examples/ruGPT3XL_finetune_example.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "%load_ext autoreload\n",
    "\n",
    "%autoreload 2"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Run finetuning"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "* Prepare your files and split to shards. All files will be splitted between all gpus. All paths to text files should be stored at file.list\n",
    "\n",
    "Example of `files.list` for single gpu:\n",
    "\n",
    "```text\n",
    "data/train.txt\n",
    "```\n",
    "\n",
    "* Add your path to template script in our repo [scripts/deepspeed_gpt3_xl_finetune.sh](https://github.com/sberbank-ai/ru-gpts/blob/master/scripts/deepspeed_gpt3_xl_finetune.sh)\n",
    "\n",
    "* `cd scripts` for go to scripts dir\n",
    "\n",
    "* run command\n",
    "\n",
    "```bash\n",
    "sh deepspeed_gpt3_xl_finetune.sh\n",
    "\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "7TbybJfIpBVa"
   },
   "source": [
    "### Test model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Load model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "id": "EnTy1SEajpPV"
   },
   "outputs": [],
   "source": [
    "import warnings\n",
    "warnings.filterwarnings(\"ignore\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "id": "Z-DSEz0ljpPV"
   },
   "outputs": [],
   "source": [
    "import sys\n",
    "sys.path.append(\"../../ru-gpts/\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "id": "_9GABoxNVpH4"
   },
   "outputs": [],
   "source": [
    "import os\n",
    "os.environ[\"USE_DEEPSPEED\"] = \"1\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "_HeCFnJEjpPV"
   },
   "outputs": [],
   "source": [
    "from src.xl_wrapper import RuGPT3XL"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "OdyughHDjpPV"
   },
   "source": [
    "Note! seq_len is max sequence length for generation used in generation process. Max avialable seq_len is 2048 (in tokens).\n",
    "Also inference takes around 10 Gb GPU memory."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "56aNJNPYjpPW",
    "outputId": "aa154214-82b6-4604-d0b7-8ad7629aa704",
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "> initializing model parallel with size 1\n",
      "> initializing model parallel cuda seeds on global rank 0, model parallel rank 0, and data parallel rank 0 with model parallel seed: 3952 and data parallel seed: 1234\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/user/conda/lib/python3.7/site-packages/ipykernel/ipkernel.py:283: DeprecationWarning: `should_run_async` will not call `transform_cell` automatically in the future. Please pass the result to `transformed_cell` argument and any exception that happen during thetransform in `preprocessing_exc_tuple` in IPython 7.17 and above.\n",
      "  and should_run_async(code)\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Use alternating sparse & dense attention layers\n"
     ]
    }
   ],
   "source": [
    "gpt = RuGPT3XL.from_pretrained(\n",
    "    \"sberbank-ai/rugpt3xl\",\n",
    "    weights_path=\"/path/2/save/model/train-iters/mp_rank_00_model_states.pt\",\n",
    "    deepspeed_config_path=\"../src/deepspeed_config/gpt3_xl_sparse_2048.json\",\n",
    "    seq_len=512\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Example of generation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "7Qw65CVzjpPZ",
    "outputId": "79a31fb8-656d-4923-f208-1589d675d7c7"
   },
   "outputs": [],
   "source": [
    "def generate(text, additional_len=20):\n",
    "    return filter_resuls(gpt.generate(\n",
    "        text=text,\n",
    "        max_length=len(gpt.tokenizer.encode(text)) + additional_len,\n",
    "        num_beams=10,\n",
    "        eos_token_id=gpt.tokenizer.eos_token_id,\n",
    "        num_return_sequences=1,\n",
    "    ))[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "def filter_resuls(nr):\n",
    "    return [x[:x.find(\"<|endoftext|>\")] for x in nr]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'Фрегат «Адмирал Макаров» вошел в Средиземное море\\u2028Большой противолодочный корабль «Вице-адмирал Кулак'"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "generate(\"Фрегат «Адмирал Макаров»\")"
   ]
  }
 ],
 "metadata": {
  "accelerator": "GPU",
  "colab": {
   "collapsed_sections": [],
   "name": "ruGPT3XL_generation",
   "provenance": []
  },
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.7.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}