分布式优化器

Author: Wenxuan Tan, Junwen Duan, Renjie Mao

相关论文

介绍

除了广泛采用的Adam和SGD外，许多现代优化器需要逐层统计信息以有效更新参数，因此无法直接应用于模型层在多个设备上分片的并行设置。我们以提供了优化的分布式实现，，并且通过plugin与Tensor Parallel、DDP和ZeRO无缝集成。

优化器

Adafactor 是一种首次采用非负矩阵分解（NMF）的 Adam 变体，用于减少内存占用。CAME 通过引入一个置信度矩阵来改进 NMF 的效果。GaLore 通过将梯度投影到低秩空间，并使用 8 位块状量化进一步减少内存占用。Lamb 允许使用巨大的批量大小而不失准确性，通过按其 Lipschitz 常数的倒数界定的逐层自适应更新实现

使用

现在我们展示如何使用分布式 Adafactor 与 booster API 结合 Tensor Parallel 和 ZeRO 2。即使您不使用distributed optimizer，plugin 也会自动将optimizer转换为分布式版本以方便使用。

step 1. 导包

from transformers import LlamaModel, LlamaConfig
from colossalai.nn.optimizer.distributed_adafactor import DistributedAdaFactor
from colossalai.booster import Booster
from colossalai.booster.plugin import HybridParallelPlugin
import colossalai
import torch

step 2. 初始化分布式

我们需要先初始化分布式环境. 为了展示, 我们使用 colossal run --nproc_per_node 4. 更多初始化方式请参考 Launch Colossal-AI

colossalai.launch_from_torch()

step 3. 初始化模型和优化器

configuration = LlamaConfig()
model = LlamaModel(configuration).cuda()
criterion = lambda x: x.mean()
dist_optim = DistributedAdaFactor(model.parameters())

step 4.初始化booster和plugin

plugin = HybridParallelPlugin(tp_size=2, zero_stage=2, pp_size=1, enable_all_optimization=True)
booster = Booster(plugin=plugin)
# You should also pass in your own dataset.
model, dist_optim, criterion, dataloader, _ = booster.boost(model, dist_optim, criterion)

step 5.训练

steps = 10
for step in range(steps):
    input_ids = torch.ones(1, 100, device="cuda", dtype=torch.int)
    attention_mask = input_ids.clone()
    outputs = model(input_ids.cuda(), attention_mask.cuda())
    loss = criterion(outputs.last_hidden_state)
    booster.backward(loss, dist_optim)
    dist_optim.step()
    dist_optim.zero_grad()

GaLore的特殊初期

对于 GaLore，我们需要为每个参数组指定投影rank，以及量化和分页优化器参数。有关量化的详细信息，请参考 bitandbytes.

from colossalai.nn.optimizer.galore import get_galore_param_groups
from colossalai.nn.optimizer import DistGaloreAwamW
optim = DistGaloreAwamW(
    get_galore_param_groups(model, decay=1e-2, rank=8),
    lr=lr,
    betas=(beta1, beta2),
    eps=eps,
    nbits=8,
    percentile_clipping=100,
    block_wise=True,
    min_8bit_size=4096,
)

兼容性

Model/Feature	Lamb	GaLore	Adafactor	CAME
Hybrid Parallel Plugin	✔️	✔️	✔️	✔️
Low Level Zero Plugin	✔️	❌	✔️	✔️
Torch DDP Plugin	✔️	✔️	✔️	✔️
Gemini Plugin	❌	❌	❌	❌
Moe Hybrid Plugin	❌	❌	❌	❌

API 参考

class

colossalai.nn.DistributedAdaFactor

(params, lr = None, eps = (1e-30, 0.001), clip_threshold = 1.0, decay_rate = -0.8, beta1 = None, weight_decay = 0.0, scale_parameter = True, relative_step = True, warmup_init = False)

Description

function

setup_distributed

(tp_group: ProcessGroup = None, dp_group: ProcessGroup = None, shard_to_working_param: typing.Dict = {}, padding_map = None, use_zero: bool = True)

Parameters

tp_group -- The devices group for tensor parallel; dp_group -- The devices group for data parallel;

shard_to_working_param (Dict) -- ZeRO 2 feeds the optimizer a sharded param view as grads are sharded. This maps from id(view) to working params used in forward & backward. padding_map -- An empty interface placeholder; use_zero -- Whether or not to use zero;

Description

Setup process groups for TP and ZeRO 2. Inject features to the Optimizer

function

step

(closure = None)

Parameters

closure (callable, optional) -- A closure that reevaluates the model and returns the loss.

Description

Performs a single optimization steps

class

colossalai.nn.DistributedLamb

(params, lr = 0.001, betas = (0.9, 0.999), eps = 1e-06, weight_decay = 0, bias_correction = True)

Parameters

params (iterable) -- iterable of parameters to optimize or dicts defining parameter groups
lr (float, optional) -- learning rate (default: 1e-3)
betas (Tuple[float, float], optional) -- coefficients used for computing running averages of gradient and its square (default: (0.9, 0.999))
eps (float, optional) -- term added to the denominator to improve numerical stability (default: 1e-8)
weight_decay (float, optional) -- weight decay (L2 penalty) (default: 0)

Description

Implements the Lamb algorithm, with extra support for ZeRO 2 and Tensor Parallel. Proposed in `Large Batch Optimization for Deep Learning: Training BERT in 76 minutes`_. It's recommended to use this with HybridParallelPlugin/ZeRO plugin and booster, which will take care of setup_distributed. Example with 4 devices: >>> optim = DistributedLamb(model.parameters(), lr=1e-3) >>> proc_mesh = ProcessGroupMesh(tp_size, zero_size) >>> tp_group = proc_mesh.get_group_along_axis(0) >>> dp_group = proc_mesh.get_group_along_axis(1) >>> optim.setup_distributed(tp_group, dp_group)

.. _Large Batch Optimization for Deep Learning: Training BERT in 76 minutes: https://arxiv.org/abs/1904.00962

function

setup_distributed

(tp_group: typing.Optional[torch.distributed.distributed_c10d.ProcessGroup] = None, dp_group: typing.Optional[torch.distributed.distributed_c10d.ProcessGroup] = None, shard_to_working_param: typing.Optional[typing.Dict] = {}, padding_map = None, is_zero: typing.Optional[bool] = False)

Parameters

tp_group (dist.ProcessGroup) -- Tensor Parallel process group
dp_group (dist.ProcessGroup) -- ZeRO 2 process group
shard_to_working_param (Dict) -- ZeRO 2 feeds the optimizer a sharded param view as grads are sharded. This maps from id(view) to working params used in forward & backward. padding_map -- An empty interface placeholder
is_zero (bool) -- Whether to use ZeRO 2.

Description

Assign process groups for TP and ZeRO 2.

function

step

(closure = None)

Parameters

closure (callable, optional) -- A closure that reevaluates the model and returns the loss.

Description

Performs a single optimization step.

class

colossalai.nn.DistGaloreAwamW

(params, lr = 0.01, betas = (0.9, 0.999), eps = 1e-08, weight_decay = 0.01, nbits = 8, min_8bit_size = 4096, percentile_clipping = 100, block_wise = True, is_paged = False, args = None)

Parameters

params (iterable) -- iterable of parameters to optimize or dicts defining parameter groups.
lr (float, optional) -- learning rate. (default: 1e-3)
betas (Tuple[float, float], optional) -- coefficients used for computing running averages of gradient and its norm. (default: (0.9, 0.999))
eps (float, optional) -- term added to the denominator to improve numerical stability. (default: 1e-6)
weight_decay (float, optional) -- weight decay (L2 penalty) (default: 0.01) nbits -- Number of bits for quantization optim states. Only 32 and 8 are supported.
min_8bit_size (int, defaults to 4096) -- The minimum number of elements of the parameter tensors for 8-bit optimization.
percentile_clipping (int, defaults to 100) -- Adapts clipping threshold automatically by tracking the last 100 gradient norms and clipping the gradient at a certain percentile to improve stability.
block_wise (bool, defaults to True) -- Whether to independently quantize each block of tensors to reduce outlier effects and improve stability.
is_paged (bool, defaults to False) -- Whether the optimizer is a paged optimizer (handle memory spike via CPU-GPU transfer) or not.
args (dict, optional) -- quantization-related arguments. If passed, will override all quantization args above.

Description

Implements Galore, a optimizer-agonistic gradient compression technique on 8-bit AdamW. It largely compresses gradient via low-rank projection and is claimed to be insensitive to hyperparams like lr. Supports Tensor Parallel and ZeRO stage 1 and 2 via booster and plugin. Proposed in `GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection` https://arxiv.org/abs/2403.03507

function

setup_distributed

Parameters

tp_group (dist.ProcessGroup) -- Tensor Parallel process group
dp_group (dist.ProcessGroup) -- ZeRO 2 process group
shard_to_working_param (Dict) -- ZeRO 2 feeds the optimizer a sharded param view as grads are sharded. This maps from id(view) to working params used in forward & backward.
padding_map (Dict) -- Padding size of each param from ZeRO's param store. Required if ZeRO is used.
is_zero (bool) -- Whether to use ZeRO 2.

Description

Setup process groups for TP and ZeRO 2.

function

step

(closure = None)

Parameters

closure (callable, optional) -- A closure that reevaluates the model and returns the loss.

Description

Performs a single optimization step.

function

to_master_shape

(data, padding)

Description

Pad to master (optimizer) param shape

class

colossalai.nn.DistributedCAME

(params, lr = None, eps = (1e-30, 1e-16), clip_threshold = 1.0, betas = (0.9, 0.999, 0.9999), weight_decay = 0.0)

Parameters

params (iterable) -- iterable of parameters to optimize or dicts defining parameter groups
lr (float, optional) -- external learning rate (default: None)
eps (tuple[float, float]) -- regularization constants for square gradient and instability respectively (default: (1e-30, 1e-16))
clip_threshold (float) -- threshold of root-mean-square of final gradient update (default: 1.0)
betas (tuple[float, float, float]) -- coefficient used for computing running averages of
update, square gradient and instability (default -- (0.9, 0.999, 0.9999)))
weight_decay (float, optional) -- weight decay (L2 penalty) (default: 0)

Description

Implements CAME algorithm. This implementation is based on: `CAME: Confidence-guided Adaptive Memory Efficient Optimization`

function

setup_distributed

(tp_group: ProcessGroup = None, dp_group: ProcessGroup = None, shard_to_working_param: typing.Dict = {}, padding_map = None, use_zero: bool = True)

Parameters

tp_group -- The devices group for tensor parallel; dp_group -- The devices group for data parallel;

shard_to_working_param (Dict) -- ZeRO 2 feeds the optimizer a sharded param view as grads are sharded. This maps from id(view) to working params used in forward & backward. padding_map -- Interface placeholder use_zero -- Whether or not to use zero;

Description

Inject features to the Optimizer

function

step

(closure = None)

Parameters

closure (callable, optional) -- A closure that reevaluates the model and returns the loss.

Description

Performs a single optimization step.

分布式优化器

介绍​

优化器​

使用​

step 1. 导包​

step 2. 初始化分布式​

step 3. 初始化模型和优化器​

step 4.初始化booster和plugin​

step 5.训练​

GaLore的特殊初期​

兼容性​

API 参考​

class

colossalai.nn.DistributedAdaFactor

function

setup_distributed

function

step

class

colossalai.nn.DistributedLamb

function

setup_distributed

function

step

class

colossalai.nn.DistGaloreAwamW

function

setup_distributed

function

step

function

to_master_shape

class

colossalai.nn.DistributedCAME

function

setup_distributed

function

step

介绍

优化器

使用

step 1. 导包

step 2. 初始化分布式

step 3. 初始化模型和优化器

step 4.初始化booster和plugin

step 5.训练

GaLore的特殊初期

兼容性

API 参考