[LLM] DeepSpeed distributed training framework

1. Introduction to DeepSpeed

1. Distributed background introduction

• In a distributed computing environment, the master node is responsible for coordinating the work of other nodes and processes
• Accelerate, an official distributed training tool provided by pytorch, only supports nvlink, while graphics cards such as T4 and 3090 are PIX, detection method: nvidia-smi topo -m; deepspeed supports larger-scale model training
• Mixed precision training
• ZeRO can reduce memory usage, optimize large model training, and divide model parameters into three parts: Optimizer States, Gradient, and Model Parameter. When using ZeRO for distributed training, different optimization techniques such as ZeRO-Offload and ZeRO-Stage3 can be selected.

2. Introduction to deepspeed

• In DeepSpeed, you can enable BF16 mixed precision training by setting "bf16.enabled": true in the configuration file to reduce memory usage.
  • Mixed-precision training refers to the technique of using both FP16 (half-precision floating-point number) and FP32 (single-precision floating-point number) precision during training.
• deepspeed can choose the appropriate communication library according to the specific situation. For example, for distributed training on a CPU cluster, you can choose mpi and gloo; if you want to perform distributed training on a GPU, you can choose nccl.
  • mpi is a cross-node communication library, often used for distributed training on CPU clusters;
  • Gloo is a high-performance distributed training framework that supports distributed training on CPU and GPU;
  • nccl is a GPU-specific communication library provided by NVIDIA, which is widely used in distributed training on GPU.
• DeepSpeed's core technology:
  • Zero (Zero Redundancy Optimizer, 3D optimization and unloading): in deepspeed through zero_optimization.stage=0/1/2/3the setting, unloading through zero_optimization.offload_optimizer.devicethe setting
• DeepSpeed's reasoning optimization technology:
  • Deep fusion: as shown in the figure below, the red dotted box is the optimized Kernel in this unit, and the corresponding number is the optimized efficiency multiple
  • Inference-customized GeMM

Two, deepspeed+transformer code combat

1. Preprocessing and Json files

• The first is to use huggingface's datasets.mapsample custom operation on the dataset; transformers can trainerintegrate the deepspeed function. This usage requires a configuration file, such as the deepspeed configuration ds_config.jsonfile below. For the specific configuration of this config, please refer to the documentation .
• The FLAN-T5 model used here; start deepspeed: deepspeed --include=localhost:1,2 train.py, start the first two graphics cards; note that enough memory is required to use ZeRO3
• If not used trianerto integrate deepspeed, core functions such as from_pretrained and from_config should contain important parts of DeepSpeed, such as zero. It should be stage3 or higher when initializing Zero. Reference documentation .

{
    
    
  "bf16": {
    
    
    "enabled": "auto"
  },
  "optimizer": {
    
    
    "type": "AdamW",
    "params": {
    
    
      "lr": "auto",
      "betas": "auto",
      "eps": "auto",
      "weight_decay": "auto"
    }
  },
  "scheduler": {
    
    
    "type": "WarmupLR",
    "params": {
    
    
      "warmup_min_lr": "auto",
      "warmup_max_lr": "auto",
      "warmup_num_steps": "auto"
    }
  },
  "zero_optimization": {
    
    
    "stage": 3,
    "offload_optimizer": {
    
    
      "device": "cpu",
      "pin_memory": true
    },
    "offload_param": {
    
    
      "device": "cpu",
      "pin_memory": true
    },
    "overlap_comm": true,
    "contiguous_gradients": true,
    "sub_group_size": 1e9,
    "reduce_bucket_size": "auto",
    "stage3_prefetch_bucket_size": "auto",
    "stage3_param_persistence_threshold": "auto",
    "stage3_max_live_parameters": 1e9,
    "stage3_max_reuse_distance": 1e9,
    "stage3_gather_16bit_weights_on_model_save": false
  },
  "gradient_accumulation_steps": "auto",
  "gradient_clipping": "auto",
  "steps_per_print": 2000,
  "train_batch_size": "auto",
  "train_micro_batch_size_per_gpu": "auto",
  "wall_clock_breakdown": false
}

2. Training code

• Data: samsum dataset
• Model: google/flan-t5-xxl large model

# !/usr/bin/python
# -*- coding: utf-8 -*-
"""
@Author    : guomiansheng
@Software  : Pycharm
@Contact   : [email protected]
@File      : deepspeed_test.py
"""
import nltk
import torch
import evaluate
import datasets
import numpy as np
from nltk.tokenize import sent_tokenize
from torch.utils.data import DataLoader
from torch.nn.utils.rnn import pad_sequence
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from transformers import Seq2SeqTrainer, Seq2SeqTrainingArguments

nltk.download("punkt")

dataset_name = "samsum" # 数据集名称
model_name="google/flan-t5-xxl" # 模型名称
max_input_length = 512
max_gen_length = 128
output_dir = "checkpoints"
num_train_epochs = 5
learning_rate = 5e-5
deepspeed_config = "./ds_config.json" # deepspeed配置文件
per_device_train_batch_size=1 # batch size设置为1，因为太大导致OOM
per_device_eval_batch_size=1
gradient_accumulation_steps=2 # 由于单卡的batch size为1，为了扩展batch size，使用梯度累加

tokenizer = AutoTokenizer.from_pretrained(model_name)

# 加载数据
dataset = datasets.load_dataset(dataset_name)
print(dataset["train"][0])

# tokenize
def preprocess(examples):
    dialogues = ["summarize:" + dia for dia in examples["dialogue"]]
    # summaries = [summ for summ in examples["summary"]]
    model_inputs = tokenizer(dialogues, max_length=max_input_length, truncation=True)
    labels = tokenizer(text_target=examples["summary"], max_length=max_gen_length, truncation=True)
    model_inputs["labels"] = labels["input_ids"]
    return model_inputs

tokenized_dataset = dataset.map(preprocess, batched=True, remove_columns=["dialogue", "summary", "id"])
# print(tokenized_dataset["train"]["input_ids"][0]) # 打印结果


# 对batch进行padding
def collate_fn(features):
    batch_input_ids = [torch.LongTensor(feature["input_ids"]) for feature in features]
    batch_attention_mask = [torch.LongTensor(feature["attention_mask"]) for feature in features]
    batch_labels = [torch.LongTensor(feature["labels"]) for feature in features]

    batch_input_ids = pad_sequence(batch_input_ids, batch_first=True, padding_value=tokenizer.pad_token_id)
    batch_attention_mask = pad_sequence(batch_attention_mask, batch_first=True, padding_value=0)
    batch_labels = pad_sequence(batch_labels, batch_first=True, padding_value=-100)

    return {
    
    
        "input_ids": batch_input_ids,
        "attention_mask": batch_attention_mask,
        "labels": batch_labels
    }
# 用于测试的代码
# dataloader = DataLoader(tokenized_dataset["test"], shuffle=False, batch_size=4, collate_fn=collate_fn)
# batch = next(iter(dataloader))
# print(batch)


# 加载模型
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
# 用于测试的代码
# dataloader = DataLoader(tokenized_dataset["test"], shuffle=False, batch_size=4, collate_fn=collate_fn)
# batch = next(iter(dataloader))
# output = model(**batch)
# print(output)


# 定义评估函数
metric = evaluate.load("rouge")

def compute_metrics(eval_preds):
    preds, labels = eval_preds
    if isinstance(preds, tuple):
        preds = preds[0]
    decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
    labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
    decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
    decoded_preds = ["\n".join(sent_tokenize(pred.strip())) for pred in decoded_preds]
    decoded_labels = ["\n".join(sent_tokenize(label.strip())) for label in decoded_labels]
    result = metric.compute(predictions=decoded_preds, references=decoded_labels, use_stemmer=True)
    result = {
    
    k: round(v * 100, 4) for k, v in result.items()}
    prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
    result["gen_len"] = np.mean(prediction_lens)
    return result


# 设置训练参数
training_args = Seq2SeqTrainingArguments(
    output_dir=output_dir,
    per_device_train_batch_size=per_device_train_batch_size,
    per_device_eval_batch_size=per_device_eval_batch_size,
    gradient_accumulation_steps=gradient_accumulation_steps,
    eval_accumulation_steps=1, # 防止评估时导致OOM
    predict_with_generate=True,
    fp16=False,
    learning_rate=learning_rate,
    num_train_epochs=num_train_epochs,
    # logging & evaluation strategies
    logging_dir="logs",
    logging_strategy="steps",
    logging_steps=50, # 每50个step打印一次log
    evaluation_strategy="steps",
    eval_steps=500, # 每500个step进行一次评估
    save_steps=500,
    save_total_limit=2,
    load_best_model_at_end=True,
    deepspeed=deepspeed_config, # deepspeed配置文件的位置
    report_to="all"
)


# 模型训练
trainer = Seq2SeqTrainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset["train"],
    eval_dataset=tokenized_dataset["validation"],
    data_collator=collate_fn,
    compute_metrics=compute_metrics,
)

trainer.train()
# 打印验证集上的结果
print(trainer.evaluate(tokenized_dataset["validation"]))
# 打印测试集上的结果
print(trainer.evaluate(tokenized_dataset["test"]))
# 保存最优模型
trainer.save_model("best")

Accelerated training methods: quantization toolkit bitsandbytes, deepspeed (first read torch.distributed and ColossalAI are working on it), llama.cpp quantization model

3. Deepspeed accelerates Bloom lora fine-tuning

1. Configuration file

{
    
    
  "train_micro_batch_size_per_gpu": "auto",
  "gradient_accumulation_steps": "auto",
  "steps_per_print": 50,
  "gradient_clipping": 1.0,
  "zero_optimization": {
    
    
    "stage": 2,
    "offload_optimizer": {
    
    
            "device": "cpu"
    },
    "contiguous_gradients": true,
    "overlap_comm": true
  },
  "zero_allow_untested_optimizer": true,
  "fp16": {
    
    
    "enabled": true,
    "loss_scale": 0,
    "loss_scale_window": 1000,
    "hysteresis": 2,
    "min_loss_scale": 1
  },
  "optimizer": {
    
    
    "type": "Adam",
    "params": {
    
    
      "lr": "auto",
      "betas": "auto",
      "eps": "auto",
      "weight_decay": "auto"
    }
  },
  "activation_checkpointing": {
    
    
    "partition_activations": true,
    "contiguous_memory_optimization": true
  },
  "wall_clock_breakdown": false
}

2. Training code

• Data: fine-tuning data using 1 million instructions provided by BELLE
• Model: bloomz-7b1-mt model

#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
@Author : andy
@Date   : 2023/7/10 10:07
@Contact: [email protected] 
@File   : bloom_lora.py 
"""
import os
import torch
import random
import datasets
import numpy as np
from tqdm import tqdm
from typing import Dict
from torch.utils.data import DataLoader
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    DataCollatorForSeq2Seq,
    TrainingArguments,
    Trainer
)
from peft import (
    LoraConfig,
    TaskType,
    get_peft_model,
    get_peft_model_state_dict,
    set_peft_model_state_dict
)

def set_random_seed(seed):
    if seed is not None and seed > 0:
        random.seed(seed)
        np.random.seed(seed)
        torch.manual_seed(seed)
        torch.random.manual_seed(seed)
        torch.cuda.manual_seed(seed)
        torch.cuda.manual_seed_all(seed)
        torch.backends.cudnn.deterministic = True

set_random_seed(1234)

# 1. 设置参数
# LoRA参数
LORA_R = 8
LORA_ALPHA = 32
LORA_DROPOUT = 0.1
# 训练参数
EPOCHS=3
LEARNING_RATE=5e-5
OUTPUT_DIR="./checkpoints"
BATCH_SIZE=4 # 2
GRADIENT_ACCUMULATION_STEPS=3
# 其他参数
MODEL_PATH = "bigscience/bloomz-7b1-mt"
DATA_PATH = "./data/belle_open_source_1M.train.json"
MAX_LENGTH = 512
PATTERN = "{}\n{}"
DS_CONFIG = "ds_zero2_config.json"
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH) # 加载tokenizer
# 加载数据
dataset = datasets.load_dataset("json", data_files=DATA_PATH)
# print(dataset["train"][0])


# 2. tokenize
def tokenize(text: str, add_eos_token=True):
    result = tokenizer(
        text,
        truncation=True,
        max_length=MAX_LENGTH,
        padding=False,
        return_tensors=None)
    # 判断是否要添加eos_token
    if (result["input_ids"][-1] != tokenizer.eos_token_id
        and len(result["input_ids"]) < MAX_LENGTH
        and add_eos_token):
        result["input_ids"].append(tokenizer.eos_token_id)
        result["attention_mask"].append(1)
    result["labels"] = result["input_ids"].copy()
    return result


def preprocess(example: Dict, train_on_inputs: bool = False):
    prompt = example["input"]
    response = example["target"]
    text = PATTERN.format(prompt, response)
    tokenized_inp = tokenize(text)
    # 若train_on_inputs为False，则将label中与input相关的token替换为-100
    if not train_on_inputs:
        tokenized_prompt = tokenize(prompt,add_eos_token=False)
        prompt_tokens_len = len(tokenized_prompt["input_ids"])
        tokenized_inp["labels"] = [-100]*prompt_tokens_len + tokenized_inp["labels"][prompt_tokens_len:]
    return tokenized_inp


train_data = dataset["train"].shuffle().map(preprocess, remove_columns=["id", "input", "target"])
print(train_data[0])

# pad_to_multiple_of=8表示padding的长度是8的倍数
collate_fn = DataCollatorForSeq2Seq(tokenizer, pad_to_multiple_of=8, return_tensors="pt", padding=True)

# 2. 加载模型
evice_map = {
    
    "": int(os.environ.get("LOCAL_RANK") or 0)}
# device_map指定模型加载的GPU;troch_dtype=torch.float16表示半精度加载模型
model = AutoModelForCausalLM.from_pretrained(MODEL_PATH, torch_dtype=torch.float16, device_map=device_map)


# 3. LoRA相关
lora_config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    inference_mode=False,
    r=LORA_R, # LoRA中低秩近似的秩
    lora_alpha=LORA_ALPHA, # 见上文中的低秩矩阵缩放超参数
    lora_dropout=LORA_DROPOUT, # LoRA层的dropout
)
# 转换模型
model = get_peft_model(model, lora_config)
model.config.use_cache = False
old_state_dict = model.state_dict
model.state_dict = (
    lambda self, *_, **__: get_peft_model_state_dict(self, old_state_dict())
).__get__(model, type(model))
# 打印模型中的可训练参数
model.print_trainable_parameters()


# 4. 训练参数
args = TrainingArguments(
    output_dir=OUTPUT_DIR, # checkpoint的存储目录
    per_device_train_batch_size=BATCH_SIZE, # 单设备上的batch size
    gradient_accumulation_steps=GRADIENT_ACCUMULATION_STEPS, # 梯度累加的step数
    warmup_steps=100,
    num_train_epochs=EPOCHS,
    learning_rate=LEARNING_RATE,
    fp16=True, # 使用混合精度训练
    logging_steps=50,
    evaluation_strategy="no", # 不进行评估
    save_strategy="steps",
    save_steps=2000, # 保存checkpoint的step数
    save_total_limit=5, # 最多保存5个checkpoint
    deepspeed=DS_CONFIG
)


# 5. 模型训练
trainer = Trainer(
    model=model,
    train_dataset=train_data,
    eval_dataset=None,
    args=args,
    data_collator=collate_fn
)
trainer.train()
model.save_pretrained("best_model")

deepspeed --include=localhost:0,1,2,3 train.pystart up.

Reference

[1] The framework of large-scale model training
[2] NLP large-scale language model fine-tuning practice: DeepSpeed+Transformers realizes simple and fast start-up tens of billions of parameter model fine-tuning
[3] Straightforward illustration of GPT2 model Self Attention attention mechanism: implementation process and MTB Language model core code reading summary
[4] Essential knowledge of language model distributed parallel training: Collective communication operation and Pytorch example
[5] Key technologies for large-scale language model training: mixed precision training, video memory analysis and DeepSpeed ​​distributed training practice
[ 6] Tensor parallel tools in large model training must read: Megatron-DeepSpeed ​​tool code mpu detailed explanation and practice
[7] https://github.com/microsoft/DeepSpeed
​​[8] deepspeed official tutorial: https://www. deepspeed.ai/getting-started/
[9] DeepSpeed ​​accelerates large-scale model inference through system optimization
[10] DeepSpeed: Accelerating large-scale model inference and training via system optimizations and compression. Microsoft Research Blog
[11]Graphical large model training: Pipeline Parallelism (Pipeline Parallelism), taking Gpipe as an example
[12] Where do Q, K, and V in the deep learning attention mechanism come from?
[13] "ZeRO: Memory Optimizations Toward Training Trillion Parameter Models "
[14] [Distributed training] DeepSpeed: AllReduce and ZeRO-DP
[15] How to generate text: using different decoding methods for language generation with Transformers
[16] Detailed explanation of Transformer model (the most complete version of the diagram). First knowledge of CV
[ 17] Large (language) model reasoning principle and acceleration. Jin
[18] Transformers DeepSpeed ​​official document interpretation
[19] NVIDIA Megatron-LM: Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism
[20] https:// github.com/NVIDIA/Megatron-LM