{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "19e284d2-74ad-4b0c-86d4-8279fcd69e8e",
   "metadata": {},
   "source": [
    "# 探究非对称量化和对称量化 - INT8\n",
    "\n",
    "> 指导文章：[17. 浅谈 RTN 模型量化：非对称 vs 对称](https://github.com/Hoper-J/AI-Guide-and-Demos-zh_CN/blob/master/Guide/17.%20浅谈%20RTN%20模型量化：非对称%20vs%20对称.md)\n",
    "\n",
    "在线链接：[Kaggle](https://www.kaggle.com/code/aidemos/15-int8) | [Colab](https://colab.research.google.com/drive/1jssltiV0avGynLPHKbRrKypy9W223fm-?usp=sharing)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1a8ce26a-af40-4356-9752-e8440b6973ae",
   "metadata": {},
   "source": [
    "## 安装库"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "b2059734-fbe3-412d-8b73-2d94a96af741",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Looking in indexes: http://mirrors.aliyun.com/pypi/simple\n",
      "Requirement already satisfied: bitsandbytes in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (0.49.2)\n",
      "Requirement already satisfied: torch<3,>=2.3 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from bitsandbytes) (2.7.1)\n",
      "Requirement already satisfied: numpy>=1.17 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from bitsandbytes) (2.4.4)\n",
      "Requirement already satisfied: packaging>=20.9 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from bitsandbytes) (26.1)\n",
      "Requirement already satisfied: filelock in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (3.29.0)\n",
      "Requirement already satisfied: typing-extensions>=4.10.0 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (4.15.0)\n",
      "Requirement already satisfied: setuptools in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (82.0.1)\n",
      "Requirement already satisfied: sympy>=1.13.3 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (1.14.0)\n",
      "Requirement already satisfied: networkx in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (3.6.1)\n",
      "Requirement already satisfied: jinja2 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (3.1.6)\n",
      "Requirement already satisfied: fsspec in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (2025.3.0)\n",
      "Requirement already satisfied: nvidia-cuda-nvrtc-cu12==12.6.77 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (12.6.77)\n",
      "Requirement already satisfied: nvidia-cuda-runtime-cu12==12.6.77 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (12.6.77)\n",
      "Requirement already satisfied: nvidia-cuda-cupti-cu12==12.6.80 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (12.6.80)\n",
      "Requirement already satisfied: nvidia-cudnn-cu12==9.5.1.17 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (9.5.1.17)\n",
      "Requirement already satisfied: nvidia-cublas-cu12==12.6.4.1 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (12.6.4.1)\n",
      "Requirement already satisfied: nvidia-cufft-cu12==11.3.0.4 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (11.3.0.4)\n",
      "Requirement already satisfied: nvidia-curand-cu12==10.3.7.77 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (10.3.7.77)\n",
      "Requirement already satisfied: nvidia-cusolver-cu12==11.7.1.2 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (11.7.1.2)\n",
      "Requirement already satisfied: nvidia-cusparse-cu12==12.5.4.2 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (12.5.4.2)\n",
      "Requirement already satisfied: nvidia-cusparselt-cu12==0.6.3 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (0.6.3)\n",
      "Requirement already satisfied: nvidia-nccl-cu12==2.26.2 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (2.26.2)\n",
      "Requirement already satisfied: nvidia-nvtx-cu12==12.6.77 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (12.6.77)\n",
      "Requirement already satisfied: nvidia-nvjitlink-cu12==12.6.85 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (12.6.85)\n",
      "Requirement already satisfied: nvidia-cufile-cu12==1.11.1.6 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (1.11.1.6)\n",
      "Requirement already satisfied: triton==3.3.1 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from torch<3,>=2.3->bitsandbytes) (3.3.1)\n",
      "Requirement already satisfied: mpmath<1.4,>=1.1.0 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from sympy>=1.13.3->torch<3,>=2.3->bitsandbytes) (1.3.0)\n",
      "Requirement already satisfied: MarkupSafe>=2.0 in /root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages (from jinja2->torch<3,>=2.3->bitsandbytes) (3.0.3)\n",
      "Note: you may need to restart the kernel to use updated packages.\n"
     ]
    }
   ],
   "source": [
    "%pip install bitsandbytes"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d5e0797e-dd3e-4df3-82df-5eeaeeb084f9",
   "metadata": {},
   "source": [
    "## 应用 INT8 量化并打印模型参数\n",
    "\n",
    "观察 `transformer.h.0` 的模块。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "e8f450da-f688-423c-abaa-b76b1b2188b9",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "GPT2LMHeadModel(\n",
      "  (transformer): GPT2Model(\n",
      "    (wte): Embedding(50257, 1280)\n",
      "    (wpe): Embedding(1024, 1280)\n",
      "    (drop): Dropout(p=0.1, inplace=False)\n",
      "    (h): ModuleList(\n",
      "      (0-35): 36 x GPT2Block(\n",
      "        (ln_1): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)\n",
      "        (attn): GPT2Attention(\n",
      "          (c_attn): Linear8bitLt(in_features=1280, out_features=3840, bias=True)\n",
      "          (c_proj): Linear8bitLt(in_features=1280, out_features=1280, bias=True)\n",
      "          (attn_dropout): Dropout(p=0.1, inplace=False)\n",
      "          (resid_dropout): Dropout(p=0.1, inplace=False)\n",
      "        )\n",
      "        (ln_2): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)\n",
      "        (mlp): GPT2MLP(\n",
      "          (c_fc): Linear8bitLt(in_features=1280, out_features=5120, bias=True)\n",
      "          (c_proj): Linear8bitLt(in_features=5120, out_features=1280, bias=True)\n",
      "          (act): NewGELUActivation()\n",
      "          (dropout): Dropout(p=0.1, inplace=False)\n",
      "        )\n",
      "      )\n",
      "    )\n",
      "    (ln_f): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)\n",
      "  )\n",
      "  (lm_head): Linear(in_features=1280, out_features=50257, bias=False)\n",
      ")\n",
      "\n",
      "打印模型 gpt2-large 的参数信息:\n",
      "\n",
      "参数名称: transformer.wte.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.wpe.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.0.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.0.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.0.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.0.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.0.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.0.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.0.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.0.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.0.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.0.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.0.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.0.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.1.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.1.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.1.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.1.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.1.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.1.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.1.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.1.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.1.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.1.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.1.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.1.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.2.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.2.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.2.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.2.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.2.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.2.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.2.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.2.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.2.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.2.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.2.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.2.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.3.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.3.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.3.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.3.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.3.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.3.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.3.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.3.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.3.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.3.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.3.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.3.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.4.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.4.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.4.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.4.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.4.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.4.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.4.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.4.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.4.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.4.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.4.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.4.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.5.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.5.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.5.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.5.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.5.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.5.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.5.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.5.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.5.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.5.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.5.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.5.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.6.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.6.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.6.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.6.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.6.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.6.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.6.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.6.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.6.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.6.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.6.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.6.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.7.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.7.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.7.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.7.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.7.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.7.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.7.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.7.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.7.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.7.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.7.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.7.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.8.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.8.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.8.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.8.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.8.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.8.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.8.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.8.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.8.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.8.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.8.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.8.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.9.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.9.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.9.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.9.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.9.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.9.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.9.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.9.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.9.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.9.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.9.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.9.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.10.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.10.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.10.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.10.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.10.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.10.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.10.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.10.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.10.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.10.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.10.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.10.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.11.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.11.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.11.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.11.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.11.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.11.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.11.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.11.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.11.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.11.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.11.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.11.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.12.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.12.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.12.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.12.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.12.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.12.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.12.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.12.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.12.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.12.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.12.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.12.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.13.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.13.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.13.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.13.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.13.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.13.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.13.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.13.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.13.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.13.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.13.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.13.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.14.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.14.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.14.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.14.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.14.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.14.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.14.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.14.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.14.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.14.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.14.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.14.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.15.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.15.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.15.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.15.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.15.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.15.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.15.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.15.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.15.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.15.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.15.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.15.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.16.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.16.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.16.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.16.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.16.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.16.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.16.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.16.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.16.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.16.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.16.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.16.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.17.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.17.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.17.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.17.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.17.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.17.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.17.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.17.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.17.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.17.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.17.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.17.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.18.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.18.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.18.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.18.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.18.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.18.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.18.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.18.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.18.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.18.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.18.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.18.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.19.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.19.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.19.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.19.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.19.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.19.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.19.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.19.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.19.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.19.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.19.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.19.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.20.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.20.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.20.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.20.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.20.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.20.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.20.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.20.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.20.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.20.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.20.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.20.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.21.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.21.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.21.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.21.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.21.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.21.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.21.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.21.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.21.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.21.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.21.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.21.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.22.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.22.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.22.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.22.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.22.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.22.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.22.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.22.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.22.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.22.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.22.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.22.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.23.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.23.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.23.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.23.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.23.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.23.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.23.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.23.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.23.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.23.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.23.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.23.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.24.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.24.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.24.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.24.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.24.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.24.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.24.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.24.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.24.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.24.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.24.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.24.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.25.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.25.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.25.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.25.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.25.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.25.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.25.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.25.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.25.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.25.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.25.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.25.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.26.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.26.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.26.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.26.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.26.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.26.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.26.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.26.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.26.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.26.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.26.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.26.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.27.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.27.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.27.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.27.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.27.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.27.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.27.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.27.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.27.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.27.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.27.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.27.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.28.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.28.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.28.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.28.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.28.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.28.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.28.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.28.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.28.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.28.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.28.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.28.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.29.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.29.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.29.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.29.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.29.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.29.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.29.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.29.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.29.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.29.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.29.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.29.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.30.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.30.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.30.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.30.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.30.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.30.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.30.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.30.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.30.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.30.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.30.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.30.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.31.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.31.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.31.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.31.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.31.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.31.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.31.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.31.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.31.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.31.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.31.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.31.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.32.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.32.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.32.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.32.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.32.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.32.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.32.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.32.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.32.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.32.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.32.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.32.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.33.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.33.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.33.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.33.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.33.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.33.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.33.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.33.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.33.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.33.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.33.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.33.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.34.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.34.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.34.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.34.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.34.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.34.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.34.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.34.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.34.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.34.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.34.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.34.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.35.ln_1.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.35.ln_1.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.35.attn.c_attn.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.35.attn.c_attn.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.35.attn.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.35.attn.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.35.ln_2.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.h.35.ln_2.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.35.mlp.c_fc.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.35.mlp.c_fc.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.h.35.mlp.c_proj.weight, 数据类型: torch.int8\n",
      "参数名称: transformer.h.35.mlp.c_proj.bias, 数据类型: torch.float16\n",
      "参数名称: transformer.ln_f.weight, 数据类型: torch.float16\n",
      "参数名称: transformer.ln_f.bias, 数据类型: torch.float16\n"
     ]
    }
   ],
   "source": [
    "import os\n",
    "# 设置模型下载镜像（注意，需要在导入 transformers 等模块前进行设置才能起效）\n",
    "os.environ['HF_ENDPOINT'] = 'https://hf-mirror.com'\n",
    "\n",
    "\n",
    "from transformers import AutoModelForCausalLM, BitsAndBytesConfig\n",
    "import torch.nn as nn\n",
    "\n",
    "model_name = 'gpt2-large'\n",
    "\n",
    "# 将模型配置为以 8-bit 量化的方式加载\n",
    "bnb_config = BitsAndBytesConfig(load_in_8bit=True)\n",
    "\n",
    "# 加载模型，并根据设备自动分配\n",
    "model = AutoModelForCausalLM.from_pretrained(\n",
    "    model_name,\n",
    "    quantization_config=bnb_config,\n",
    "    device_map='auto'\n",
    ")\n",
    "\n",
    "# 打印模型的架构\n",
    "print(model)\n",
    "\n",
    "# 打印模型的参数名称和数据类型\n",
    "print(f\"\\n打印模型 {model_name} 的参数信息:\\n\")\n",
    "for name, param in model.named_parameters():\n",
    "    print(f\"参数名称: {name}, 数据类型: {param.dtype}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "936ada0a-9a06-44a1-a8eb-17c17d053861",
   "metadata": {},
   "source": [
    "## 代码示例\n",
    "\n",
    "尝试运行三种不同的 `fp32_values`。\n",
    "```python\n",
    "# 0\n",
    "fp32_values = torch.tensor([3.0, -5.5, 0.0, 6.0, -6.0, 2.5], dtype=torch.float32)\n",
    "# 1\n",
    "fp32_values = torch.tensor([3.0, -5.5, 0.0, 4.0, -6.0, 2.5], dtype=torch.float32)\n",
    "# 2\n",
    "fp32_values = torch.tensor([3.0, -5.5, 0.0, 8.0, -6.0, 2.5], dtype=torch.float32)\n",
    "```"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "3ffeba75-da11-494d-98f5-268dde4df385",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "FP32 的示例张量: tensor([ 3.0000, -5.5000,  0.0000,  6.0000, -6.0000,  2.5000])\n",
      "\n",
      "PyTorch 量化后的 INT8 数值： tensor([  64, -117,    0,  127, -128,   53], dtype=torch.int8)\n",
      "PyTorch 使用的 scale: 0.0470588244497776 zero_point: tensor([0], dtype=torch.int32)\n",
      "PyTorch 反量化恢复后的 FP32 数值： tensor([ 3.0118, -5.5059,  0.0000,  5.9765, -6.0235,  2.4941])\n",
      "\n",
      "=====================\n",
      "\n",
      "自定义计算的 scale: 0.047058823529411764 zero_point: 0\n",
      "自定义量化后的 INT8 数值： tensor([  64, -117,    0,  127, -128,   53], dtype=torch.int8)\n",
      "自定义反量化恢复后的 FP32 数值： tensor([ 3.0118, -5.5059,  0.0000,  5.9765, -6.0235,  2.4941])\n",
      "\n",
      "=====================\n",
      "\n",
      "FP32 模型的权重:\n",
      " Parameter containing:\n",
      "tensor([[ 3.0000, -5.5000,  0.0000,  6.0000, -6.0000,  2.5000]],\n",
      "       requires_grad=True)\n",
      "量化后的 INT8 模型的权重:\n",
      " tensor([[ 3.0118, -5.5059,  0.0000,  5.9765, -6.0235,  2.4941]], size=(1, 6),\n",
      "       dtype=torch.qint8, quantization_scheme=torch.per_channel_affine,\n",
      "       scale=tensor([0.0471], dtype=torch.float64), zero_point=tensor([0]),\n",
      "       axis=0)\n",
      "量化权重的 scale: tensor([0.0471], dtype=torch.float64)\n",
      "量化权重的 zero_point: tensor([0])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/root/autodl-tmp/AI-Guide-and-Demos-zh_CN/.venv/lib/python3.12/site-packages/torch/ao/quantization/observer.py:244: UserWarning: Please use quant_min and quant_max to specify the range for observers.                     reduce_range will be deprecated in a future release of PyTorch.\n",
      "  warnings.warn(\n"
     ]
    }
   ],
   "source": [
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.quantization\n",
    "\n",
    "# 假设 FP32 的张量\n",
    "fp32_values = torch.tensor([3.0, -5.5, 0.0, 6.0, -6.0, 2.5], dtype=torch.float32)\n",
    "print(f\"FP32 的示例张量: {fp32_values}\\n\")\n",
    "\n",
    "# 定义 PyTorch 的量化和反量化函数\n",
    "def pytorch_quantize(fp32_tensor):\n",
    "    # 使用 min-max 范围计算缩放因子，指定 dtype 为 torch.qint8\n",
    "    q_params = torch.quantization.MinMaxObserver(dtype=torch.qint8)\n",
    "    q_params(fp32_tensor)\n",
    "    scale, zero_point = q_params.calculate_qparams()\n",
    "\n",
    "    # 量化\n",
    "    int8_tensor = torch.quantize_per_tensor(fp32_tensor, scale.item(), zero_point.item(), dtype=torch.qint8)\n",
    "    return int8_tensor, scale.item(), zero_point\n",
    "\n",
    "def pytorch_dequantize(int8_tensor):\n",
    "    # 反量化\n",
    "    fp32_tensor = int8_tensor.dequantize()\n",
    "    return fp32_tensor\n",
    "\n",
    "# 量化并获取 PyTorch 结果\n",
    "int8_tensor, scale, zero_point = pytorch_quantize(fp32_values)\n",
    "print(\"PyTorch 量化后的 INT8 数值：\", int8_tensor.int_repr())\n",
    "print(\"PyTorch 使用的 scale:\", scale, \"zero_point:\", zero_point)\n",
    "\n",
    "# 反量化\n",
    "recovered_fp32_pytorch = pytorch_dequantize(int8_tensor)\n",
    "print(\"PyTorch 反量化恢复后的 FP32 数值：\", recovered_fp32_pytorch)\n",
    "\n",
    "print(\"\\n=====================\\n\")\n",
    "\n",
    "# 对比与自定义的量化方式\n",
    "def custom_quantize_compare(fp32_values):\n",
    "    # 获取张量数值的最小值和最大值\n",
    "    x_min, x_max = fp32_values.min().item(), fp32_values.max().item()\n",
    "    \n",
    "    # 定义量化后整数数值的范围\n",
    "    qmin, qmax = -128, 127  # 对应 torch.qint8\n",
    "    \n",
    "    # 计算 scale\n",
    "    scale_custom = (x_max - x_min) / (qmax - qmin)\n",
    "    \n",
    "    # 非对称量化\n",
    "    initial_zero_point = qmin - x_min / scale_custom\n",
    "    zero_point_custom = int(round(initial_zero_point))\n",
    "    \n",
    "    # 将 zero_point 限制在 [qmin, qmax] 范围内\n",
    "    zero_point_custom = max(qmin, min(qmax, zero_point_custom))\n",
    "    \n",
    "    print(\"自定义计算的 scale:\", scale_custom, \"zero_point:\", zero_point_custom)\n",
    "    \n",
    "    def quantize(fp32_tensor, scale, zero_point):\n",
    "        # 计算量化值\n",
    "        int8_tensor = torch.round(fp32_tensor / scale) + zero_point\n",
    "        # 限制在 [qmin, qmax] 范围内\n",
    "        int8_tensor = torch.clamp(int8_tensor, qmin, qmax)\n",
    "        return int8_tensor.to(torch.int8)\n",
    "    \n",
    "    def dequantize(int8_tensor, scale, zero_point):\n",
    "        # 反量化\n",
    "        fp32_tensor = (int8_tensor.float() - zero_point) * scale\n",
    "        return fp32_tensor\n",
    "    \n",
    "    # 量化\n",
    "    int8_values_custom = quantize(fp32_values, scale_custom, zero_point_custom)\n",
    "    print(\"自定义量化后的 INT8 数值：\", int8_values_custom)\n",
    "    \n",
    "    # 反量化\n",
    "    recovered_fp32_custom = dequantize(int8_values_custom, scale_custom, zero_point_custom)\n",
    "    print(\"自定义反量化恢复后的 FP32 数值：\", recovered_fp32_custom)\n",
    "\n",
    "# 运行自定义量化并比较\n",
    "custom_quantize_compare(fp32_values)\n",
    "\n",
    "print(\"\\n=====================\\n\")\n",
    "\n",
    "# 使用 fp32_values 作为线性层参数\n",
    "\n",
    "# 定义一个简单的线性模型\n",
    "class SimpleLinearModel(nn.Module):\n",
    "    def __init__(self, weights, bias=None):\n",
    "        super(SimpleLinearModel, self).__init__()\n",
    "        # 假设输入特征数为6，输出特征数为1，用于匹配之前定义的张量，你也可以试试 (1, 6)，记得对应修改 weights.view(6, 1)\n",
    "        self.linear = nn.Linear(6, 1, bias=False)\n",
    "        # 初始化权重\n",
    "        self.linear.weight = nn.Parameter(weights.view(1, 6))  # 权重形状为 [out_features, in_features]\n",
    "    \n",
    "    def forward(self, x):\n",
    "        return self.linear(x)\n",
    "\n",
    "# 创建 FP32 模型\n",
    "fp32_weights = fp32_values  # [6]\n",
    "model_fp32 = SimpleLinearModel(fp32_weights)\n",
    "\n",
    "# 打印 FP32 模型的权重\n",
    "print(\"FP32 模型的权重:\\n\", model_fp32.linear.weight)\n",
    "\n",
    "# 使用默认量化配置\n",
    "model_fp32.qconfig = torch.quantization.get_default_qconfig('fbgemm')\n",
    "\n",
    "# 准备量化\n",
    "torch.quantization.prepare(model_fp32, inplace=True)\n",
    "\n",
    "# 校准：用示例数据跑一次前向，让 activation observer 统计到激活范围\n",
    "# 不做会触发 UserWarning: must run observer before calling calculate_qparams\n",
    "model_fp32(fp32_values.view(1, 6))\n",
    "\n",
    "# 量化权重\n",
    "torch.quantization.convert(model_fp32, inplace=True)\n",
    "\n",
    "# 打印量化后的权重\n",
    "print(\"量化后的 INT8 模型的权重:\\n\", model_fp32.linear.weight())\n",
    "\n",
    "# 获取量化参数\n",
    "weight_observer = model_fp32.linear.weight().q_per_channel_scales()\n",
    "weight_zero_points = model_fp32.linear.weight().q_per_channel_zero_points()\n",
    "print(\"量化权重的 scale:\", weight_observer)\n",
    "print(\"量化权重的 zero_point:\", weight_zero_points)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "84d27d79-424a-404c-b076-58545107fe00",
   "metadata": {},
   "source": [
    "### 非对称量化\n",
    "\n",
    "模型量化方式修改为非对称量化。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "64256bcf-3e75-4b46-957f-ffdf8a5fdac4",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "FP32 模型的权重:\n",
      " Parameter containing:\n",
      "tensor([[ 3.0000, -5.5000,  0.0000,  6.0000, -6.0000,  2.5000]],\n",
      "       requires_grad=True)\n",
      "量化后的 INT8 模型的权重（int_repr）：\n",
      " tensor([[  64, -117,    0,  127, -128,   53]], dtype=torch.int8)\n",
      "量化权重的 scale: 0.0470588244497776\n",
      "量化权重的 zero_point: 0\n"
     ]
    }
   ],
   "source": [
    "model_fp32 = SimpleLinearModel(fp32_weights)\n",
    "\n",
    "# 打印 FP32 模型的权重\n",
    "print(\"FP32 模型的权重:\\n\", model_fp32.linear.weight)\n",
    "\n",
    "# ------------------------ 修改部分开始 ------------------------\n",
    "\n",
    "import torch.quantization as quant\n",
    "\n",
    "# 自定义的 qconfig，使用非对称量化\n",
    "custom_qconfig = quant.QConfig(\n",
    "    activation=quant.MinMaxObserver.with_args(dtype=torch.quint8, qscheme=torch.per_tensor_affine),\n",
    "    weight=quant.MinMaxObserver.with_args(dtype=torch.qint8, qscheme=torch.per_tensor_affine)\n",
    ")\n",
    "\n",
    "# 应用自定义的 qconfig 到模型\n",
    "model_fp32.qconfig = custom_qconfig\n",
    "\n",
    "# 插入量化准备步骤\n",
    "quant.prepare(model_fp32, inplace=True)\n",
    "\n",
    "# 校准：用示例数据跑一次前向，让 activation observer 统计到激活范围\n",
    "# 不做会触发 UserWarning: must run observer before calling calculate_qparams\n",
    "model_fp32(fp32_values.view(1, 6))\n",
    "\n",
    "# 量化权重\n",
    "quant.convert(model_fp32, inplace=True)\n",
    "\n",
    "# 打印量化后的权重\n",
    "quantized_weight = model_fp32.linear.weight()\n",
    "print(\"量化后的 INT8 模型的权重（int_repr）：\\n\", quantized_weight.int_repr())\n",
    "print(\"量化权重的 scale:\", quantized_weight.q_scale())\n",
    "print(\"量化权重的 zero_point:\", quantized_weight.q_zero_point())"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f266a2fb-a305-468e-9d27-795e898ab0a8",
   "metadata": {},
   "source": [
    "### 对称量化\n",
    "\n",
    "所有方法修改为对称量化。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "26be6efc-3c53-4c0c-89ad-f9452cc28882",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "FP32 的示例张量: tensor([ 3.0000, -5.5000,  0.0000,  6.0000, -6.0000,  2.5000])\n",
      "\n",
      "PyTorch 对称量化后的 INT8 数值： tensor([  64, -117,    0,  127, -128,   53], dtype=torch.int8)\n",
      "PyTorch 使用的 scale: 0.0470588244497776 zero_point: tensor([0])\n",
      "PyTorch 反量化恢复后的 FP32 数值： tensor([ 3.0118, -5.5059,  0.0000,  5.9765, -6.0235,  2.4941])\n",
      "\n",
      "=====================\n",
      "\n",
      "运行自定义对称量化并比较：\n",
      "自定义计算的 scale: 0.047058823529411764 zero_point: 0\n",
      "自定义对称量化后的 INT8 数值： tensor([  64, -117,    0,  127, -128,   53], dtype=torch.int8)\n",
      "自定义对称反量化恢复后的 FP32 数值： tensor([ 3.0118, -5.5059,  0.0000,  5.9765, -6.0235,  2.4941])\n",
      "\n",
      "=====================\n",
      "\n",
      "FP32 模型的权重:\n",
      " Parameter containing:\n",
      "tensor([[ 3.0000, -5.5000,  0.0000,  6.0000, -6.0000,  2.5000]],\n",
      "       requires_grad=True)\n",
      "量化后的 INT8 模型的权重（int_repr）：\n",
      " tensor([[  64, -117,    0,  127, -128,   53]], dtype=torch.int8)\n",
      "量化权重的 scale: 0.0470588244497776\n",
      "量化权重的 zero_point: 0\n"
     ]
    }
   ],
   "source": [
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.quantization\n",
    "\n",
    "# 假设 FP32 的张量\n",
    "fp32_values = torch.tensor([3.0, -5.5, 0.0, 6.0, -6.0, 2.5], dtype=torch.float32)\n",
    "print(f\"FP32 的示例张量: {fp32_values}\\n\")\n",
    "\n",
    "# 定义 PyTorch 的量化和反量化函数（对称量化）\n",
    "def pytorch_quantize_symmetric(fp32_tensor):\n",
    "    # 使用 MinMaxObserver 计算缩放因子，指定 dtype 为 torch.qint8，使用对称量化\n",
    "    q_params = torch.quantization.MinMaxObserver(\n",
    "        dtype=torch.qint8, qscheme=torch.per_tensor_symmetric\n",
    "    )\n",
    "    q_params(fp32_tensor)\n",
    "    scale, zero_point = q_params.calculate_qparams()\n",
    "\n",
    "    # 量化\n",
    "    int8_tensor = torch.quantize_per_tensor(fp32_tensor, scale.item(), zero_point.item(), dtype=torch.qint8)\n",
    "    return int8_tensor, scale.item(), zero_point\n",
    "\n",
    "def pytorch_dequantize(int8_tensor):\n",
    "    # 反量化\n",
    "    fp32_tensor = int8_tensor.dequantize()\n",
    "    return fp32_tensor\n",
    "\n",
    "# 量化并获取 PyTorch 结果（对称量化）\n",
    "int8_tensor, scale, zero_point = pytorch_quantize_symmetric(fp32_values)\n",
    "print(\"PyTorch 对称量化后的 INT8 数值：\", int8_tensor.int_repr())\n",
    "print(\"PyTorch 使用的 scale:\", scale, \"zero_point:\", zero_point)\n",
    "\n",
    "# 反量化\n",
    "recovered_fp32_pytorch = pytorch_dequantize(int8_tensor)\n",
    "print(\"PyTorch 反量化恢复后的 FP32 数值：\", recovered_fp32_pytorch)\n",
    "\n",
    "print(\"\\n=====================\\n\")\n",
    "\n",
    "# 对比与自定义的量化方式（对称量化）\n",
    "def custom_quantize_compare_symmetric(fp32_values):\n",
    "    # 获取张量的绝对最大值\n",
    "    x_max = fp32_values.abs().max().item()\n",
    "    \n",
    "    # 定义量化后整数数值的范围\n",
    "    qmin, qmax = -128, 127  # 对应 torch.qint8\n",
    "    \n",
    "    # 使用 PyTorch 的方式计算 scale\n",
    "    scale_custom = x_max / 127.5  # PyTorch源码：scale = max_val_pos / (float(quant_max - quant_min) / 2)\n",
    "    \n",
    "    # 对称量化时，zero_point 固定为 0\n",
    "    zero_point_custom = 0\n",
    "    \n",
    "    print(\"自定义计算的 scale:\", scale_custom, \"zero_point:\", zero_point_custom)\n",
    "    \n",
    "    # 使用 torch 的 round 函数以匹配 PyTorch 的量化行为\n",
    "    def quantize(fp32_tensor, scale, zero_point):\n",
    "        # 计算量化值\n",
    "        int8_tensor = torch.round(fp32_tensor / scale) + zero_point\n",
    "        # 限制在 [qmin, qmax] 范围内\n",
    "        int8_tensor = torch.clamp(int8_tensor, qmin, qmax)\n",
    "        return int8_tensor.to(torch.int8)\n",
    "    \n",
    "    def dequantize(int8_tensor, scale, zero_point):\n",
    "        # 反量化\n",
    "        fp32_tensor = (int8_tensor.float() - zero_point) * scale\n",
    "        return fp32_tensor\n",
    "    \n",
    "    # 量化\n",
    "    int8_values_custom = quantize(fp32_values, scale_custom, zero_point_custom)\n",
    "    print(\"自定义对称量化后的 INT8 数值：\", int8_values_custom)\n",
    "    \n",
    "    # 反量化\n",
    "    recovered_fp32_custom = dequantize(int8_values_custom, scale_custom, zero_point_custom)\n",
    "    print(\"自定义对称反量化恢复后的 FP32 数值：\", recovered_fp32_custom)\n",
    "    \n",
    "    return recovered_fp32_custom  # 返回以便后续使用\n",
    "\n",
    "# 运行自定义对称量化并比较\n",
    "print(\"运行自定义对称量化并比较：\")\n",
    "recovered_fp32_custom = custom_quantize_compare_symmetric(fp32_values)\n",
    "\n",
    "print(\"\\n=====================\\n\")\n",
    "\n",
    "# 使用 fp32_values 作为线性层参数\n",
    "\n",
    "# 定义一个简单的线性模型\n",
    "class SimpleLinearModel(nn.Module):\n",
    "    def __init__(self, weights, bias=None):\n",
    "        super(SimpleLinearModel, self).__init__()\n",
    "        # 假设输入特征数为6，输出特征数为1\n",
    "        self.linear = nn.Linear(6, 1, bias=False)\n",
    "        # 初始化权重\n",
    "        with torch.no_grad():\n",
    "            self.linear.weight = nn.Parameter(weights.view(1, 6))  # 权重形状为 [out_features, in_features]\n",
    "    \n",
    "    def forward(self, x):\n",
    "        return self.linear(x)\n",
    "\n",
    "# 创建 FP32 模型\n",
    "fp32_weights = fp32_values  # [6]\n",
    "model_fp32 = SimpleLinearModel(fp32_weights)\n",
    "\n",
    "# 打印 FP32 模型的权重\n",
    "print(\"FP32 模型的权重:\\n\", model_fp32.linear.weight)\n",
    "\n",
    "# 和之前非对称保持一致方便对比\n",
    "\n",
    "import torch.quantization as quant\n",
    "\n",
    "# 自定义的 qconfig，使用对称量化\n",
    "custom_qconfig = quant.QConfig(\n",
    "    activation=quant.MinMaxObserver.with_args(\n",
    "        dtype=torch.quint8, qscheme=torch.per_tensor_symmetric\n",
    "    ),\n",
    "    weight=quant.MinMaxObserver.with_args(\n",
    "        dtype=torch.qint8, qscheme=torch.per_tensor_symmetric\n",
    "    )\n",
    ")\n",
    "\n",
    "# 应用自定义的 qconfig 到模型，结果和model_fp32.qconfig = torch.quantization.get_default_qconfig('fbgemm')一致\n",
    "model_fp32.qconfig = custom_qconfig\n",
    "\n",
    "# 准备量化\n",
    "quant.prepare(model_fp32, inplace=True)\n",
    "\n",
    "# 校准：用示例数据跑一次前向，让 activation observer 统计到激活范围\n",
    "# 不做会触发 UserWarning: must run observer before calling calculate_qparams\n",
    "model_fp32(fp32_values.view(1, 6))\n",
    "\n",
    "# 量化权重\n",
    "quant.convert(model_fp32, inplace=True)\n",
    "\n",
    "# 打印量化后的权重\n",
    "quantized_weight = model_fp32.linear.weight()\n",
    "print(\"量化后的 INT8 模型的权重（int_repr）：\\n\", quantized_weight.int_repr())\n",
    "print(\"量化权重的 scale:\", quantized_weight.q_scale())\n",
    "print(\"量化权重的 zero_point:\", quantized_weight.q_zero_point())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "2e89242a-deda-48eb-819b-ea62b2ab40dd",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "ai",
   "language": "python",
   "name": "ai"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.12.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}