add finetune nb example

examples/ruGPT3XL_finetune_example.ipynb

@@ -0,0 +1,250 @@
+{
+ "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
+}