[源码解析] 模型并行分布式训练 Megatron (4) — 如何设置各种并行

[源码解析] 模型并行分布式训练 Megatron (4) — 如何设置各种并行

0x00 摘要

NVIDIA Megatron 是一个基于 PyTorch 的分布式训练框架,用来训练超大Transformer语言模型,其通过综合应用了数据并行,Tensor并行和Pipeline并行来复现 GPT3,值得我们深入分析其背后机理。

本系列大概有 5 篇文章,通过论文和源码和大家一起学习研究。本文将看看 Megatron 如何处理设置并行。

本系列其他文章为:

[源码解析] 模型并行分布式训练Megatron (1) — 论文 & 基础

[源码解析] 模型并行分布式训练Megatron (2) — 整体架构

[源码解析] 模型并行分布式训练 Megatron (3) —模型并行实现

0x01 前文回顾

前文我们对模型并行的原理和代码进行了分析,对于给定的模型,现在还需要解决几个问题:

  • 如何把模型切分给节点,比如哪个节点负责哪些层。
  • 数据并行,模型并行,流水线并行这几种并行之中,每个节点分别属于哪个部分?
  • 如何避免流水线带来的问题。

我们接下来就仔细分析一下。

0x02 初始化

initialize_model_parallel 方法用来设置模型并行,所以我们接下来就具体分析。

2.1 全局变量

因为前文_initialize_distributed之中调用了torch.distributed.init_process_group 初始化分布式环境,所以我们知道,每个进程都有自己的 gloabl rank 和 local rank,都有自己的全局变量。

主要变量如下(具体例子可以结合 initialize_model_parallel 之中的注释来看):

  • _TENSOR_MODEL_PARALLEL_GROUP :当前 rank 所属于的Intra-layer model parallel group,就是tensor 并行进程组。
    • 假如每一层分为两个tensor,则 _TENSOR_MODEL_PARALLEL_GROUP 例子为:[g0, g1], [g2, g3], [g4, g5], [g6, g7], [g8, g9], [g10, g11], [g12, g13], [g14, g15]。
  • _PIPELINE_MODEL_PARALLEL_GROUP :当前 rank 所属于的Intra-layer model parallel group,就是流水线进程组。
    • 假如流水线深度为4,则例子为 [g0, g4, g8, g12], [g1, g5, g9, g13], [g2, g6, g10, g14], [g3, g7, g11, g15]。
  • _MODEL_PARALLEL_GROUP :当前 rank 所属于的模型并行进程组,包括了以上两组。
    • 针对我们例子,就是完整模型被复制了两份,其 GPU 节点具体是[0, 1, 4, 5, 8, 9, 12, 13],[2, 3, 6, 7, 10, 11, 14, 15]
  • _EMBEDDING_GROUP : 嵌入对应的进程组。
  • _DATA_PARALLEL_GROUP :当前 rank 所属于的Data parallel group。
    • 假如数据并行度数为2,则例子为[g0, g2], [g1, g3], [g4, g6], [g5, g7], [g8, g10], [g9, g11], [g12, g14], [g13, g15]。

具体如下:

# Intra-layer model parallel group that the current rank belongs to.
_TENSOR_MODEL_PARALLEL_GROUP = None
# Inter-layer model parallel group that the current rank belongs to.
_PIPELINE_MODEL_PARALLEL_GROUP = None
# Model parallel group (both intra- and pipeline) that the current rank belongs to.
_MODEL_PARALLEL_GROUP = None
# Embedding group.
_EMBEDDING_GROUP = None
# Data parallel group that the current rank belongs to.
_DATA_PARALLEL_GROUP = None

_VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK = None
_VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = None
_PIPELINE_MODEL_PARALLEL_SPLIT_RANK = None

# These values enable us to change the mpu sizes on the fly.
_MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE = None
_MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = None
_MPU_TENSOR_MODEL_PARALLEL_RANK = None
_MPU_PIPELINE_MODEL_PARALLEL_RANK = None

# A list of ranks that have a copy of the embedding.
_EMBEDDING_GLOBAL_RANKS = None

# A list of global ranks for each pipeline group to ease calculation of the source
# rank when broadcasting from the first or last pipeline stage.
_PIPELINE_GLOBAL_RANKS = None

2.2 初始化代码

我们首先把 initialize_model_parallel 代码摘录出来。initialize_model_parallel 作用就是对模型进行分组,然后初始化进程组相关的各种全局变量。

def initialize_model_parallel(tensor_model_parallel_size_=1,
                              pipeline_model_parallel_size_=1,
                              virtual_pipeline_model_parallel_size_=None,
                              pipeline_model_parallel_split_rank_=None):
    """
    Initialize model data parallel groups.

    Arguments:
        tensor_model_parallel_size: number of GPUs used for tensor model parallelism.
        pipeline_model_parallel_size: number of GPUs used for pipeline model parallelism.
        virtual_pipeline_model_parallel_size: number of virtual stages (interleaved
                                              pipeline).
        pipeline_model_parallel_split_rank: for models with both encoder and decoder,
                                            rank in pipeline with split point.


    Let's say we have a total of 16 GPUs denoted by g0 ... g15 and we
    use 2 GPUs to parallelize the model tensor, and 4 GPUs to parallelize
    the model pipeline. The present function will
    create 8 tensor model-parallel groups, 4 pipeline model-parallel groups
    and 8 data-parallel groups as:
        8 data_parallel groups:
            [g0, g2], [g1, g3], [g4, g6], [g5, g7], [g8, g10], [g9, g11], [g12, g14], [g13, g15]
        8 tensor model-parallel groups:
            [g0, g1], [g2, g3], [g4, g5], [g6, g7], [g8, g9], [g10, g11], [g12, g13], [g14, g15]
        4 pipeline model-parallel groups:
            [g0, g4, g8, g12], [g1, g5, g9, g13], [g2, g6, g10, g14], [g3, g7, g11, g15]
    Note that for efficiency, the caller should make sure adjacent ranks
    are on the same DGX box. For example if we are using 2 DGX-1 boxes
    with a total of 16 GPUs, rank 0 to 7 belong to the first box and
    ranks 8 to 15 belong to the second box.
    """
    if torch.distributed.get_rank() == 0:
        print('> initializing tensor model parallel with size {}'.format(
            tensor_model_parallel_size_))
        print('> initializing pipeline model parallel with size {}'.format(
            pipeline_model_parallel_size_))
    # Get world size and rank. Ensure some consistencies.
    world_size = torch.distributed.get_world_size()
    tensor_model_parallel_size = min(tensor_model_parallel_size_, world_size)
    pipeline_model_parallel_size = min(pipeline_model_parallel_size_, world_size)
    ensure_divisibility(world_size,
                        tensor_model_parallel_size * pipeline_model_parallel_size)
    data_parallel_size = world_size // (tensor_model_parallel_size *
                                        pipeline_model_parallel_size)

    num_tensor_model_parallel_groups = world_size // tensor_model_parallel_size
    num_pipeline_model_parallel_groups = world_size // pipeline_model_parallel_size
    num_data_parallel_groups = world_size // data_parallel_size

    if virtual_pipeline_model_parallel_size_ is not None:
        global _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
        global _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
        _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK = 0
        _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = virtual_pipeline_model_parallel_size_

    if pipeline_model_parallel_split_rank_ is not None:
        global _PIPELINE_MODEL_PARALLEL_SPLIT_RANK
        _PIPELINE_MODEL_PARALLEL_SPLIT_RANK = pipeline_model_parallel_split_rank_

    rank = torch.distributed.get_rank()

    # Build the data-parallel groups.
    global _DATA_PARALLEL_GROUP
    all_data_parallel_group_ranks = []
    for i in range(pipeline_model_parallel_size):
        start_rank = i * num_pipeline_model_parallel_groups
        end_rank = (i + 1) * num_pipeline_model_parallel_groups
        for j in range(tensor_model_parallel_size):
            ranks = range(start_rank + j, end_rank,
                          tensor_model_parallel_size)
            all_data_parallel_group_ranks.append(list(ranks))
            group = torch.distributed.new_group(ranks)
            if rank in ranks:
                _DATA_PARALLEL_GROUP = group

    # Build the model-parallel groups.
    global _MODEL_PARALLEL_GROUP
    for i in range(data_parallel_size):
        ranks = [data_parallel_group_ranks[i]
                 for data_parallel_group_ranks in all_data_parallel_group_ranks]
        group = torch.distributed.new_group(ranks)
        if rank in ranks:
            _MODEL_PARALLEL_GROUP = group

    # Build the tensor model-parallel groups.
    global _TENSOR_MODEL_PARALLEL_GROUP
    for i in range(num_tensor_model_parallel_groups):
        ranks = range(i * tensor_model_parallel_size,
                      (i + 1) * tensor_model_parallel_size)
        group = torch.distributed.new_group(ranks)
        if rank in ranks:
            _TENSOR_MODEL_PARALLEL_GROUP = group

    # Build the pipeline model-parallel groups and embedding groups
    # (first and last rank in each pipeline model-parallel group).
    global _PIPELINE_MODEL_PARALLEL_GROUP
    global _PIPELINE_GLOBAL_RANKS
    global _EMBEDDING_GROUP
    global _EMBEDDING_GLOBAL_RANKS
    for i in range(num_pipeline_model_parallel_groups):
        ranks = range(i, world_size,
                      num_pipeline_model_parallel_groups)
        group = torch.distributed.new_group(ranks)
        if rank in ranks:
            _PIPELINE_MODEL_PARALLEL_GROUP = group
            _PIPELINE_GLOBAL_RANKS = ranks
        # Setup embedding group (to exchange gradients between
        # first and last stages).
        if len(ranks) > 1:
            embedding_ranks = [ranks[0], ranks[-1]]
            if pipeline_model_parallel_split_rank_ is not None and \
                    pipeline_model_parallel_split_rank_ not in embedding_ranks:
                embedding_ranks = [ranks[0],
                                   ranks[pipeline_model_parallel_split_rank_],
                                   ranks[-1]]
        else:
            embedding_ranks = ranks
        group = torch.distributed.new_group(embedding_ranks)
        if rank in embedding_ranks:
            _EMBEDDING_GROUP = group
        if rank in ranks:
            _EMBEDDING_GLOBAL_RANKS = embedding_ranks

0x03 切分样例

我们使用注释内容来进行学习如何切分模型,如何把多种并行模式组合在一起。

3.1 注释

initialize_model_parallel 的注释值得我们深入学习,具体如下:

Let's say we have a total of 16 GPUs denoted by g0 ... g15 and we
use 2 GPUs to parallelize the model tensor, and 4 GPUs to parallelize
the model pipeline. The present function will
create 8 tensor model-parallel groups, 4 pipeline model-parallel groups
and 8 data-parallel groups as:
    8 data_parallel groups:
        [g0, g2], [g1, g3], [g4, g6], [g5, g7], [g8, g10], [g9, g11], [g12, g14], [g13, g15]
    8 tensor model-parallel groups:
        [g0, g1], [g2, g3], [g4, g5], [g6, g7], [g8, g9], [g10, g11], [g12, g13], [g14, g15]
    4 pipeline model-parallel groups:
        [g0, g4, g8, g12], [g1, g5, g9, g13], [g2, g6, g10, g14], [g3, g7, g11, g15]
Note that for efficiency, the caller should make sure adjacent ranks
are on the same DGX box. For example if we are using 2 DGX-1 boxes
with a total of 16 GPUs, rank 0 to 7 belong to the first box and
ranks 8 to 15 belong to the second box.

从注释可以知道如下信息:

  • 假定目前有16个GPU,属于两个node,rank 0 ~7 属于第一个节点,rank 8 ~ 15 属于第二个节点。

  • create 8 tensor model-parallel groups, 4 pipeline model-parallel groups,这说明将一个完整模型切分如下:

    • 沿着行横向切了一刀:tensor_model_parallel_size = 16 / 8 = 2,就是2个 GPUs 来进行模型张量并行。
    • 沿着列纵向切了三刀:pipeline_model_parallel_size = 16 /4 = 4,就是4个GPUs 进行流水线并行。
    • 因此,一个模型分为8块,每一块放在一个GPU之上,就是8个GPU。而通过如下计算可以知 16 GPUs / 8 GPUs = 2 models。即,16张卡可以放置两个完整模型。
  • 因为张量模型并行组大小是2,即16个GPU被分成8组,则这8组内容是 [g0, g1], [g2, g3], [g4, g5], [g6, g7], [g8, g9], [g10, g11], [g12, g13], [g14, g15]。

  • 因为流水线并行组大小是4,即16个GPU被分成4组,则这4组内容是[g0, g4, g8, g12], [g1, g5, g9, g13], [g2, g6, g10, g14], [g3, g7, g11, g15]。

  • 因为数据并行组大小是2,16个GPU被分成8组,则这8组内容是[g0, g2], [g1, g3], [g4, g6], [g5, g7], [g8, g10], [g9, g11], [g12, g14], [g13, g15]。

  • 以上这些进程组都是通过 torch.distributed.new_group 来完成,这样组内进程之间就知道哪些进程是在同一个组内,是在一起训练的,也知道怎么通信。

3.2 切分情况

模型原始图如下

模型切分之后如下,一共被分成8块。其中,第一层被切分为 A,B,所以 A,B 之间就是 Tensor Model parallel。后面 C,D 之间也是 Tensor Model parallel,把两层都做了切分,依次类推。

我们的目标就是用代码来看看如何生成注释里面的各种模型组

3.3 切分策略

我们接下来看看具体切分的策略,也就是GPU分配策略。切分需要综合考虑多种情况,首先看看模型并行的通信状况。

  • 张量并行:通信发生在每层的前向传播和后向传播过程之中,通信类型是all-reduce,不但单次通信数据量大,并且通信频繁。
  • 流水线并行:通信在流水线阶段相邻的切分点之上,通信类型是P2P通信,单词通信数据量较少但是比较频繁,而且因为流水线的特点,会产生GPU空闲时间,这里称为流水线气泡(Bubble)。

我们接下来看看各种并行机制的对比。

  • Tensor versus Pipeline Parallelism. 张量模型的并行性在节点内是最好的,因为它会减少通信量。另一方面,流水线模型并行使用更便宜的点对点通信,可以跨节点执行,而不会限制整个计算。然而,流水线并行性会在流水线气泡中花费大量时间,因此,应限制流水线级的总数,以便流水线中的microbatches数量是流水线深度的合理倍数。当张量并行大小等于单个节点中的GPU数量时会达到峰值性能。
  • Pipeline versus Data Parallelism. 对于每个batch size,吞吐量随着流水线并行规模的增加而降低。流水线模型并行应该主要用于支持不适合单个 worker 的大型模型训练。而数据并行应该用于扩大训练规模。
  • Tensor versus Data Parallelism. 接下来看看数据和张量模型的并行性对性能的影响。在较大的批处理量和微批处理量为1的情况下,数据并行通信并不频繁;张量模型并行需要对批处理中的每个微批进行all-to-all通信。这种all-to-all的通信主导了端到端的训练时间,特别是当通信需要在多GPU节点上进行时。此外,随着张量模型并行规模的增加,我们在每个GPU上执行较小的矩阵乘法(因为会把模型张量进行切分),这降低了每个GPU的利用率。

最后看看结论

  • Tensor模型并行被用于intra-node transformer 层,因为张量并行计算密集且是耗费大量带宽,这样会在HGX based系统上高效运行。
  • Pipeline 模型并行主要被用于inter-node transformer 层,因为Pipeline 并行的通信带宽占用少,其可以有效利用集群中多网卡设计。
  • 数据并行则在前两者基础之上进行加持,使得训练可以扩展到更大规模和更快的速度。我们应该注意到,尽管数据并行可以带来高效的扩展,但我们不能单独使用数据并行来处理训练超大模型,因为 a)内存容量不足,b)数据并行的扩展限制。

3.4 实验

我们接下来做一个实验看看。

import torch

world_size = 16
tensor_model_parallel_size = 2 # 2 GPUs to parallelize the model tensor
pipeline_model_parallel_size = 4 # 4 GPUs to parallelize the model pipeline
data_parallel_size = world_size // (tensor_model_parallel_size *
                                    pipeline_model_parallel_size) # 2
num_tensor_model_parallel_groups = world_size // tensor_model_parallel_size # 8
num_pipeline_model_parallel_groups = world_size // pipeline_model_parallel_size # 4
num_data_parallel_groups = world_size // data_parallel_size # 8

# Build the data-parallel groups.
print("------ Build the data-parallel groups -----")
all_data_parallel_group_ranks = []
for i in range(pipeline_model_parallel_size):
    start_rank = i * num_pipeline_model_parallel_groups
    end_rank = (i + 1) * num_pipeline_model_parallel_groups
    for j in range(tensor_model_parallel_size):
        ranks = range(start_rank + j, end_rank,
                      tensor_model_parallel_size)
        all_data_parallel_group_ranks.append(list(ranks))
print(all_data_parallel_group_ranks)

# Build the model-parallel groups.
print("------ Build the model-parallel groups -----")
for i in range(data_parallel_size):
    ranks = [data_parallel_group_ranks[i]
             for data_parallel_group_ranks in all_data_parallel_group_ranks]
    print(list(ranks))

# Build the tensor model-parallel groups.
print("------ Build the tensor model-parallel groups -----")
for i in range(num_tensor_model_parallel_groups):
    ranks = range(i * tensor_model_parallel_size,
                  (i + 1) * tensor_model_parallel_size)
    print(list(ranks))

# Build the pipeline model-parallel groups and embedding groups
# (first and last rank in each pipeline model-parallel group).
print("------ Build the pipeline model-parallel groups -----")
for i in range(num_pipeline_model_parallel_groups):
    ranks = range(i, world_size,
                  num_pipeline_model_parallel_groups)
    print(list(ranks))

输出如下。需要注意,这里都是 GPU 的序列号,[0,2] 就是 [g0, g2]:

------ Build the data-parallel groups -----
[[0, 2], [1, 3], [4, 6], [5, 7], [8, 10], [9, 11], [12, 14], [13, 15]]
------ Build the model-parallel groups -----
[0, 1, 4, 5, 8, 9, 12, 13]
[2, 3, 6, 7, 10, 11, 14, 15]
------ Build the tensor model-parallel groups -----
[0, 1]
[2, 3]
[4, 5]
[6, 7]
[8, 9]
[10, 11]
[12, 13]
[14, 15]
------ Build the pipeline model-parallel groups -----
[0, 4, 8, 12]
[1, 5, 9, 13]
[2, 6, 10, 14]
[3, 7, 11, 15]

我们对比一下注释,发现代码打印结果可以和注释对应上:
    Let's say we have a total of 16 GPUs denoted by g0 ... g15 and we
    use 2 GPUs to parallelize the model tensor, and 4 GPUs to parallelize
    the model pipeline. The present function will
    create 8 tensor model-parallel groups, 4 pipeline model-parallel groups
    and 8 data-parallel groups as:
        8 data_parallel groups:
            [g0, g2], [g1, g3], [g4, g6], [g5, g7], [g8, g10], [g9, g11], [g12, g14], [g13, g15]
        8 tensor model-parallel groups:
            [g0, g1], [g2, g3], [g4, g5], [g6, g7], [g8, g9], [g10, g11], [g12, g13], [g14, g15]
        4 pipeline model-parallel groups:
            [g0, g4, g8, g12], [g1, g5, g9, g13], [g2, g6, g10, g14], [g3, g7, g11, g15]

我们接下来会进行具体分析。

0x04 起始状态

4.1 GPU 状况

从注释中可以看到:

Note that for efficiency, the caller should make sure adjacent ranks are on the same DGX box. For example if we are using 2 DGX-1 boxes with a total of 16 GPUs, rank 0 to 7 belong to the first box and ranks 8 to 15 belong to the second box.

意思就是:调用者需要确保相邻的rank在同一个节点上,我们例子有两个Node,其中第一个Node拥有 GPU 0 ~ 7,就是 rank 0 ~ 7,第二个Node是 GPU 8~15,就是 rank 8 ~ 15。

具体如下,这里每行4个GPU,是因为 4 GPUs to parallelize the model pipeline,所以流水线每个stage是4个GPU

4.2 符号说明

下面是论文之中提到的一些符号,这里有必要再取出来温习一下:

  • (𝑝, 𝑡, 𝑑): Parallelization dimensions.

  • 𝑝 for the pipeline-modelparallel size,

  • 𝑡 for the tensor-model-parallel size, and 𝑑 for the data-parallel size.

  • 𝑛: Number of GPUs. We require 𝑝 · 𝑡 · 𝑑 = 𝑛.

4.3 初始分组

依据注释,我们得出目前分组情况和一些全局信息。

  • 一共16个GPU,所以 world_size 为 16。就是 Notation 之中的 n。
  • 使用两个GPU进行 model tensor 并行,所以 tensor_model_parallel_size = 2。就是 Notation 之中的 t。
  • 使用四个GPU进行模型流水线并行,所以 pipeline_model_parallel_size = 4。就是 Notation 之中的 p。其实,就是流水线深度为 4,即,4 个 GPU 是串行的。
  • 依据上面定义,d = n / ( t * p) = 2,就是 data_parallel_size = 2。因为 t * p 就是一个模型所需要的 GPU,d = (总 GPU / 一个模型需要的 GPU),结果是这些GPU可以训练 d 个模型,就是可以用 d 个 mini-batches 进行这个 d个模型一起训练,所以数据并行度为 d。

接下来结合代码看看需要分成多少个process groups,他们在代码之中的变量是什么。

  • num_tensor_model_parallel_groups 就是从 tensor model 并行角度看,分成8 个进程roup。
  • num_pipeline_model_parallel_groups = world_size // pipeline_model_parallel_size 就是从 model 并行角度看,分成 4 个 进程group。
  • num_data_parallel_groups = world_size // data_parallel_size 就是从data 并行角度看,分成8 个 进程group。就是会有 8 个 DDP,每个 DDP 包括 2 个 rank。
  • 还有一个 _MODEL_PARALLEL_GROUP,

具体如下:

world_size = 16
tensor_model_parallel_size = 2 # 2 GPUs to parallelize the model tensor
pipeline_model_parallel_size = 4 # 4 GPUs to parallelize the model pipeline
data_parallel_size = world_size // (tensor_model_parallel_size *
                                    pipeline_model_parallel_size) # 2
num_tensor_model_parallel_groups = world_size // tensor_model_parallel_size # 8
num_pipeline_model_parallel_groups = world_size // pipeline_model_parallel_size # 4
num_data_parallel_groups = world_size // data_parallel_size # 8

0x05 Tensor model-parallel

本节我们分析的是,如何将 Node 上的 GPU 分给 tensor model 并行组。

5.1 分组

对于注释例子,16 / 2 = 8,分成 8 个进程组,每个组 两个 rank。这些分组分别是:[g0, g1], [g2, g3], [g4, g5], [g6, g7], [g8, g9], [g10, g11], [g12, g13], [g14, g15],我们得到了如下信息:

  • [g0, g1] 就是某一层分切为2半,分别被 g0, g1 来执行,[g2, g3] 表示另一层被分为两层,分别被 g2,g3 来执行。

  • 我们可以看到,每一个 tensor-model-parallel group的 rank一定是相邻的,比如 [g0, g1], [g2, g3]。

  • 注意,0 ~ 7 不代表是同一个模型。0 ~ 7 是同一个 Node 上的 GPU,这点容易被混淆。

我们再看看代码:

    # Build the tensor model-parallel groups.
    global _TENSOR_MODEL_PARALLEL_GROUP
    for i in range(num_tensor_model_parallel_groups): # 8
        ranks = range(i * tensor_model_parallel_size,
                      (i + 1) * tensor_model_parallel_size)
        group = torch.distributed.new_group(ranks) # 就有生成 8 组
        if rank in ranks: 
            # 如果本rank在某一list之中,即1 在 [0,1] 之中,则本 rank 就属于 new_group([0,1])
            _TENSOR_MODEL_PARALLEL_GROUP = group 

我们实验之中在这里得到:

------ Build the tensor model-parallel groups -----
[0, 1]
[2, 3]
[4, 5]
[6, 7]
[8, 9]
[10, 11]
[12, 13]
[14, 15]

对应我们图上如下,每个 tensor model group 用一个虚线小矩形框标示,一共8个:

_TENSOR_MODEL_PARALLEL_GROUP = group 就记录了本rank的进程组信息,比如 rank 2,它的 _TENSOR_MODEL_PARALLEL_GROUP 内容就是:group([g2, g3])。

5.2 使用

我们接下来看看如何使用。

get_tensor_model_parallel_group 返回了自己 rank 对应的 tensor model group。

def get_tensor_model_parallel_group():
    """Get the tensor model parallel group the caller rank belongs to."""
    return _TENSOR_MODEL_PARALLEL_GROUP

在 megatron/mpu/mappings.py 之中有对 tensor model group 的使用:

def _reduce(input_):
    """All-reduce the input tensor across model parallel group."""

    # Bypass the function if we are using only 1 GPU.
    if get_tensor_model_parallel_world_size()==1:
        return input_

    # All-reduce.
    torch.distributed.all_reduce(input_, group=get_tensor_model_parallel_group())

    return input_

就是当流水线反向传播时候,利用 _TENSOR_MODEL_PARALLEL_GROUP 进行在组内进行集合通信

0x06 Pipe-parallel

本节我们分析的是,如何将 Node 上的 GPU 分给 pipeline model 并行组。

6.1 分组

从注释可以看到,流水线分组就是把这个16个GPU 分成 4 组,每组 4 个 GPU,得到 [g0, g4, g8, g12], [g1, g5, g9, g13], [g2, g6, g10, g14], [g3, g7, g11, g15],我们得到了如下信息:

  • 每组的四个GPU进行模型流水线并行,所以 pipeline_model_parallel_size = 4。就是 Notation 之中的 p。其实,就是流水线深度为 4, 每组内 4 个 GPU 是串行的。即, [g0, g4, g8, g12] 这4个 GPU是串行的。

  • 再看看流水线的每一层,含有 16 / 4 = 4 个 GPU,能看到第一层是 0 ~ 4,第二层是 5 ~ 8,……。

  • 可以看到,流水线的 group是隔 n // p个取一个,比如[0, 4, 8, 12]。

  • 对于流水线每个stage,则是stage i 的 rank 范围是:[(i-1) * n//p, (i) * n//p],即 rank 2 所在的stage 的rank是 [0,1,2,3]。

  • _PIPELINE_MODEL_PARALLEL_GROUP 得到了本rank对应的流水线进程组。

  • _PIPELINE_GLOBAL_RANKS 得到了进程组的ranks。

  • 假如本进程是 rank 2,则流水线进程组 ranks 是 [g2, g6, g10, g14]

具体代码如下:

    # Build the pipeline model-parallel groups and embedding groups
    # (first and last rank in each pipeline model-parallel group).
    global _PIPELINE_MODEL_PARALLEL_GROUP
    global _PIPELINE_GLOBAL_RANKS
    global _EMBEDDING_GROUP
    for i in range(num_pipeline_model_parallel_groups): # 4
        ranks = range(i, world_size, # 每隔 n // p个取一个
                      num_pipeline_model_parallel_groups)
        group = torch.distributed.new_group(ranks)
        if rank in ranks:
            _PIPELINE_MODEL_PARALLEL_GROUP = group
            _PIPELINE_GLOBAL_RANKS = ranks
        # Setup embedding group (to exchange gradients between
        # first and last stages).
        if len(ranks) > 1:
            embedding_ranks = [ranks[0], ranks[-1]]
        else:
            embedding_ranks = ranks
        group = torch.distributed.new_group(embedding_ranks)
        if rank in embedding_ranks:
            _EMBEDDING_GROUP = group

我们拓展之前图如下,现在看到增加了 4 条从上到下的虚线箭头,分别对应了 4 组流水线串行。横向层是从 Stage 0 ~ Stage 3。

6.2 使用

接下来看看如何使用。

get_pipeline_model_parallel_group 返回了自己 rank 对应的 pipeline model group。

def get_pipeline_model_parallel_group():
    """Get the pipeline model parallel group the caller rank belongs to."""
    return _PIPELINE_MODEL_PARALLEL_GROUP

具体使用是在 megatron/p2p_communication.py,_communicate 之中会用流水线组信息来进行通信。这里省略了大部分代码。

def _communicate(tensor_send_next, tensor_send_prev, recv_prev, recv_next,
                 use_ring_exchange=False, tensor_shape=None,
                 override_scatter_gather_tensors_in_pipeline=False,
                 dtype_=None):
    """Communicate tensors between stages. Used as helper method in other
    communication methods that are used in megatron/schedules.py.
    """

    # Send tensors in both the forward and backward directions as appropriate.
    if use_ring_exchange: # 这里使用get_pipeline_model_parallel_group 进行通信
        torch.distributed.ring_exchange(tensor_send_prev=tensor_send_prev,
                                        tensor_recv_prev=tensor_recv_prev,
                                        tensor_send_next=tensor_send_next,
                                        tensor_recv_next=tensor_recv_next,
                                        group=mpu.get_pipeline_model_parallel_group())
    else:
        ops = []
        if tensor_send_prev is not None:
            send_prev_op = torch.distributed.P2POp(
                torch.distributed.isend, tensor_send_prev,
                mpu.get_pipeline_model_parallel_prev_rank()) # 得到流水线前一个rank
            ops.append(send_prev_op)
        if tensor_recv_prev is not None:
            recv_prev_op = torch.distributed.P2POp(
                torch.distributed.irecv, tensor_recv_prev,
                mpu.get_pipeline_model_parallel_prev_rank())
            ops.append(recv_prev_op)
        if tensor_send_next is not None:
            send_next_op = torch.distributed.P2POp(
                torch.distributed.isend, tensor_send_next,
                mpu.get_pipeline_model_parallel_next_rank()) # 得到流水线下一个rank
            ops.append(send_next_op)
        if tensor_recv_next is not None:
            recv_next_op = torch.distributed.P2POp(
                torch.distributed.irecv, tensor_recv_next,
                mpu.get_pipeline_model_parallel_next_rank())
            ops.append(recv_next_op)


6.2.1 上下游rank

具体如何得到流水线上下游的rank?是通过 get_pipeline_model_parallel_next_rank 和 get_pipeline_model_parallel_prev_rank 来完成。其中_PIPELINE_GLOBAL_RANKS 得到了进程组的ranks,假如本进程是 rank 2,则流水线进程组 ranks 是 [g2, g6, g10, g14]。

def get_pipeline_model_parallel_next_rank():
    rank_in_pipeline = get_pipeline_model_parallel_rank()
    world_size = get_pipeline_model_parallel_world_size()
    return _PIPELINE_GLOBAL_RANKS[(rank_in_pipeline + 1) % world_size]

def get_pipeline_model_parallel_prev_rank():
    rank_in_pipeline = get_pipeline_model_parallel_rank()
    world_size = get_pipeline_model_parallel_world_size()
    return _PIPELINE_GLOBAL_RANKS[(rank_in_pipeline - 1) % world_size]

6.2.2 world size

get_pipeline_model_parallel_world_size 得到了进程组的 world size。

def get_pipeline_model_parallel_world_size():
    """Return world size for the pipeline model parallel group."""
    global _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
    if _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE is not None:
        return _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
    return torch.distributed.get_world_size(group=get_pipeline_model_parallel_group())

0x07 Data-parallel

我们接下来看看数据并行。

7.1 分组

对于注释例子,16 / 2 = 8,分成 8 个进程组,每个组 两个 rank。这些分组分别是:[g0, g2], [g1, g3], [g4, g6], [g5, g7], [g8, g10], [g9, g11], [g12, g14], [g13, g15],我们得到了如下信息:

  • 依据上面分析, t * p 就是一个模型所需要的 GPU,因此,d = (总 GPU 数目 / 一个模型需要的 GPU 数目) = n / ( t * p),就是说,目前提供的这 n 个GPU可以同时训练 d 个模型,就是可以用 d 个 mini-batches 输入到这 d 个模型一起训练,所以数据并行度为 d。
  • 对应注释例子,就是data_parallel_size = 16 / (2 * 4) = 2。
  • rank 2 对应的数据并行进程组是[g0, g2]。

我们再看看用代码怎么确定有哪些group,每个group里面包含什么。

  • 首先,流水线被分成了 p 个 stage,对于流水线每个stage,其有 n // p 个GPU,stage i 的 rank 范围是:[i * n//p, (i+1) * n//p],即 rank 2所在的stage 的rank是 [0,1,2,3]。
  • 其次,在每一个stage之中,ranks = range(start_rank + j, end_rank, tensor_model_parallel_size) ,意思是这stage的n//p个GPUs中,每隔 t 个取一个作为数据并行 group 之中的一份子,因此每个data-parallel group大小为 n // p // t = d。

具体代码如下:

    # Build the data-parallel groups.
    global _DATA_PARALLEL_GROUP
    assert _DATA_PARALLEL_GROUP is None, \
        'data parallel group is already initialized'
    all_data_parallel_group_ranks = []
    for i in range(pipeline_model_parallel_size): # 遍历流水线深度
        start_rank = i * num_pipeline_model_parallel_groups # 找到每个stage的起始rank
        end_rank = (i + 1) * num_pipeline_model_parallel_groups # 找到每个stage的终止rank
        for j in range(tensor_model_parallel_size): # 遍历tensor model分组size
            ranks = range(start_rank + j, end_rank, # 每隔 t 个取一个作为数据并行group中的一份子
                          tensor_model_parallel_size)
            all_data_parallel_group_ranks.append(list(ranks))
            group = torch.distributed.new_group(ranks)
            if rank in ranks:
                _DATA_PARALLEL_GROUP = group

打印输出如下,和注释一致。

------ Build the data-parallel groups -----
[[0, 2], [1, 3], [4, 6], [5, 7], [8, 10], [9, 11], [12, 14], [13, 15]]

对应图片拓展如下:其中,每个新增的双箭头对应一个DDP(两个rank),比如[2, 3]对应一个DDP。

7.2 如何使用

我们接下来看看如何使用。

get_data_parallel_group 会得到本rank对应的 _DATA_PARALLEL_GROUP。

def get_data_parallel_group():
    """Get the data parallel group the caller rank belongs to."""
    return _DATA_PARALLEL_GROUP

在 allreduce_gradients之中,会对本数据并行组进行all-reduce。

    def allreduce_gradients(self):
        """Reduce gradients across data parallel ranks."""
        # If we have buffers, simply reduce the data in the buffer.
        if self._grad_buffers is not None:
            for _, buffer_ in self._grad_buffers.items():
                buffer_.data /= mpu.get_data_parallel_world_size() # 数据并行 world size
                torch.distributed.all_reduce(
                    buffer_.data, group=mpu.get_data_parallel_group()) # 数据并行组
        else:
            # Otherwise, bucketize and all-reduce
            buckets = {}
            # Pack the buckets.
            for param in self.module.parameters():
                if param.requires_grad and param.grad is not None:
                    tp = param.data.type()
                    if tp not in buckets:
                        buckets[tp] = []
                    buckets[tp].append(param)
                    param.main_grad = param.grad

            # For each bucket, all-reduce and copy all-reduced grads.
            for tp in buckets:
                bucket = buckets[tp]
                grads = [param.grad.data for param in bucket]
                coalesced = _flatten_dense_tensors(grads)
                coalesced /= mpu.get_data_parallel_world_size()
                torch.distributed.all_reduce(
                    coalesced, group=mpu.get_data_parallel_group())
                for buf, synced in zip(grads, _unflatten_dense_tensors(
                        coalesced, grads)):
                    buf.copy_(synced)

0x08 模型组

前面实验中,我们得到模型并行组如下:[0, 1, 4, 5, 8, 9, 12, 13] [2, 3, 6, 7, 10, 11, 14, 15]。生成代码如下:

    # Build the model-parallel groups.
    global _MODEL_PARALLEL_GROUP
    for i in range(data_parallel_size):
        ranks = [data_parallel_group_ranks[i]
                 for data_parallel_group_ranks in all_data_parallel_group_ranks]
        group = torch.distributed.new_group(ranks)
        if rank in ranks:
            _MODEL_PARALLEL_GROUP = group

_MODEL_PARALLEL_GROUP 会得到本rank对应的模型组。

def get_model_parallel_group():
    """Get the model parallel group the caller rank belongs to."""
    return _MODEL_PARALLEL_GROUP

这里是裁剪梯度会用到,就是在本模型的全部rank之中进行梯度裁剪相关操作。

def clip_grad_norm_fp32(parameters, max_norm, norm_type=2):
    """Clips gradient norm of an iterable of parameters whose gradients
       are in fp32.

    This is adapted from torch.nn.utils.clip_grad.clip_grad_norm_ and
    added functionality to handle model parallel parameters. Note that
    the gradients are modified in place.

    Arguments:
        parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a
            single Tensor that will have gradients normalized
        max_norm (float or int): max norm of the gradients
        norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for
            infinity norm.

    Returns:
        Total norm of the parameters (viewed as a single vector).
    """

    if isinstance(parameters, torch.Tensor):
        parameters = [parameters]

    # Filter parameters based on:
    #   - grad should not be none
    #   - parameter should not be shared
    #   - should not be a replica due to tensor model parallelism
    grads = []
    grads_for_norm = []
    for param in parameters:
        grad_not_none = param.grad is not None
        is_not_shared = param_is_not_shared(param)
        is_not_tp_duplicate = param_is_not_tensor_parallel_duplicate(param)
        grad = param.grad.detach()
        if grad_not_none:
            # Make sure the grads are in fp32
            grads.append(grad)
        if grad_not_none and is_not_shared and is_not_tp_duplicate:
            grads_for_norm.append(grad)

    # Norm parameters.
    max_norm = float(max_norm)
    norm_type = float(norm_type)
    total_norm = 0.0

    # Calculate norm.
    if norm_type == inf:
        total_norm = max(grad.abs().max() for grad in grads_for_norm)
        total_norm_cuda = torch.cuda.FloatTensor([float(total_norm)])
        # Take max across all model-parallel GPUs.
        torch.distributed.all_reduce(total_norm_cuda,
                                     op=torch.distributed.ReduceOp.MAX,
                                     group=mpu.get_model_parallel_group()) # 模型组信息
        total_norm = total_norm_cuda[0].item()

    else:
        if norm_type == 2.0:
            dummy_overflow_buf = torch.cuda.IntTensor([0])
            # Use apex's multi-tensor applier for efficiency reasons.
            # Multi-tensor applier takes a function and a list of list
            # and performs the operation on that list all in one kernel.
            grad_norm, _ = multi_tensor_applier(
                amp_C.multi_tensor_l2norm,
                dummy_overflow_buf,
                [grads_for_norm],
                False # no per-parameter norm
            )
            # Since we will be summing across data parallel groups,
            # we need the pow(norm-type).
            total_norm = grad_norm ** norm_type

        else:
            for grad in grads_for_norm:
                grad_norm = torch.norm(grad, norm_type)
                total_norm += grad_norm ** norm_type

        # Sum across all model-parallel GPUs.
        torch.distributed.all_reduce(total_norm,
                                     op=torch.distributed.ReduceOp.SUM,
                                     group=mpu.get_model_parallel_group()) # 模型组信息
        total_norm = total_norm.item() ** (1.0 / norm_type)

    # Scale.
    clip_coeff = max_norm / (total_norm + 1.0e-6)
    if clip_coeff < 1.0:
        dummy_overflow_buf = torch.cuda.IntTensor([0])
        multi_tensor_applier(amp_C.multi_tensor_scale,
                             dummy_overflow_buf,
                             [grads, grads],
                             clip_coeff)

    return total_norm

之前的图如下,利用看到分成两组,左边是Model 0 对应的全部ranks,右面是model 1 的ranks。

0x09 如何把模型分到GPU

我们最后还有一个问题没有涉及,就是如何把模型分块放到对应的GPU之上。就是如何与最初分成A,B,…, H 的那个图对应起来。其实,不是根据模型来把模型部分拷贝到对应的rank或者GPU,而是rank或者GPU主动过来拷贝自己对应的层

  • 因为调用了 mpu.initialize_model_parallel 来设置模型并行,数据并行等各种进程组,所以每个 rank 对应的进程都有自己的全局变量,具体其实就是进程自动就被映射到GPU上了。比如 rank 2 对应的进程在启动之后才知道自己是 rank 2,然后从初始化的全局变量之中知道自己的 data_parallel group 是 [g0, g2],tensor model-parallel group 是[g2, g3],pipeline model-parallel group 是 [g2, g6, g10, g14]。
  • ParallelTransformer 的初始化之中,offset 就是根据 rank 知道自己应该生成模型的那些层,然后通过 self.layers = torch.nn.ModuleList([build_layer(i + 1 + offset) for i in range(self.num_layers)]) 来生成对应的层。
  • get_model 方法也会根据自己的 pipeline rank 和 is_pipeline_first_stage 来知道是不是第一层或者最后一层,然后做相应处理。
  • 最后把模型参数拷贝到了自己对应的 GPU 之上。

具体 ParallelTransformer 初始化代码如下:

class ParallelTransformer(MegatronModule):
    """Transformer class."""

    def __init__(self, init_method, output_layer_init_method,
                 layer_type=LayerType.encoder,
                 self_attn_mask_type=AttnMaskType.padding,
                 pre_process=True, post_process=True):
        super(ParallelTransformer, self).__init__()
        args = get_args()
        
        # 省略代码
        
        # Transformer layers.
        def build_layer(layer_number):
            return ParallelTransformerLayer(
                init_method,
                output_layer_init_method,
                layer_number,
                layer_type=layer_type,
                self_attn_mask_type=self_attn_mask_type)
      
        # 下面 offset 就是根据rank知道自己应该生成模型的那些层
        if args.virtual_pipeline_model_parallel_size is not None:
            # Number of layers in each model chunk is the number of layers in the stage,
            # divided by the number of model chunks in a stage.
            self.num_layers = self.num_layers // args.virtual_pipeline_model_parallel_size
            # With 8 layers, 2 stages, and 4 model chunks, we want an assignment of
            # layers to stages like (each list is a model chunk):
            # Stage 0: [0]  [2]  [4]  [6]
            # Stage 1: [1]  [3]  [5]  [7]
            # With 8 layers, 2 stages, and 2 virtual stages, we want an assignment of
            # layers to stages like (each list is a model chunk):
            # Stage 0: [0, 1]  [4, 5]
            # Stage 1: [2, 3]  [6, 7]
            offset = mpu.get_virtual_pipeline_model_parallel_rank() * (
                args.num_layers // args.virtual_pipeline_model_parallel_size) + \
                (mpu.get_pipeline_model_parallel_rank() * self.num_layers)
        else:
            # Each stage gets a contiguous set of layers.
            offset = mpu.get_pipeline_model_parallel_rank() * self.num_layers

        self.layers = torch.nn.ModuleList(
            [build_layer(i + 1 + offset) for i in range(self.num_layers)])

        if self.post_process:
            # Final layer norm before output.
            self.final_layernorm = LayerNorm(
                args.hidden_size,
                eps=args.layernorm_epsilon)

所以,最终效果如下,其中同名子模块具有同样的参数,可以数据并行,即两个A可以数据并行。一列上的层之间可以流水线串行,比如 A–> C –> E –> G 就是串行,而一个横行4个是流水线的一个stage,其中从0开始,横向相邻两个GPU是 tensor model 并行。

0xFF 参考

[细读经典]Megatron论文和代码详细分析(2)

[细读经典]Megatron论文和代码详细分析(1)

Megatron-LM源码阅读(一)

Megatron-LM源码阅读(二)

megatron学习总结

GTC 2020: Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism

如何评价 NVIDIA 发布的 DGX-1?