{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7653981e-5b39-45d7-ad23-aa18de28af12",
   "metadata": {},
   "outputs": [],
   "source": [
    "%cd .."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "5e995d5e-f8bd-411f-a78d-ba03edc60968",
   "metadata": {},
   "outputs": [],
   "source": [
    "base_model = \"unsloth/gemma-3-27b-it\"\n",
    "adapter_path = \"gemma-27b-thinker/checkpoint-2400\"\n",
    "output_path = \"output-merged\"\n",
    "\n",
    "def merge_lora(base_model, adapter_path, output_path, max_shard_size, precision):\n",
    "    \"\"\"Merge LoRA adapter into base model and save the merged model.\"\"\"\n",
    "    try:\n",
    "        # Load LoRA adapter\n",
    "        adapter_model = PeftModel.from_pretrained(base_model, adapter_path)\n",
    "        \n",
    "        # Merge LoRA weights and free memory\n",
    "        merged_model = adapter_model.merge_and_unload()\n",
    "        del adapter_model\n",
    "        torch.cuda.empty_cache()\n",
    "\n",
    "        # Convert model precision\n",
    "        dtype = torch.bfloat16 if precision == \"bfloat16\" else torch.float16\n",
    "        for param in merged_model.parameters():\n",
    "            if param.dtype == torch.float32:\n",
    "                param.data = param.data.to(dtype)\n",
    "\n",
    "        # Save merged model\n",
    "        os.makedirs(output_path, exist_ok=True)\n",
    "        merged_model.save_pretrained(output_path, safe_serialization=True, max_shard_size=max_shard_size)\n",
    "\n",
    "        # Save tokenizer\n",
    "        tokenizer = AutoTokenizer.from_pretrained(base_model.name_or_path)\n",
    "        tokenizer.save_pretrained(output_path)\n",
    "\n",
    "        return True\n",
    "\n",
    "    except Exception as e:\n",
    "        print(f\"Merge failed: {e}\")\n",
    "        return False\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "0d59c8fc-9db6-492c-ba98-0e87e728e929",
   "metadata": {},
   "outputs": [],
   "source": [
    "from unsloth.chat_templates import standardize_data_formats\n",
    "formatted_dataset = standardize_data_formats(formatted_dataset)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "749a0361-a7f9-4d1a-967d-6f651f4eb1bb",
   "metadata": {},
   "outputs": [],
   "source": [
    "def apply_chat_template(examples):\n",
    "    texts = tokenizer.apply_chat_template(examples[\"messages\"])\n",
    "    return { \"text\" : texts }\n",
    "pass\n",
    "formatted_dataset = formatted_dataset.map(apply_chat_template, batched = True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7bb95cac-5d93-4e0f-a124-c0ff48096411",
   "metadata": {},
   "outputs": [],
   "source": [
    "print(formatted_dataset[1])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "ed846b5d-b1f6-4197-8401-8797b3ac3508",
   "metadata": {},
   "outputs": [
    {
     "ename": "NameError",
     "evalue": "name 'os' is not defined",
     "output_type": "error",
     "traceback": [
      "\u001b[31m---------------------------------------------------------------------------\u001b[39m",
      "\u001b[31mNameError\u001b[39m                                 Traceback (most recent call last)",
      "\u001b[36mCell\u001b[39m\u001b[36m \u001b[39m\u001b[32mIn[10]\u001b[39m\u001b[32m, line 1\u001b[39m\n\u001b[32m----> \u001b[39m\u001b[32m1\u001b[39m \u001b[43mos\u001b[49m\n",
      "\u001b[31mNameError\u001b[39m: name 'os' is not defined"
     ]
    }
   ],
   "source": [
    "!git clone --depth=1 https://github.com/ggerganov/llama.cpp\n",
    "%cd llama.cpp\n",
    "!cmake -Bbuild\n",
    "!cmake --build build -D...\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f1fa7b18-c2b0-47a4-a46a-38bdbf5e4582",
   "metadata": {},
   "outputs": [],
   "source": [
    "%cd build\n",
    "!cpack -G DEB\n",
    "!dpkg -i *.deb"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.11.5"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}