{
 "cells": [
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 使用 DeepSpeed 和 Hugging Face Transformer 微调 FLAN-T5 XL/XXL\n",
    "\n",
    "[Scaling Instruction-Finetuned Language Models](https://arxiv.org/pdf/2210.11416.pdf) 论文发布了 FLAN-T5 模型，它是 T5 模型的增强版。FLAN-T5 由很多各种各样的任务微调而得，因此，简单来讲，它就是个方方面面都更优的 T5 模型。相同参数量的条件下，FLAN-T5 的性能相比 T5 而言有两位数的提高。Google 在 Hugging Face 上开源了 [5 个 FLAN-T5 的 checkpoints](https://huggingface.co/models?other=arxiv:2210.11416)，参数量范围从 8000 万 到 110 亿。\n",
    "\n",
    "在之前的一篇博文中，我们已经学习了如何 [针对聊天对话数据摘要生成任务微调 FLAN-T5](https://www.philschmid.de/fine-tune-flan-t5)，那时我们使用的是 [Base (250M参数)](https://huggingface.co/google/flan-t5-base)模型。本文，我们将研究如何将训练从 Base 扩展到 [XL (30 亿参数)](https://huggingface.co/google/flan-t5-xl) 或 [XXL (110 亿参数)](https://huggingface.co/google/flan-t5-xxl)。\n",
    "\n",
    "这意味着我们将学习如何利用模型并行、多 GPU 以及 [DeepSpeed ZeRO](https://www.deepspeed.ai/tutorials/zero/) 来微调 FLAN-T5 XL 和 XXL。\n",
    "\n",
    "除了作为教程的部分之外，我们还跑了一系列实验，这些实验数据可以帮助你选择正确的硬件设置。你可以在*结果和实验*部分找到详细信息。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "# install git lfs for pushing artifacts\n",
    "!sudo apt install git-lfs\n",
    "# install torch with the correct cuda version, check nvcc --version\n",
    "!pip install torch --extra-index-url https://download.pytorch.org/whl/cu116 --upgrade\n",
    "# install Hugging Face Libraries\n",
    "!pip install \"transformers==4.26.0\" \"datasets==2.9.0\" \"accelerate==0.16.0\" \"evaluate==0.4.0\" --upgrade\n",
    "# install deepspeed and ninja for jit compilations of kernels\n",
    "!pip install \"deepspeed==0.8.0\" ninja --upgrade\n",
    "# install additional dependencies needed for training\n",
    "!pip install rouge-score nltk py7zr tensorboard"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 处理数据集"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "与[针对聊天对话的摘要生成任务微调 FLAN-T5](https://www.philschmid.de/fine-tune-flan-t5)一文中类似，我们需要先准备一个用于微调的数据集。本文，我们将在 [CNN Dailymail 数据集](https://huggingface.co/datasets/cnn_dailymail) 上微调 [FLAN-T5-XXL](https://huggingface.co/google/flan-t5-xxl)。我们不会赘述如何生成数据集，如果你想了解数据集生成的详细步骤，请参阅[上一篇文章](https://www.philschmid.de/fine-tune-flan-t5)。\n",
    "\n",
    "我们定义了一些参数，本文的示例都会基于这些参数，但你可以根据实际需要进行调整。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 实验配置\n",
    "model_id = \"google/flan-t5-xxl\" # Hugging Face 模型 Id\n",
    "dataset_id = \"cnn_dailymail\" # Hugging Face 数据集 Id\n",
    "dataset_config = \"3.0.0\" # 数据集版本\n",
    "save_dataset_path = \"data\" # 存放处理后数据的本地路径\n",
    "text_column = \"article\" # 输入文本所属列\n",
    "summary_column = \"highlights\" # 输出文本所属列\n",
    "# 定制指令提示格式\n",
    "prompt_template = f\"Summarize the following news article:\\n{{input}}\\nSummary:\\n\""
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "与 [之前的示例](https://www.philschmid.de/fine-tune-flan-t5) 不同，这次我们把预处理和训练分开。这样我们就可以在非 GPU 实例上运行预处理。我们先对数据集进行预处理（即分词）并将其保存到磁盘，然后训练脚本再从磁盘中加载预处理后的数据集。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from datasets import load_dataset\n",
    "from transformers import AutoTokenizer\n",
    "import numpy as np \n",
    "\n",
    "# Load dataset from the hub\n",
    "dataset = load_dataset(dataset_id,name=dataset_config)\n",
    "# Load tokenizer of FLAN-t5-base\n",
    "tokenizer = AutoTokenizer.from_pretrained(model_id)\n",
    "\n",
    "print(f\"Train dataset size: {len(dataset['train'])}\")\n",
    "print(f\"Test dataset size: {len(dataset['test'])}\")\n",
    "\n",
    "# Train dataset size: 287113\n",
    "# Test dataset size: 11490"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "我们在配置文件中定义了一个 `prompt_template`，其可用于来构建指令提示，以提高我们模型的性能。 `prompt_template` 有“固定”的开始词和结束词，文档放在中间。这意味着我们需要确保 *“固定”模板词 + 文档* 总长不超过模型支持的最大序列长度。因此我们需要计算模型支持的最大文档长度，稍后我们会根据它来填充或截断模板中的文档。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Prompt length: 12\n",
      "Max input length: 500\n"
     ]
    }
   ],
   "source": [
    "prompt_length = len(tokenizer(prompt_template.format(input=\"\"))[\"input_ids\"])\n",
    "max_sample_length = tokenizer.model_max_length - prompt_length\n",
    "print(f\"Prompt length: {prompt_length}\")\n",
    "print(f\"Max input length: {max_sample_length}\")\n",
    "\n",
    "# Prompt length: 12\n",
    "# Max input length: 500"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "现在我们知道，模型支持的最大输入文档长度为 500。除了输入之外，我们还需要知道最大“目标”序列长度，我们可以通过遍历数据集中的摘要长度来得到。（代码需要运行几分钟）"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/json": {
       "ascii": false,
       "bar_format": null,
       "colour": null,
       "elapsed": 0.012465238571166992,
       "initial": 0,
       "n": 0,
       "ncols": null,
       "nrows": null,
       "postfix": null,
       "prefix": "",
       "rate": null,
       "total": 299,
       "unit": "ba",
       "unit_divisor": 1000,
       "unit_scale": false
      },
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "32577879b38640f898e798ea8f88a801",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/299 [00:00<?, ?ba/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Max source length: 500\n"
     ]
    },
    {
     "data": {
      "application/json": {
       "ascii": false,
       "bar_format": null,
       "colour": null,
       "elapsed": 0.011892318725585938,
       "initial": 0,
       "n": 0,
       "ncols": null,
       "nrows": null,
       "postfix": null,
       "prefix": "",
       "rate": null,
       "total": 299,
       "unit": "ba",
       "unit_divisor": 1000,
       "unit_scale": false
      },
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "724cc7afe0ba49a3b8a6a763a189e380",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/299 [00:00<?, ?ba/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Max target length: 129\n"
     ]
    }
   ],
   "source": [
    "from datasets import concatenate_datasets\n",
    "import numpy as np\n",
    "\n",
    "\n",
    "# The maximum total input sequence length after tokenization. \n",
    "# Sequences longer than this will be truncated, sequences shorter will be padded.\n",
    "tokenized_inputs = concatenate_datasets([dataset[\"train\"], dataset[\"test\"]]).map(lambda x: tokenizer(x[text_column], truncation=True), batched=True, remove_columns=[text_column, summary_column])\n",
    "max_source_length = max([len(x) for x in tokenized_inputs[\"input_ids\"]])\n",
    "max_source_length = min(max_source_length, max_sample_length)\n",
    "print(f\"Max source length: {max_source_length}\")\n",
    "\n",
    "# The maximum total sequence length for target text after tokenization. \n",
    "# Sequences longer than this will be truncated, sequences shorter will be padded.\"\n",
    "tokenized_targets = concatenate_datasets([dataset[\"train\"], dataset[\"test\"]]).map(lambda x: tokenizer(x[summary_column], truncation=True), batched=True, remove_columns=[text_column, summary_column])\n",
    "target_lenghts = [len(x) for x in tokenized_targets[\"input_ids\"]]\n",
    "# use 95th percentile as max target length\n",
    "max_target_length = int(np.percentile(target_lenghts, 95))\n",
    "print(f\"Max target length: {max_target_length}\")"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "现在一切准备就绪，可以处理数据集了。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "\n",
    "def preprocess_function(sample, padding=\"max_length\"):\n",
    "    # created prompted input\n",
    "    inputs = [prompt_template.format(input=item) for item in sample[text_column]]\n",
    "\n",
    "    # tokenize inputs\n",
    "    model_inputs = tokenizer(inputs, max_length=tokenizer.model_max_length, padding=padding, truncation=True)\n",
    "\n",
    "    # Tokenize targets with the `text_target` keyword argument\n",
    "    labels = tokenizer(text_target=sample[summary_column], max_length=max_target_length, padding=padding, truncation=True)\n",
    "\n",
    "    # If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore\n",
    "    # padding in the loss.\n",
    "    if padding == \"max_length\":\n",
    "        labels[\"input_ids\"] = [\n",
    "            [(l if l != tokenizer.pad_token_id else -100) for l in label] for label in labels[\"input_ids\"]\n",
    "        ]\n",
    "\n",
    "    model_inputs[\"labels\"] = labels[\"input_ids\"]\n",
    "    return model_inputs\n",
    "\n",
    "# process dataset\n",
    "tokenized_dataset = dataset.map(preprocess_function, batched=True, remove_columns=list(dataset[\"train\"].features))\n",
    "\n",
    "# save dataset to disk\n",
    "tokenized_dataset[\"train\"].save_to_disk(os.path.join(save_dataset_path,\"train\"))\n",
    "tokenized_dataset[\"test\"].save_to_disk(os.path.join(save_dataset_path,\"eval\"))"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 使用 `deepspeed` 微调模型\n",
    "\n",
    "准备完毕！我们现在可以开始训练模型了！如前所述，我们将使用集成了 DeepSpeed 的 Hugging Face Trainer。因此我们需要创建一个 `deespeed_config.json`。[DeepSpeed 配置](https://www.deepspeed.ai/docs/config-json/) 定义了要使用的 ZeRO 策略以及是否要使用混合精度训练等配置项。 Hugging Face Trainer 允许我们从 `deepspeed_config.json` 中的 `TrainingArguments` 继承相关配置以避免重复设置，查看[文档了解更多信息](https://huggingface.co/docs/transformers/v4.26.1/en/main_classes/deepspeed#configuration)。\n",
    "\n",
    "我们创建了 4 组 deepspeed 配置文件用于实验，包括 `CPU 卸载`和`混合精度`：\n",
    "\n",
    "- [ds_flan_t5_z3_config.json](https://github.com/philschmid/deep-learning-pytorch-huggingface/blob/main/training/configs/ds_flan_t5_z3_config.json)\n",
    "- [ds_flan_t5_z3_config_bf16.json](https://github.com/philschmid/deep-learning-pytorch-huggingface/blob/main/training/configs/ds_flan_t5_z3_config_bf16.json)\n",
    "- [ds_flan_t5_z3_offload.json](https://github.com/philschmid/deep-learning-pytorch-huggingface/blob/main/training/configs/ds_flan_t5_z3_offload.json)\n",
    "- [ds_flan_t5_z3_offload_bf16.json](https://github.com/philschmid/deep-learning-pytorch-huggingface/blob/main/training/configs/ds_flan_t5_z3_offload_bf16.json)\n",
    "\n",
    "你可以根据你的运行环境选择，例如如果在 NVIDIA V100s 上运行，你就不能使用带 `bf16` 的配置，因为 V100 不支持 `bfloat16` 数据类型。\n",
    "\n",
    "> 在微调 `T5` 模型时，不能使用 `fp16`，因为它会导致精度溢出问题，参见：[#4586](https://github.com/huggingface/transformers/issues/4586)，[#10830](https://github.com/huggingface/transformers/issues/10830), [#10956](https://github.com/huggingface/transformers/pull/10956)\n",
    ">\n",
    "\n",
    "如开头所述，我们使用的是 p4dn.24xlarge AWS EC2 实例，该实例包含 8 张显存为 40GB 的 NVIDIA A100。这意味着我们可以使用 `bf16`，它将减少近一半的模型显存占用，使我们能够在不卸载的情况下高效训练。\n",
    "\n",
    "我们将使用 [ds_flan_t5_z3_config_bf16.json](https://github.com/philschmid/deep-learning-pytorch-huggingface/blob/main/training/configs/ds_flan_t5_z3_config_bf16.json)。如果你不想用 `auto` 值，可以查看 [文档](https://huggingface.co/docs/transformers/v4.26.1/en/main_classes/deepspeed#configuration)。"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "```\n",
    "{\n",
    "  \"bf16\": {\n",
    "    \"enabled\": \"auto\"\n",
    "  },\n",
    "  \"optimizer\": {\n",
    "    \"type\": \"AdamW\",\n",
    "    \"params\": {\n",
    "      \"lr\": \"auto\",\n",
    "      \"betas\": \"auto\",\n",
    "      \"eps\": \"auto\",\n",
    "      \"weight_decay\": \"auto\"\n",
    "    }\n",
    "  },\n",
    "  \"scheduler\": {\n",
    "    \"type\": \"WarmupLR\",\n",
    "    \"params\": {\n",
    "      \"warmup_min_lr\": \"auto\",\n",
    "      \"warmup_max_lr\": \"auto\",\n",
    "      \"warmup_num_steps\": \"auto\"\n",
    "    }\n",
    "  },\n",
    "  \"zero_optimization\": {\n",
    "    \"stage\": 3,\n",
    "    \"overlap_comm\": true,\n",
    "    \"contiguous_gradients\": true,\n",
    "    \"sub_group_size\": 1e9,\n",
    "    \"reduce_bucket_size\": \"auto\",\n",
    "    \"stage3_prefetch_bucket_size\": \"auto\",\n",
    "    \"stage3_param_persistence_threshold\": \"auto\",\n",
    "    \"stage3_max_live_parameters\": 1e9,\n",
    "    \"stage3_max_reuse_distance\": 1e9,\n",
    "    \"stage3_gather_16bit_weights_on_model_save\": false\n",
    "  },\n",
    "  \"gradient_accumulation_steps\": \"auto\",\n",
    "  \"gradient_clipping\": \"auto\",\n",
    "  \"steps_per_print\": 2000,\n",
    "  \"train_batch_size\": \"auto\",\n",
    "  \"train_micro_batch_size_per_gpu\": \"auto\",\n",
    "  \"wall_clock_breakdown\": false\n",
    "}\n",
    "```"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "现在，该训练脚本上场了。我们根据[之前的博文](https://www.philschmid.de/fine-tune-flan-t5)准备了一个 [run_seq2seq_deepspeed.py](https://github.com/philschmid/deep-learning-pytorch-huggingface/blob/main/training/scripts/run_seq2seq_deepspeed.py) 训练脚本，它支持我们配置 deepspeed 和其他超参数，包括 `google/flan-t5-xxl` 的模型 ID。\n",
    "\n",
    "我们使用 `deepspeed` 启动器触发训练，输入给启动器的参数包括 GPU 数量、deepspeed 配置及其它超参数（如 `google/flan-t5-xxl` 的模型 ID）。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\n",
      "To disable this warning, you can either:\n",
      "\t- Avoid using `tokenizers` before the fork if possible\n",
      "\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\n",
      "deepspeed --num_gpus=8 scripts/run_seq2seq_deepspeed.py --model_id google/flan-t5-xxl --dataset_path data --epochs 3 --per_device_train_batch_size 8 --per_device_eval_batch_size 8 --generation_max_length 129 --lr 1e-4 --deepspeed configs/ds_flan_t5_z3_config_bf16.json\n"
     ]
    }
   ],
   "source": [
    "!deepspeed --num_gpus=8 scripts/run_seq2seq_deepspeed.py \\\n",
    "    --model_id $model_id \\\n",
    "    --dataset_path $save_dataset_path \\\n",
    "    --epochs 3 \\\n",
    "    --per_device_train_batch_size 8 \\\n",
    "    --per_device_eval_batch_size 8 \\\n",
    "    --generation_max_length $max_target_length \\\n",
    "    --lr 1e-4 \\\n",
    "    --deepspeed configs/ds_flan_t5_z3_config_bf16.json "
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "DeepSpeed 先将模型加载到 CPU 上，然后将其拆分到 8 张 A100 上然后开始训练。使用 [CNN Dailymail 数据集](https://huggingface.co/datasets/cnn_dailymail)进行训练大约需要 10 个小时，费用约为 `322 美元`。"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 结果与实验\n",
    "\n",
    "为了更好地了解硬件要求，我们对 FLAN-T5 XL 和 XXL 进行了一系列实验，以帮助我们评估和了解硬件需求以及训练这些模型的成本。\n",
    "\n",
    "下表列出了实验和相关设置的详细信息。\n",
    "\n",
    "数据集: `\"cnn_dailymail\"`\n",
    "-  训练样本数: `287113`\n",
    "-  验证样本数: `13368`\n",
    "\n",
    "超参:\n",
    "- epochs: `3`\n",
    "- 学习率: `1e-4`\n",
    "\n",
    "运行环境设置: \n",
    "- 4x V100 16GB: p3.8xlarge\n",
    "- 4x A10G 24GB: g5.24xlarge\n",
    "- 8x V100 16GB: p3.16xlarge\n",
    "- 8x A100 40GB: p4dn.24xlarge\n",
    "\n",
    "\n",
    "| 模型             | DeepSpeed 卸载 | 硬件     | GPU每卡batch size | 精度 | 时长 | 费用   |\n",
    "|-------------------|------------|--------------|--------------------|-----------|----------|--------|\n",
    "| FLAN-T5-XL (3B)   | No         | 4x V100 16GB | OOM                | fp32      | -        | -      |\n",
    "| FLAN-T5-XL (3B)   | No         | 8x V100 16GB | 1                  | fp32      | 105h     | ~$2570 |\n",
    "| FLAN-T5-XL (3B)   | No         | 8x A100 40GB | 72                 | bf16     |   2.5h       | ~$81       |\n",
    "| FLAN-T5-XL (3B)   | Yes        | 4x V100 16GB | 8                  | fp32      | 69h      | ~$828  |\n",
    "| FLAN-T5-XL (3B)   | Yes        | 8x V100 16GB | 8                  | fp32      | 32h      | ~$768  |\n",
    "| FLAN-T5-XXL (11B) | No        | 8x A100 40GB | 8                | bf16      | 10h        | ~$322      |\n",
    "| FLAN-T5-XXL (11B) | Yes        | 4x V100 16GB | OOM                | fp32      | -        | -      |\n",
    "| FLAN-T5-XXL (11B) | Yes        | 8x V100 16GB | OOM                | fp32      | -        | -      |\n",
    "| FLAN-T5-XXL (11B) | Yes        | 4x A10G 24GB | 24                | bf16      | 90h      | ~$732  |\n",
    "| FLAN-T5-XXL (11B) | Yes        | 8x A100 40GB | 48                | bf16      | 19h      | ~$613  |\n",
    "\n",
    "我们可以看到 `bf16` 与 `fp32` 相比具有显著优势。FLAN-T5-XXL 能放进 4 张 A10G (24GB)，但放不进 8 张 V100 16GB。\n",
    "\n",
    "我们的实验还表明，如果模型可以无需卸载同时以 batch size 大于 4 的配置跑在 GPU 上，其速度将比卸载模型和减小 batch size 的配置快约 2 倍且更具成本效益。"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "> 英文原文: <url> https://www.philschmid.de/fine-tune-flan-t5-deepspeed </url>\n",
    "> 原文作者：Philipp Schmid\n",
    "> 译者: Matrix Yao (姚伟峰)，英特尔深度学习工程师，工作方向为 transformer-family 模型在各模态数据上的应用及大规模模型的训练推理。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.8"
  },
  "orig_nbformat": 4,
  "vscode": {
   "interpreter": {
    "hash": "916dbcbb3f70747c44a77c7bcd40155683ae19c65e1c03b4aa3499c5328201f1"
   }
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}