Upload setup_fine_tune_mps.ipynb

setup_fine_tune_mps.ipynb

@@ -0,0 +1,107 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "5aa6bfeb",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# === 1. 在所有导入之前设置环境变量 ===\n",
+    "import os\n",
+    "from pathlib import Path\n",
+    "\n",
+    "# 强制指定模板路径\n",
+    "custom_template_dir = Path(\"./custom_templates\").resolve()\n",
+    "os.environ[\"HUGGINGFACE_HUB_TEMPLATES_PATH\"] = str(custom_template_dir)\n",
+    "os.environ[\"PYTORCH_MPS_HIGH_WATERMARK_RATIO\"] = \"0.0\"\n",
+    "\n",
+    "# === 2. 初始化模板 ===\n",
+    "custom_template_dir.mkdir(parents=True, exist_ok=True)\n",
+    "template_file = custom_template_dir / \"modelcard_template.md\"\n",
+    "template_content = \"\"\"---\n",
+    "language: en\n",
+    "tags:\n",
+    "- generated_from_trainer\n",
+    "model-index:\n",
+    "- name: fine-tuned\n",
+    "  results: []\n",
+    "---\n",
+    "\n",
+    "# Model Card\n",
+    "\n",
+    "This model was fine-tuned using PEFT/LoRA.\"\"\"\n",
+    "\n",
+    "if not template_file.exists():\n",
+    "    with open(template_file, \"w\") as f:\n",
+    "        f.write(template_content)\n",
+    "    print(\"✅ 自定义模板已创建\")\n",
+    "\n",
+    "# === 3. 强制禁用默认模板 ===\n",
+    "import huggingface_hub.repocard as hf_card\n",
+    "def _disable_default_template():\n",
+    "    # 禁用默认模板加载\n",
+    "    hf_card.ModelCard._default_template_path = None\n",
+    "    # 重定向模板加载方法\n",
+    "    hf_card.ModelCard.from_template = classmethod(\n",
+    "        lambda cls, template_path=None, **kwargs: cls(template_content))\n",
+    "    \n",
+    "_disable_default_template()\n",
+    "\n",
+    "# === 4. 强化猴子补丁 ===\n",
+    "# === 正确应用猴子补丁的方法 ===\n",
+    "from functools import wraps\n",
+    "from transformers import Trainer\n",
+    "import shutil\n",
+    "\n",
+    "# 1. 保存原始方法到不同变量名\n",
+    "_original_trainer_save = Trainer.save_model  # 使用唯一变量名\n",
+    "\n",
+    "@wraps(_original_trainer_save)\n",
+    "def _patched_save_model(self, output_dir, **kwargs):\n",
+    "    # 调用原始保存方法（使用保存的引用）\n",
+    "    _original_trainer_save(self, output_dir, **kwargs)\n",
+    "    \n",
+    "    # 原子化写入模板\n",
+    "    output_path = Path(output_dir)\n",
+    "    temp_file = output_path / \"README.tmp\"\n",
+    "    target_file = output_path / \"README.md\"\n",
+    "    \n",
+    "    try:\n",
+    "        # 写入临时文件\n",
+    "        with open(temp_file, \"w\") as f:\n",
+    "            f.write(template_content)\n",
+    "        # 原子替换\n",
+    "        temp_file.rename(target_file)\n",
+    "    except Exception as e:\n",
+    "        print(f\"⚠️ 模板写入失败: {str(e)}\")\n",
+    "    \n",
+    "    print(f\"✅ 检查点 {output_path.name} 保存完成\")\n",
+    "\n",
+    "# 2. 应用补丁前确保变量未被覆盖\n",
+    "if not hasattr(Trainer, '_hf_original_save_model'):\n",
+    "    Trainer._hf_original_save_model = Trainer.save_model\n",
+    "    Trainer.save_model = _patched_save_model\n",
+    "\n",
+    "# === 5. 验证 Hugging Face 配置 ===\n",
+    "import huggingface_hub\n",
+    "print(f\"当前模板路径: {os.environ.get('HUGGINGFACE_HUB_TEMPLATES_PATH', '未设置')}\")\n",
+    "\n",
+    "\n",
+    "# === 剩余代码保持不变，但需确保以下顺序 ===\n",
+    "import torch\n",
+    "from huggingface_hub import login\n",
+    "from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments\n",
+    "from peft import LoraConfig\n",
+    "from trl import SFTTrainer"
+   ]
+  }
+ ],
+ "metadata": {
+  "language_info": {
+   "name": "python"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}