0.47% hfu:0.47% Train per step time: 1.317 ms, mfu:48.59% hfu:48.59% """ def __init__(self, data_size, computility=1, full_flops=True): super(FlopsUtilizationCollector, self).__init__() self.step_time = time.time() self.computility = computility self.full_mfu = 0.0 self.full_hfu = 0.0 self.shard_mfu = 0.0 self.shard_hfu = 0.0 self.full_model_flops = 0.0 self.shard_model_flops = 0.0 self.full_hardware_flops = 0.0 self.shard_hardware_flops = 0.0 self.mfu_calculated = False self.data_size = data_size self.time_step_path = '' self.verbose = full_flops self.ma = os.environ.get("ENABLE_FLOPS_UTILIZATION_COLLECTOR") == "1" Validator.check_bool(full_flops, "verbose") Validator.check_positive_int(data_size, "data_size")
[文档] def step_begin(self, run_context): """ Record time at the beginning of step. Args: run_context (RunContext): Context of the process running. For more details, please refer to :class:`mindspore.train.RunContext`. """ self.step_time = time.time()
def _get_pipeline_group(self): """ Calculate the communication group between all pipeline stages """ rank = get_rank() stage_nums = auto_parallel_context().get_pipeline_stages() device_nums = get_group_size() per_stage_device_nums = device_nums // stage_nums local_stage_rank_id = rank % per_stage_device_nums group = range(0, stage_nums) rank_list = [local_stage_rank_id + x * per_stage_device_nums for x in group] rank_str_list = [str(local_stage_rank_id + x * per_stage_device_nums) for x in group] rank_list_str = "-".join(rank_str_list) return rank_list, rank_list_str
[文档] def step_end(self, run_context): """ Print mfu and hfu time at the end of step. Args: run_context (RunContext): Context of the process running. For more details, please refer to :class:`mindspore.train.RunContext`. """ if context.get_context("mode") != context.GRAPH_MODE: logger.warning("FlopsUtilizationCollector now only support graph mode.") return step_seconds = (time.time() - self.step_time) * 1000 if not self.mfu_calculated: cb_params = run_context.original_args() if cb_params.mode == 'train': network = cb_params.train_network elif cb_params.mode == 'eval': network = cb_params.eval_network else: logger.warning('FlopsUtilizationCollector only support train and eval mode!') return full_model_flops, full_hardware_flops, shard_model_flops, \ shard_hardware_flops, is_dynamic_shape = flops_collection(network.current_phase) if is_dynamic_shape: logger.warning("FlopsUtilizationCollector now do not support dynamic shape.") return self.full_mfu = full_model_flops / self.computility self.full_hfu = full_hardware_flops / self.computility self.shard_mfu = shard_model_flops / self.computility self.shard_hfu = shard_hardware_flops / self.computility self.full_model_flops = full_model_flops self.full_hardware_flops = full_hardware_flops self.shard_model_flops = shard_model_flops self.shard_hardware_flops = shard_hardware_flops self.mfu_calculated = True shard_mf_dir = os.path.realpath(os.getenv('MA_LOG_DIR', './')) if self.ma: flops_path = os.path.join(shard_mf_dir, "flops.txt") self.time_step_path = os.path.join(shard_mf_dir, "time_step.txt") if auto_parallel_context().get_parallel_mode() != "stand_alone": flops_path = os.path.join( shard_mf_dir, "flops_rank_" + str(get_rank())) + ".txt" self.time_step_path = os.path.join( shard_mf_dir, "time_step_rank_" + str(get_rank())) + ".txt" time_stamp = time.time() model_flops_log = "flops{{type=\"model_flops\"}} {} {}\n".\ format(shard_model_flops, int(round(time_stamp * 1000))) hardware_flops_log = "flops{{type=\"hardware_flops\"}} {} {}\n".\ format(shard_hardware_flops, int(round(time_stamp * 1000))) flags = os.O_WRONLY | os.O_CREAT modes = stat.S_IWUSR | stat.S_IRUSR with os.fdopen(os.open(flops_path, flags, modes), 'w') as f: f.write(model_flops_log) f.write(hardware_flops_log) if self.verbose: flops_log = f"Full model flops is {full_model_flops}, Full hardware flops is {full_hardware_flops}, " \ f"Shard model flops is {shard_model_flops}, Shard hardware flops is {shard_hardware_flops}." print(flops_log, flush=True) if auto_parallel_context().get_pipeline_stages() > 1: pipeline_group_list, pipeline_group_name = self._get_pipeline_group() auto_parallel_context().set_pipeline_stages(1) hashed = hashlib.md5( pipeline_group_name.encode()).hexdigest()[:48] pipeline_group_name = str(hashed) create_group(pipeline_group_name, pipeline_group_list) self.full_mfu = AllReduceNet(pipeline_group_name)( Tensor([self.full_mfu])).asnumpy()[0] self.full_hfu = AllReduceNet(pipeline_group_name)( Tensor([self.full_hfu])).asnumpy()[0] if auto_parallel_context().get_parallel_mode() != "stand_alone": self.full_mfu = self.full_mfu / get_group_size() self.full_hfu = self.full_hfu / get_group_size() step_size = self.data_size cb_params = run_context.original_args() if hasattr(cb_params, "batch_num"): batch_num = cb_params.batch_num if isinstance(batch_num, int) and batch_num > 0: step_size = cb_params.batch_num Validator.check_positive_int(step_size) if cb_params.dataset_sink_mode: step_seconds = step_seconds / step_size time_stamp = time.time() train_log = "time_monitor{{type=\"per_step_time\"}} {} {}".format(step_seconds, int(round(time_stamp * 1000))) if self.ma: flags = os.O_WRONLY | os.O_CREAT modes = stat.S_IWUSR | stat.S_IRUSR with os.fdopen(os.open(self.time_step_path, flags, modes), 'w') as f: f.write(train_log + '\n') train_log = "{} per step time: {:5.3f} ms".format( cb_params.mode.title(), step_seconds) if self.verbose: mfu = 1000 * self.full_mfu / step_seconds hfu = 1000 * self.full_hfu / step_seconds def floored_percentage(index, val, digits): val *= 10 ** (digits + 2) return index + '{1:.{0}f}%'.format(digits, floor(val) / 10 ** digits) train_log += floored_percentage(' mfu:', mfu, 2) train_log += floored_percentage(' hfu:', hfu, 2) print(train_log, flush=True)