Upload merge.ipynb

merge.ipynb

@@ -0,0 +1,504 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "from transformers import AutoModelForCausalLM, AutoTokenizer"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from safetensors.torch import save_file, load_file"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "CACHE_DIR = \"/huggingface/cache\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "NEKOMATA_MODEL = \"rinna/nekomata-14b\"\n",
+    "QARASU_MODEL = \"lightblue/qarasu-14B-chat-plus-unleashed\"\n",
+    "QWEN_14B_MODEL = \"Qwen/Qwen-14B\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "nekomata = AutoModelForCausalLM.from_pretrained(\n",
+    "    NEKOMATA_MODEL,\n",
+    "    cache_dir=CACHE_DIR,\n",
+    "    torch_dtype=torch.bfloat16,\n",
+    "    device_map=\"cpu\",\n",
+    "    offload_folder=\"nekomata\",\n",
+    "    offload_state_dict=True,\n",
+    "    trust_remote_code=True,\n",
+    ")\n",
+    "nekomata.eval()\n",
+    "nekomata.hf_device_map"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "nekomata_state_dict = nekomata.state_dict().copy()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "for key in nekomata_state_dict.keys():\n",
+    "    nekomata_value = nekomata_state_dict[key].clone().to(\"cpu\")\n",
+    "    print(key, nekomata_value.dtype, nekomata_value.shape, nekomata_value)\n",
+    "    break"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "save_file(nekomata_state_dict, \"./nekomata_state.safetensors\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "**Restart Runtime**\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "qarasu = AutoModelForCausalLM.from_pretrained(\n",
+    "    QARASU_MODEL,\n",
+    "    cache_dir=CACHE_DIR,\n",
+    "    torch_dtype=torch.bfloat16,\n",
+    "    device_map=\"cpu\",\n",
+    "    offload_folder=\"qarasu\",\n",
+    "    offload_state_dict=True,\n",
+    "    trust_remote_code=True,\n",
+    ")\n",
+    "qarasu.eval()\n",
+    "qarasu.hf_device_map"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "**Restart Runtime**\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "qarasu_state_dict = qarasu.state_dict().copy()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "for key in qarasu_state_dict.keys():\n",
+    "    qarasu_value = qarasu_state_dict[key].clone().to(\"cpu\")\n",
+    "    print(key, qarasu_value.dtype, qarasu_value.shape, qarasu_value)\n",
+    "    break"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "save_file(qarasu_state_dict, \"./qarasu_state.safetensors\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "**Restart Runtime**\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "qwen14b = AutoModelForCausalLM.from_pretrained(\n",
+    "    QWEN_14B_MODEL,\n",
+    "    cache_dir=CACHE_DIR,\n",
+    "    torch_dtype=torch.bfloat16,\n",
+    "    device_map=\"cpu\",\n",
+    "    offload_folder=\"qwen\",\n",
+    "    offload_state_dict=True,\n",
+    "    trust_remote_code=True,\n",
+    ")\n",
+    "qwen14b.eval()\n",
+    "qwen14b.hf_device_map"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "qwen14b_state_dict = qwen14b.state_dict().copy()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "save_file(qwen14b_state_dict, \"./qwen14b_state.safetensors\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "**Restart Runtime**\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from safetensors.torch import save_file, load_file"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "nekomata_state_dict = load_file(\"./nekomata_state.safetensors\", device=\"cpu\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "qarasu_state_dict = load_file(\"./qarasu_state.safetensors\", device=\"cpu\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "new_state_dict = nekomata_state_dict\n",
+    "\n",
+    "with torch.no_grad():\n",
+    "    for key in nekomata_state_dict.keys():\n",
+    "        print(key)\n",
+    "\n",
+    "        new_state_dict[key] = (\n",
+    "            new_state_dict[key].to(\"cuda\") + qarasu_state_dict[key].to(\"cuda\")\n",
+    "        ).to(\"cpu\")\n",
+    "\n",
+    "new_state_dict"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "del nekomata_state_dict, qarasu_state_dict"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "torch.cuda.empty_cache()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "save_file(new_state_dict, \"./nekomata+qarasu_state.safetensors\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "**Restart Runtime**\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "from safetensors.torch import load_file, save_file"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "nekomata_qarasu_state_dict = load_file(\n",
+    "    \"./nekomata+qarasu_state.safetensors\", device=\"cpu\"\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "qwen14b_state_dict = load_file(\"./qwen14b_state.safetensors\", device=\"cpu\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# キー名が同じことを確認\n",
+    "for neko_key, qwen14b_key in zip(\n",
+    "    nekomata_qarasu_state_dict.keys(), qwen14b_state_dict.keys()\n",
+    "):\n",
+    "    assert neko_key == qwen14b_key"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "new_state_dict = nekomata_qarasu_state_dict\n",
+    "\n",
+    "with torch.no_grad():\n",
+    "    for key in new_state_dict.keys():\n",
+    "        print(key)\n",
+    "\n",
+    "        new_state_dict[key] = (\n",
+    "            new_state_dict[key].to(\"cuda\") - qwen14b_state_dict[key].to(\"cuda\")\n",
+    "        ).to(\"cpu\")\n",
+    "\n",
+    "new_state_dict"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "save_file(new_state_dict, \"./nekoqarasu_state.safetensors\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "**Restart Runtime**\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "CACHE_DIR = \"/huggingface/cache\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "QWEN_14B_CHAT_MODEL = \"Qwen/Qwen-14B-Chat\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "qwen_chat_config = AutoConfig.from_pretrained(\n",
+    "    QWEN_14B_CHAT_MODEL, trust_remote_code=True, cache_dir=CACHE_DIR\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "nekoqarasu = AutoModelForCausalLM.from_config(\n",
+    "    qwen_chat_config,\n",
+    "    trust_remote_code=True,\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from safetensors.torch import load_file"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "state_dict = load_file(\"./nekoqarasu_state.safetensors\", device=\"cpu\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "nekoqarasu.load_state_dict(\n",
+    "    state_dict,\n",
+    "    strict=False,\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "nekoqarasu.push_to_hub(\"nekoqarasu-14b-chat\", private=True)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "tokenizer = AutoTokenizer.from_pretrained(\n",
+    "    QWEN_14B_CHAT_MODEL, cache_dir=CACHE_DIR, trust_remote_code=True\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "tokenizer.push_to_hub(\"nekoqarasu-14b-chat\", private=True)"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "py310",
+   "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": "3.10.10"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}