{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Lora 实战"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step1 导入相关包"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/xhr/anaconda3/envs/llm/lib/python3.9/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n",
      "  from .autonotebook import tqdm as notebook_tqdm\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "'/CV/xhr/xhr_project/LLM_learn/transformers-code-master/self-llm'"
      ]
     },
     "execution_count": 1,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from datasets import Dataset\n",
    "from transformers import AutoTokenizer, AutoModelForCausalLM, DataCollatorForSeq2Seq, TrainingArguments, Trainer\n",
    "import pandas as pd"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step2 加载数据集"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 将JSON文件转换为CSV文件\n",
    "df = pd.read_json('../../dataset/huanhuan.json')\n",
    "ds = Dataset.from_pandas(df)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'instruction': ['小姐，别的秀女都在求中选，唯有咱们小姐想被撂牌子，菩萨一定记得真真儿的——',\n",
       "  '这个温太医啊，也是古怪，谁不知太医不得皇命不能为皇族以外的人请脉诊病，他倒好，十天半月便往咱们府里跑。',\n",
       "  '嬛妹妹，刚刚我去府上请脉，听甄伯母说你来这里进香了。'],\n",
       " 'input': ['', '', ''],\n",
       " 'output': ['嘘——都说许愿说破是不灵的。', '你们俩话太多了，我该和温太医要一剂药，好好治治你们。', '出来走走，也是散心。']}"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "ds[:3]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step3 数据集预处理"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "ChatGLMTokenizer(name_or_path='/CV/xhr/xhr_project/LLM_learn/transformers-code-master/model/chatglm3-6b', vocab_size=64798, model_max_length=1000000000000000019884624838656, is_fast=False, padding_side='left', truncation_side='right', special_tokens={'eos_token': '</s>', 'unk_token': '<unk>', 'pad_token': '<unk>'}, clean_up_tokenization_spaces=False),  added_tokens_decoder={\n",
       "\t\n",
       "}"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tokenizer = AutoTokenizer.from_pretrained(\"/root/autodl-tmp/ZhipuAI/chatglm3-6b\", trust_remote_code=True)\n",
    "tokenizer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[64790, 64792, 35182, 55671, 31123, 34752, 55276, 54740, 32595, 54806, 54538, 54878, 31123, 37963, 35662, 36028, 54695, 54732, 60136, 49127, 31123, 34856, 31781, 33498, 54792, 54792, 40972, 16747]\n"
     ]
    }
   ],
   "source": [
    "print(tokenizer.encode(ds[0]['instruction']))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "([64794, 30910, 13, 42579, 34526, 34975, 33690, 32587, 35524, 621, 52339],\n",
       " '<|system|> \\n 现在你要扮演皇帝身边的女人--甄嬛')"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "demo_token = tokenizer.build_single_message('system', \"\", \"现在你要扮演皇帝身边的女人--甄嬛\")\n",
    "demo_token, tokenizer.decode(demo_token)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(64790, '')"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tokenizer.get_command(\"[gMASK]\"), tokenizer._convert_id_to_token(+tokenizer.eos_token_id)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 调用api处理数据"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "instruction = \"\\n\".join([ds[0][\"instruction\"], ds[0][\"input\"]]).strip()     # query\n",
    "instruction = tokenizer.build_chat_input(instruction, history=[], role=\"user\")\n",
    "response = tokenizer(\"\\n\" + ds[0][\"output\"], add_special_tokens=False)\n",
    "input_ids = instruction[\"input_ids\"][0].numpy().tolist() + response[\"input_ids\"] + [tokenizer.eos_token_id]\n",
    "tokenizer.decode(input_ids)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 手动拆解数据处理"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "prompt = [tokenizer.get_command(\"<|system|>\")] + tokenizer.encode(\"现在你要扮演皇帝身边的女人--甄嬛\\n \", add_special_tokens=False)\n",
    "instruction_ = [tokenizer.get_command(\"<|user|>\")] + tokenizer.encode(\"\\n \" + \"\\n\".join([ds[0][\"instruction\"], ds[0][\"input\"]]).strip(), add_special_tokens=False,max_length=512) + [tokenizer.get_command(\"<|assistant|>\")]\n",
    "instruction = tokenizer.encode(prompt + instruction_)\n",
    "response = tokenizer.encode(\"\\n\" + ds[0][\"output\"], add_special_tokens=False)\n",
    "input_ids = instruction + response + [tokenizer.eos_token_id]\n",
    "tokenizer.decode(input_ids)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "def process_func(example):\n",
    "    MAX_LENGTH = 512\n",
    "    input_ids, labels = [], []\n",
    "    prompt = [tokenizer.get_command(\"<|system|>\")] + tokenizer.encode(\"现在你要扮演皇帝身边的女人--甄嬛\\n \", add_special_tokens=False)\n",
    "    instruction_ = [tokenizer.get_command(\"<|user|>\")] + tokenizer.encode(\"\\n \" + \"\\n\".join([example[\"instruction\"], example[\"input\"]]).strip(), add_special_tokens=False,max_length=512) + [tokenizer.get_command(\"<|assistant|>\")]\n",
    "    instruction = tokenizer.encode(prompt + instruction_)\n",
    "    response = tokenizer.encode(\"\\n\" + example[\"output\"], add_special_tokens=False)\n",
    "    input_ids = instruction + response + [tokenizer.eos_token_id]\n",
    "    labels = [tokenizer.pad_token_id] * len(instruction) + response + [tokenizer.eos_token_id]\n",
    "    pad_len = MAX_LENGTH - len(input_ids)\n",
    "    # print()\n",
    "    input_ids += [tokenizer.pad_token_id] * pad_len\n",
    "    labels += [tokenizer.pad_token_id] * pad_len\n",
    "    labels = [(l if l != tokenizer.pad_token_id else -100) for l in labels]\n",
    "\n",
    "    return {\n",
    "        \"input_ids\": input_ids,\n",
    "        \"labels\": labels\n",
    "    }"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Map: 100%|██████████| 3729/3729 [00:00<00:00, 4034.25 examples/s]\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "Dataset({\n",
       "    features: ['input_ids', 'labels'],\n",
       "    num_rows: 3729\n",
       "})"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tokenized_ds = ds.map(process_func, remove_columns=ds.column_names)\n",
    "tokenized_ds"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "tokenizer.decode(tokenized_ds[1][\"input_ids\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "tokenizer.decode(list(filter(lambda x: x != -100, tokenized_ds[1][\"labels\"])))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step4 创建模型"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Loading checkpoint shards: 100%|██████████| 7/7 [00:03<00:00,  1.96it/s]\n"
     ]
    }
   ],
   "source": [
    "model = AutoModelForCausalLM.from_pretrained(\"/root/autodl-tmp/ZhipuAI/chatglm3-6b\", trust_remote_code=True, low_cpu_mem_usage=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Lora"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### PEFT Step1 配置文件"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- target_modules也可以传入正则项,比如以h.1结尾的query_key_value：\".*\\.1.*query_key_value\"  \n",
    "- modules_to_save指定的是除了拆成lora的模块，其他的模块可以完整的指定训练。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "LoraConfig(peft_type=<PeftType.LORA: 'LORA'>, auto_mapping=None, base_model_name_or_path=None, revision=None, task_type=<TaskType.CAUSAL_LM: 'CAUSAL_LM'>, inference_mode=False, r=8, target_modules={'query_key_value'}, lora_alpha=32, lora_dropout=0.0, fan_in_fan_out=False, bias='none', modules_to_save=None, init_lora_weights=True, layers_to_transform=None, layers_pattern=None, rank_pattern={}, alpha_pattern={})"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from peft import LoraConfig, TaskType, get_peft_model\n",
    "\n",
    "# model = AutoModelForCausalLM.from_pretrained(\"/root/autodl-tmp/ZhipuAI/chatglm3-6b\", low_cpu_mem_usage=True)\n",
    "config = LoraConfig(task_type=TaskType.CAUSAL_LM, target_modules={\"query_key_value\"}, r=8, lora_alpha=32)\n",
    "config"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### PEFT Step2 创建模型"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = get_peft_model(model, config)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "trainable params: 1,949,696 || all params: 6,245,533,696 || trainable%: 0.031217444255383614\n"
     ]
    }
   ],
   "source": [
    "model.print_trainable_parameters()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step5 配置训练参数"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Data collator\n",
    "data_collator = DataCollatorForSeq2Seq(\n",
    "    tokenizer,\n",
    "    model=model,\n",
    "    label_pad_token_id=-100,\n",
    "    pad_to_multiple_of=None,\n",
    "    padding=False\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "args = TrainingArguments(\n",
    "    output_dir=\"./huanhuan\",\n",
    "    per_device_train_batch_size=1,\n",
    "    gradient_accumulation_steps=8,\n",
    "    logging_steps=20,\n",
    "    num_train_epochs=1\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step6 创建训练器"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "trainer = Trainer(\n",
    "    model=model,\n",
    "    args=args,\n",
    "    train_dataset=tokenized_ds,\n",
    "    data_collator=data_collator,\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step7 模型训练"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "trainer.train()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step8 模型推理"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 61,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'[gMASK]sop <|system|>\\n现在你要扮演皇帝身边的女人--甄嬛\\n<|user|>\\n 你是谁？<|assistant|>\\n 我是甄嬛，家父是大理寺少卿甄远道。'"
      ]
     },
     "execution_count": 61,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model = model.cuda()\n",
    "ipt = tokenizer(\"<|system|>\\n现在你要扮演皇帝身边的女人--甄嬛\\n<|user|>\\n {}\\n{}\".format(\"你是谁？\", \"\").strip() + \"<|assistant|>\\n\", return_tensors=\"pt\").to(model.device)\n",
    "tokenizer.decode(model.generate(**ipt, max_length=128, do_sample=True)[0], skip_special_tokens=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "transformers",
   "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": "3.9.18"
  },
  "orig_nbformat": 4
 },
 "nbformat": 4,
 "nbformat_minor": 2
}