Saving weights and logs of step 100

.ipynb_checkpoints/Untitled-checkpoint.ipynb

@@ -0,0 +1,6 @@
+{
+ "cells": [],
+ "metadata": {},
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

Untitled.ipynb

@@ -0,0 +1,436 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 12,
+   "id": "81fd300c",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import pandas as pd\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 13,
+   "id": "1237ddf7",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "3418\n"
+     ]
+    }
+   ],
+   "source": [
+    "with open(\"Lilgpt.txt\",'r') as file:\n",
+    "    data = file.read()\n",
+    "List = data.split(\"<EOS>\")\n",
+    "print(len(List))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 14,
+   "id": "18b93c2e",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "\n",
+      "<BOS>\n",
+      "3 Headed Goat[Intro]\n",
+      "(Aviator)\n",
+      "\n",
+      "[Chorus: Lil Baby]\n",
+      "These ain't no Guess jeans\n",
+      "I dropped out of school, I'm still good at math, but, nigga, don't test me\n",
+      "I played to the left, they went to the right, they tried to finesse me\n",
+      "Still riding 'round with that blicky, I hope they don't catch me\n",
+      "Police had raided our spot, so we went to the next street\n",
+      "Play like I'm dumb, as soon as it pop, I'm goin' retarded\n",
+      "He say I'm hard and he say I'm garbage, I'm rich regardless\n",
+      "We in Miami in the middle of the winter, and we on them jet skis\n",
+      "If we in Atlanta, I'm runnin' the 'Cat and I'm workin' the red key\n",
+      "\n",
+      "[Verse 1: Lil Durk]\n",
+      "I cannot mention my homies inside of my song 'cause I know they be trappin' a lot\n",
+      "I can't keep takin' these pills, when I'm in the trenches, they say I be cappin' a lot\n",
+      "I know a nigga who say he got rich off the dope, but I know he be actin' a lot\n",
+      "I know some niggas who said that they took down the city, but niggas be lackin' a lot\n",
+      "Yeah\n",
+      "That shit was awful, nigga had that dog food\n",
+      "That day they shot you, I slid on a Mongoose\n",
+      "You cannot come back around me, you turned your back on me, I cannot forget\n",
+      "The police was lyin', they say that they caught you, but nigga, they made you admit\n",
+      "Your name was found, you put in that work, they took your stick, you a bitch\n",
+      "Fuck my opps, they be on my dick, they all be mad we rich (Turn up)\n",
+      "[Verse 2: Lil Baby]\n",
+      "Only twenty-five, livin' like a boss, ridin' 'round with a chauffeur\n",
+      "I don't sell drugs, still be paranoid, keep lookin' over my shoulder\n",
+      "Niggas lyin' like I'm stealin' swag, boy, that's my shit like I wrote it\n",
+      "\n",
+      "[Verse 3: Polo G]\n",
+      "Uh\n",
+      "These rappers really nice as hell\n",
+      "I'm a different nigga when I'm pissed off\n",
+      "Man, he say he gon' press up on who?\n",
+      "I'ma get the steel like I'm Chris Paul\n",
+      "Back to back suburbans, I'm a big dawg\n",
+      "I was in the slums servin' Fentanyl\n",
+      "Zombieland, junkies havin' withdrawals\n",
+      "I been gettin' to it, lotta missed calls\n",
+      "Turn it off, what the fuck is he talking 'bout?\n",
+      "I should slap you for sayin' he hot as me\n",
+      "I don't know who could fuck with me honestly\n",
+      "They know I'm the man, so they watchin' me\n",
+      "Different color bands like Monopoly\n",
+      "Man, he must not be usin' his head\n",
+      "If he thinkin' I don't keep a Glock with me\n",
+      "That's like suicide if you play with us\n",
+      "Got a better chance at the lottery\n",
+      "Call an ambulance when that chopper sweep\n",
+      "Make the crowd dance, choreography\n",
+      "Once I got a plan, ain't no stoppin' me\n",
+      "Three-car garage, million-dollar crib\n",
+      "With a foreign bitch ridin' on top of me\n",
+      "Lot of people done said I wouldn't be shit\n",
+      "Well, I guess they owe me an apology\n",
+      "[Chorus: Lil Baby]\n",
+      "These ain't no Guess jeans\n",
+      "I dropped out of school, I'm still good at math, but, nigga, don't test me\n",
+      "I played to the left, they went to the right, they tried to finesse me\n",
+      "Still riding 'round with that blicky, I hope they don't catch me\n",
+      "Police had raided our spot, so we went to the next street\n",
+      "Play like I'm dumb, as soon as it pop, I'm goin' retarded\n",
+      "He say I'm hard and he say I'm garbage, I'm rich regardless\n",
+      "We in Miami in the middle of the winter, and we on them jet skis\n",
+      "If we in Atlanta, I'm runnin' the 'Cat and I'm workin' the red key35EmbedShare URLCopyEmbedCopy\n",
+      "\n"
+     ]
+    }
+   ],
+   "source": [
+    "print(List[1])\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 15,
+   "id": "9c7698db",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "\n",
+      "<BOS>\n",
+      "3 Headed Goat[Intro]\n",
+      "(Aviator)\n",
+      "\n",
+      "[Chorus: Lil Baby]\n",
+      "These ain't no Guess jeans\n",
+      "I dropped out of school, I'm still good at math, but, nigga, don't test me\n",
+      "I played to the left, they went to the right, they tried to finesse me\n",
+      "Still riding 'round with that blicky, I hope they don't catch me\n",
+      "Police had raided our spot, so we went to the next street\n",
+      "Play like I'm dumb, as soon as it pop, I'm goin' retarded\n",
+      "He say I'm hard and he say I'm garbage, I'm rich regardless\n",
+      "We in Miami in the middle of the winter, and we on them jet skis\n",
+      "If we in Atlanta, I'm runnin' the 'Cat and I'm workin' the red key\n",
+      "\n",
+      "[Verse 1: Lil Durk]\n",
+      "I cannot mention my homies inside of my song 'cause I know they be trappin' a lot\n",
+      "I can't keep takin' these pills, when I'm in the trenches, they say I be cappin' a lot\n",
+      "I know a nigga who say he got rich off the dope, but I know he be actin' a lot\n",
+      "I know some niggas who said that they took down the city, but niggas be lackin' a lot\n",
+      "Yeah\n",
+      "That shit was awful, nigga had that dog food\n",
+      "That day they shot you, I slid on a Mongoose\n",
+      "You cannot come back around me, you turned your back on me, I cannot forget\n",
+      "The police was lyin', they say that they caught you, but nigga, they made you admit\n",
+      "Your name was found, you put in that work, they took your stick, you a bitch\n",
+      "Fuck my opps, they be on my dick, they all be mad we rich (Turn up)\n",
+      "[Verse 2: Lil Baby]\n",
+      "Only twenty-five, livin' like a boss, ridin' 'round with a chauffeur\n",
+      "I don't sell drugs, still be paranoid, keep lookin' over my shoulder\n",
+      "Niggas lyin' like I'm stealin' swag, boy, that's my shit like I wrote it\n",
+      "\n",
+      "[Verse 3: Polo G]\n",
+      "Uh\n",
+      "These rappers really nice as hell\n",
+      "I'm a different nigga when I'm pissed off\n",
+      "Man, he say he gon' press up on who?\n",
+      "I'ma get the steel like I'm Chris Paul\n",
+      "Back to back suburbans, I'm a big dawg\n",
+      "I was in the slums servin' Fentanyl\n",
+      "Zombieland, junkies havin' withdrawals\n",
+      "I been gettin' to it, lotta missed calls\n",
+      "Turn it off, what the fuck is he talking 'bout?\n",
+      "I should slap you for sayin' he hot as me\n",
+      "I don't know who could fuck with me honestly\n",
+      "They know I'm the man, so they watchin' me\n",
+      "Different color bands like Monopoly\n",
+      "Man, he must not be usin' his head\n",
+      "If he thinkin' I don't keep a Glock with me\n",
+      "That's like suicide if you play with us\n",
+      "Got a better chance at the lottery\n",
+      "Call an ambulance when that chopper sweep\n",
+      "Make the crowd dance, choreography\n",
+      "Once I got a plan, ain't no stoppin' me\n",
+      "Three-car garage, million-dollar crib\n",
+      "With a foreign bitch ridin' on top of me\n",
+      "Lot of people done said I wouldn't be shit\n",
+      "Well, I guess they owe me an apology\n",
+      "[Chorus: Lil Baby]\n",
+      "These ain't no Guess jeans\n",
+      "I dropped out of school, I'm still good at math, but, nigga, don't test me\n",
+      "I played to the left, they went to the right, they tried to finesse me\n",
+      "Still riding 'round with that blicky, I hope they don't catch me\n",
+      "Police had raided our spot, so we went to the next street\n",
+      "Play like I'm dumb, as soon as it pop, I'm goin' retarded\n",
+      "He say I'm hard and he say I'm garbage, I'm rich regardless\n",
+      "We in Miami in the middle of the winter, and we on them jet skis\n",
+      "If we in Atlanta, I'm runnin' the 'Cat and I'm workin' the red key35EmbedShare URLCopyEmbedCopy\n",
+      "<EOS>\n"
+     ]
+    }
+   ],
+   "source": [
+    "NewList = []\n",
+    "for l in List:\n",
+    "    n = l+\"<EOS>\"\n",
+    "    NewList.append(n)\n",
+    "print(NewList[1])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 16,
+   "id": "f31e28d8",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "330\n",
+      "3088\n"
+     ]
+    }
+   ],
+   "source": [
+    "counter = 0\n",
+    "List_val = []\n",
+    "List_train = []\n",
+    "for l in List:\n",
+    "    n = l+\"<EOS>\"\n",
+    "    if counter<330:\n",
+    "        List_val.append(n)\n",
+    "    else:\n",
+    "        List_train.append(n)\n",
+    "    counter += 1\n",
+    "print(len(List_val))\n",
+    "print(len(List_train))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 17,
+   "id": "9efd0b25",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "<BOS>\n",
+      "My Beyoncé[Chorus: Lil Durk]\n",
+      "Ooh, I like the way she move\n",
+      "Shorty my baby, my everything, she the truth\n",
+      "Together we cool, me and her can't lose\n",
+      "Keep 'em on their feet, baby, I know they so confused\n",
+      "Shorty my Beyoncé\n",
+      "Durk and DeJ, Durk and DeJ, Durk and DeJ\n",
+      "Shorty my Beyoncé\n",
+      "Durk and DeJ, Durk and DeJ, Durk and DeJ\n",
+      "My Beyoncé\n",
+      "\n",
+      "[Verse 1: Lil Durk]\n",
+      "Trippin' on that drank, but I know she worth it\n",
+      "Independent baby, I know she workin'\n",
+      "Adriana's serving drinks, 20 bottles, urgent\n",
+      "I know it can be better but nobody's perfect\n",
+      "We flirted for a minute, DeJ, that's my baby\n",
+      "I ain't trippin', I'm like Henny, yeah I'm in her kidneys\n",
+      "She like to play her songs to the way I'm hittin' it\n",
+      "Turn around like, \"Damn Durk, I like the way you hittin' it\"\n",
+      "Don't believe the rumors, girl\n",
+      "You know I'll do you, girl\n",
+      "I don't wanna hear the shit about the niggas\n",
+      "That tried to do you, girl\n",
+      "Fuck the past right now\n",
+      "Shawty got you right now\n",
+      "And you hot right now\n",
+      "You can get it right now, baby\n",
+      "[Chorus: Lil Durk]\n",
+      "Ooh, I like the way she move\n",
+      "Shorty my baby, my everything, she the truth\n",
+      "Together we cool, me and her can't lose\n",
+      "Keep 'em on their feet, baby, I know they so confused\n",
+      "Shorty my Beyoncé\n",
+      "Durk and DeJ, Durk and DeJ, Durk and DeJ\n",
+      "Shorty my Beyoncé\n",
+      "Durk and DeJ, Durk and DeJ, Durk and DeJ\n",
+      "My Beyoncé\n",
+      "[Verse 2: DeJ Loaf]\n",
+      "I let him get it when he want it, knock it down and push up on it\n",
+      "I was plottin' for a while, now I got him where I want him\n",
+      "They didn't understand none of this was planned\n",
+      "99 problems but a bitch better not be none\n",
+      "Na na, na na, yeah yeah\n",
+      "This ain't what he want, I told him that\n",
+      "Leave your girl, be through with that\n",
+      "Get with DeJ, he ain't ever goin' back\n",
+      "He was shy when I seen him, now he smile\n",
+      "Heard a few rumors but they ain't my style\n",
+      "I be hatin' when he out of town\n",
+      "Hotel, I FaceTime you, no towel\n",
+      "They ain't get it but they ain't our problem\n",
+      "What the fuck can they do about it?\n",
+      "Durk and DeJ\n",
+      "I'm thinkin' 'bout changin' my last name\n",
+      "[Chorus: Lil Durk]\n",
+      "Ooh, I like the way she move\n",
+      "Shorty my baby, my everything, she the truth\n",
+      "Together we cool, me and her can't lose\n",
+      "Keep 'em on their feet, baby, I know they so confused\n",
+      "Shorty my Beyoncé\n",
+      "Durk and DeJ, Durk and DeJ, Durk and DeJ\n",
+      "Shorty my Beyoncé\n",
+      "Durk and DeJ, Durk and DeJ, Durk and DeJ\n",
+      "My Beyoncé\n",
+      "\n",
+      "[Verse 3: Lil Durk]\n",
+      "You and I\n",
+      "White dress, flowers, and a suit and tie\n",
+      "Me and you like Bonnie and Clyde\n",
+      "No beat the case, we're do or die\n",
+      "Who am I to say you ain't natural?\n",
+      "Your haters my haters, ain't switchin' up, baby, I got you\n",
+      "I'm with her like a tattoo\n",
+      "The way you wear that dress, they gon' attack you\n",
+      "The way you look at me, baby, I got you\n",
+      "I hit it from the front, I like the back too\n",
+      "She say, \"Lay down so I can ride you\"\n",
+      "I know that she fiending\n",
+      "She scratchin' my back, I like how she screamin'\n",
+      "I fuck her and leave her, she fiending\n",
+      "Shawty my Beyoncé\n",
+      "[Chorus: Lil Durk]\n",
+      "Ooh, I like the way she move\n",
+      "Shorty my baby, my everything, she the truth\n",
+      "Together we cool, me and her can't lose\n",
+      "Keep 'em on their feet, baby, I know they so confused\n",
+      "Shorty my Beyoncé\n",
+      "Durk and DeJ, Durk and DeJ, Durk and DeJ\n",
+      "Shorty my Beyoncé\n",
+      "Durk and DeJ, Durk and DeJ, Durk and DeJ\n",
+      "My Beyoncé\n",
+      "\n",
+      "[Outro: Lil Durk]\n",
+      "Ooh, ooh\n",
+      "Durk and DeJ, Durk and DeJ, Durk and DeJ\n",
+      "Ooh, ooh43EmbedShare URLCopyEmbedCopy\n",
+      "<EOS>\n"
+     ]
+    }
+   ],
+   "source": [
+    "print(List_val[0])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 18,
+   "id": "5dfecab1",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "val_set =List_val[0]\n",
+    "for i in range(1,len(List_val)):\n",
+    "    val_set = val_set+List_val[i]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 19,
+   "id": "3a895d2f",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "train_set =List_train[0]\n",
+    "for i in range(1,len(List_train)):\n",
+    "    train_set = train_set+List_train[i]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 20,
+   "id": "74ea6efc",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "file1 = open(\"train.txt\",\"w+\")\n",
+    "file1.write(train_set)\n",
+    "file1.close()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 21,
+   "id": "3416e6da",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "file2 = open(\"val.txt\",\"w+\")\n",
+    "file2.write(val_set)\n",
+    "file2.close()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "1bd0ca86",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "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": "3.8.10"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

config.json

@@ -0,0 +1,36 @@
+{
+  "activation_function": "gelu_new",
+  "architectures": [
+    "GPT2LMHeadModel"
+  ],
+  "attn_pdrop": 0.1,
+  "bos_token_id": 50256,
+  "embd_pdrop": 0.1,
+  "eos_token_id": 50256,
+  "gradient_checkpointing": false,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "gpt2",
+  "n_ctx": 1024,
+  "n_embd": 768,
+  "n_head": 12,
+  "n_inner": null,
+  "n_layer": 12,
+  "n_positions": 1024,
+  "resid_pdrop": 0.1,
+  "scale_attn_weights": true,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "task_specific_params": {
+    "text-generation": {
+      "do_sample": true,
+      "max_length": 50
+    }
+  },
+  "transformers_version": "4.9.0.dev0",
+  "use_cache": true,
+  "vocab_size": 50257
+}

flax_model.msgpack

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e12947417fc788cb2b8252765ca2297987abde75ee501fbb7dfa41a2e5a4d68
+size 497764120

run.sh

@@ -0,0 +1,20 @@
+#!/bin/env bash
+python3 run_clm_flax.py \
+    --output_dir="./" \
+    --model_type="gpt2" \
+    --config_name="./" \
+    --tokenizer_name="./" \
+    --train_file="/home/anantshankhdhar/gpt2-rap-lyric-generator/train.txt" \
+    --validation_file="/home/anantshankhdhar/gpt2-rap-lyric-generator/val.txt" \
+    --do_train \
+    --do_eval \
+    --block_size="512" \
+    --per_device_train_batch_size="64" \
+    --per_device_eval_batch_size="32" \
+    --learning_rate="5e-3"  \
+    --adam_beta1="0.9" --adam_beta2="0.98" --weight_decay="0.01" \
+    --overwrite_output_dir \
+    --num_train_epochs="50" \
+    --save_steps="100" \
+    --eval_steps="100" \
+    --push_to_hub

run_clm_flax.py

@@ -0,0 +1,640 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2021 The HuggingFace Team All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Pre-training/Fine-tuning the library models for causal language modeling (GPT, GPT-2, CTRL, ...) on a text file or a dataset.
+
+Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
+https://huggingface.co/models?filter=causal-lm
+"""
+# You can also adapt this script on your own causal language modeling task. Pointers for this are left as comments.
+
+import logging
+import math
+import os
+import sys
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Callable, Optional
+
+import datasets
+from datasets import Dataset, load_dataset
+from tqdm import tqdm
+
+import jax
+import jax.numpy as jnp
+import optax
+import transformers
+from flax import jax_utils, traverse_util
+from flax.jax_utils import unreplicate
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard, shard_prng_key
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_CAUSAL_LM_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    FlaxAutoModelForCausalLM,
+    HfArgumentParser,
+    TrainingArguments,
+    is_tensorboard_available,
+)
+from transformers.testing_utils import CaptureLogger
+
+
+logger = logging.getLogger(__name__)
+
+MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_CAUSAL_LM_MAPPING.keys())
+MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The model checkpoint for weights initialization."
+            "Don't set if you want to train a model from scratch."
+        },
+    )
+    model_type: Optional[str] = field(
+        default=None,
+        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+    dtype: Optional[str] = field(
+        default="float32",
+        metadata={
+            "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
+        },
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: Optional[str] = field(
+        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: Optional[str] = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    validation_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
+            "value if set."
+        },
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    validation_split_percentage: Optional[int] = field(
+        default=5,
+        metadata={
+            "help": "The percentage of the train set used as validation set in case there's no validation split"
+        },
+    )
+    block_size: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "Optional input sequence length after tokenization. "
+            "The training dataset will be truncated in block of this size for training. "
+            "Default to the model max input length for single sentence inputs (take into account special tokens)."
+        },
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+
+    def __post_init__(self):
+        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
+            raise ValueError("Need either a dataset name or a training/validation file.")
+        else:
+            if self.train_file is not None:
+                extension = self.train_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
+
+
+class TrainState(train_state.TrainState):
+    dropout_rng: jnp.ndarray
+
+    def replicate(self):
+        return jax_utils.replicate(self).replace(dropout_rng=shard_prng_key(self.dropout_rng))
+
+
+def data_loader(rng: jax.random.PRNGKey, dataset: Dataset, batch_size: int, shuffle: bool = False):
+    """
+    Returns batches of size `batch_size` from truncated `dataset`, sharded over all local devices.
+    Shuffle batches if `shuffle` is `True`.
+    """
+    steps_per_epoch = len(dataset) // batch_size
+
+    if shuffle:
+        batch_idx = jax.random.permutation(rng, len(dataset))
+    else:
+        batch_idx = jnp.arange(len(dataset))
+
+    batch_idx = batch_idx[: steps_per_epoch * batch_size]  # Skip incomplete batch.
+    batch_idx = batch_idx.reshape((steps_per_epoch, batch_size))
+
+    for idx in batch_idx:
+        batch = dataset[idx]
+        batch = {k: jnp.array(v) for k, v in batch.items()}
+
+        batch = shard(batch)
+
+        yield batch
+
+
+def write_train_metric(summary_writer, train_metrics, train_time, step):
+    summary_writer.scalar("train_time", train_time, step)
+
+    train_metrics = get_metrics(train_metrics)
+    for key, vals in train_metrics.items():
+        tag = f"train_{key}"
+        for i, val in enumerate(vals):
+            summary_writer.scalar(tag, val, step - len(vals) + i + 1)
+
+
+def write_eval_metric(summary_writer, eval_metrics, step):
+    for metric_name, value in eval_metrics.items():
+        summary_writer.scalar(f"eval_{metric_name}", value, step)
+
+
+def create_learning_rate_fn(
+    train_ds_size: int, train_batch_size: int, num_train_epochs: int, num_warmup_steps: int, learning_rate: float
+) -> Callable[[int], jnp.array]:
+    """Returns a linear warmup, linear_decay learning rate function."""
+    steps_per_epoch = train_ds_size // train_batch_size
+    num_train_steps = steps_per_epoch * num_train_epochs
+    warmup_fn = optax.linear_schedule(init_value=0.0, end_value=learning_rate, transition_steps=num_warmup_steps)
+    decay_fn = optax.linear_schedule(
+        init_value=learning_rate, end_value=0, transition_steps=num_train_steps - num_warmup_steps
+    )
+    schedule_fn = optax.join_schedules(schedules=[warmup_fn, decay_fn], boundaries=[num_warmup_steps])
+    return schedule_fn
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Make one log on every process with the configuration for debugging.
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    # Setup logging, we only want one process per machine to log things on the screen.
+    logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
+    if jax.process_index() == 0:
+        datasets.utils.logging.set_verbosity_warning()
+        transformers.utils.logging.set_verbosity_info()
+    else:
+        datasets.utils.logging.set_verbosity_error()
+        transformers.utils.logging.set_verbosity_error()
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    #  Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+    #
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+    #
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        dataset = load_dataset(
+            data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir, keep_in_memory=False
+        )
+
+        if "validation" not in dataset.keys():
+            dataset["validation"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            dataset["train"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+    else:
+        data_files = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+        extension = data_args.train_file.split(".")[-1]
+        if extension == "txt":
+            extension = "text"
+        dataset = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    if model_args.config_name:
+        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
+    elif model_args.model_name_or_path:
+        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
+    else:
+        config = CONFIG_MAPPING[model_args.model_type]()
+        logger.warning("You are instantiating a new config instance from scratch.")
+
+    if model_args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    elif model_args.model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    else:
+        raise ValueError(
+            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
+            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
+        )
+
+    if model_args.model_name_or_path:
+        model = FlaxAutoModelForCausalLM.from_pretrained(
+            model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+    else:
+        model = FlaxAutoModelForCausalLM.from_config(
+            config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+    if training_args.do_train:
+        column_names = dataset["train"].column_names
+    else:
+        column_names = dataset["validation"].column_names
+    text_column_name = "text" if "text" in column_names else column_names[0]
+
+    # since this will be pickled to avoid _LazyModule error in Hasher force logger loading before tokenize_function
+    tok_logger = transformers.utils.logging.get_logger("transformers.tokenization_utils_base")
+
+    def tokenize_function(examples):
+        with CaptureLogger(tok_logger) as cl:
+            output = tokenizer(examples[text_column_name])
+        # clm input could be much much longer than block_size
+        if "Token indices sequence length is longer than the" in cl.out:
+            tok_logger.warning(
+                "^^^^^^^^^^^^^^^^ Please ignore the warning above - this long input will be chunked into smaller bits before being passed to the model."
+            )
+        return output
+
+    tokenized_datasets = dataset.map(
+        tokenize_function,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        remove_columns=column_names,
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    if data_args.block_size is None:
+        block_size = tokenizer.model_max_length
+        if block_size > config.max_position_embeddings:
+            logger.warning(
+                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
+                "Picking 1024 instead. You can change that default value by passing --block_size xxx."
+            )
+            block_size = 1024
+    else:
+        if data_args.block_size > tokenizer.model_max_length:
+            logger.warning(
+                f"The block_size passed ({data_args.block_size}) is larger than the maximum length for the model"
+                f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}."
+            )
+        block_size = min(data_args.block_size, tokenizer.model_max_length)
+
+    # Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
+    def group_texts(examples):
+        # Concatenate all texts.
+        concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
+        total_length = len(concatenated_examples[list(examples.keys())[0]])
+        # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
+        # customize this part to your needs.
+        total_length = (total_length // block_size) * block_size
+        # Split by chunks of max_len.
+        result = {
+            k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
+            for k, t in concatenated_examples.items()
+        }
+        result["labels"] = result["input_ids"].copy()
+        return result
+
+    # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder
+    # for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value might be slower
+    # to preprocess.
+    #
+    # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
+    # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
+
+    lm_datasets = tokenized_datasets.map(
+        group_texts,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    if training_args.do_train:
+        if "train" not in tokenized_datasets:
+            raise ValueError("--do_train requires a train dataset")
+        train_dataset = lm_datasets["train"]
+        if data_args.max_train_samples is not None:
+            train_dataset = train_dataset.select(range(data_args.max_train_samples))
+
+    if training_args.do_eval:
+        if "validation" not in tokenized_datasets:
+            raise ValueError("--do_eval requires a validation dataset")
+        eval_dataset = lm_datasets["validation"]
+        if data_args.max_eval_samples is not None:
+            eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))
+
+    # Enable tensorboard only on the master node
+    has_tensorboard = is_tensorboard_available()
+    if has_tensorboard and jax.process_index() == 0:
+        try:
+            from flax.metrics.tensorboard import SummaryWriter
+
+            summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
+        except ImportError as ie:
+            has_tensorboard = False
+            logger.warning(
+                f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
+            )
+    else:
+        logger.warning(
+            "Unable to display metrics through TensorBoard because the package is not installed: "
+            "Please run pip install tensorboard to enable."
+        )
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(training_args.seed)
+    rng, dropout_rng = jax.random.split(rng)
+
+    # Store some constant
+    num_epochs = int(training_args.num_train_epochs)
+    train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
+    eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
+    steps_per_epoch = len(train_dataset) // train_batch_size
+    total_train_steps = steps_per_epoch * num_epochs
+
+    # Create learning rate schedule
+    linear_decay_lr_schedule_fn = create_learning_rate_fn(
+        len(train_dataset),
+        train_batch_size,
+        training_args.num_train_epochs,
+        training_args.warmup_steps,
+        training_args.learning_rate,
+    )
+
+    # We use Optax's "masking" functionality to not apply weight decay
+    # to bias and LayerNorm scale parameters. decay_mask_fn returns a
+    # mask boolean with the same structure as the parameters.
+    # The mask is True for parameters that should be decayed.
+    # Note that this mask is specifically adapted for FlaxGPT2.
+    # For other models, one should correct the layer norm parameter naming
+    # accordingly.
+    def decay_mask_fn(params):
+        flat_params = traverse_util.flatten_dict(params)
+        flat_mask = {
+            path: (path[-1] != "bias" and path[-2:] not in [("ln_1", "scale"), ("ln_2", "scale"), ("ln_f", "scale")])
+            for path in flat_params
+        }
+        return traverse_util.unflatten_dict(flat_mask)
+
+    # create adam optimizer
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = optax.adamw(
+            learning_rate=linear_decay_lr_schedule_fn,
+            b1=training_args.adam_beta1,
+            b2=training_args.adam_beta2,
+            eps=training_args.adam_epsilon,
+            weight_decay=training_args.weight_decay,
+            mask=decay_mask_fn,
+        )
+
+    # Setup train state
+    state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer, dropout_rng=dropout_rng)
+
+    def loss_fn(logits, labels):
+        shift_logits = logits[..., :-1, :]
+        shift_labels = labels[..., 1:]
+        loss = optax.softmax_cross_entropy(shift_logits, onehot(shift_labels, shift_logits.shape[-1]))
+        return loss.mean()
+
+    # Define gradient update step fn
+    def train_step(state, batch):
+        dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
+
+        def compute_loss(params):
+            labels = batch.pop("labels")
+            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
+            loss = loss_fn(logits, labels)
+            return loss
+
+        grad_fn = jax.value_and_grad(compute_loss)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+
+        new_state = state.apply_gradients(grads=grad, dropout_rng=new_dropout_rng)
+
+        metrics = {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+
+        return new_state, metrics
+
+    # Define eval fn
+    def eval_step(params, batch):
+        labels = batch.pop("labels")
+        logits = model(**batch, params=params, train=False)[0]
+        loss = loss_fn(logits, labels)
+
+        # summarize metrics
+        metrics = {"loss": loss}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+        return metrics
+
+    # Create parallel version of the train and eval step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+    p_eval_step = jax.pmap(eval_step, "batch")
+
+    # Replicate the train state on each device
+    state = state.replicate()
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num Epochs = {num_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel & distributed) = {train_batch_size}")
+    logger.info(f"  Total optimization steps = {total_train_steps}")
+
+    train_time = 0
+    train_metrics = []
+    epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
+    for epoch in epochs:
+        # ======================== Training ================================
+        train_start = time.time()
+
+        # Create sampling rng
+        rng, input_rng = jax.random.split(rng)
+
+        # Generate an epoch by shuffling sampling indices from the train dataset
+        train_loader = data_loader(input_rng, train_dataset, train_batch_size, shuffle=True)
+        steps_per_epoch = len(train_dataset) // train_batch_size
+        # train
+        for step in tqdm(range(steps_per_epoch), desc="Training...", position=1, leave=False):
+            batch = next(train_loader)
+            state, train_metric = p_train_step(state, batch)
+            train_metrics.append(train_metric)
+
+            cur_step = epoch * (len(train_dataset) // train_batch_size) + step
+
+            if cur_step % training_args.logging_steps == 0 and cur_step > 0:
+                # Save metrics
+                train_metric = unreplicate(train_metric)
+                train_time += time.time() - train_start
+                if has_tensorboard and jax.process_index() == 0:
+                    write_train_metric(summary_writer, train_metrics, train_time, cur_step)
+
+                epochs.write(
+                    f"Step... ({cur_step} | Loss: {train_metric['loss'].mean()}, Learning Rate: {train_metric['learning_rate'].mean()})"
+                )
+
+                train_metrics = []
+
+            if cur_step % training_args.eval_steps == 0 and cur_step > 0:
+                # ======================== Evaluating ==============================
+                eval_metrics = []
+                eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size)
+                eval_steps = len(eval_dataset) // eval_batch_size
+                for _ in tqdm(range(eval_steps), desc="Evaluating...", position=2, leave=False):
+                    # Model forward
+                    batch = next(eval_loader)
+                    metrics = p_eval_step(state.params, batch)
+                    eval_metrics.append(metrics)
+
+                # normalize eval metrics
+                eval_metrics = get_metrics(eval_metrics)
+                eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
+
+                try:
+                    eval_metrics["perplexity"] = math.exp(eval_metrics["loss"])
+                except OverflowError:
+                    eval_metrics["perplexity"] = float("inf")
+
+                # Print metrics and update progress bar
+                desc = f"Step... ({cur_step} | Eval Loss: {eval_metrics['loss']} | Eval Perplexity: {eval_metrics['perplexity']})"
+                epochs.write(desc)
+                epochs.desc = desc
+
+                # Save metrics
+                if has_tensorboard and jax.process_index() == 0:
+                    cur_step = epoch * (len(train_dataset) // train_batch_size)
+                    write_eval_metric(summary_writer, eval_metrics, cur_step)
+
+            if cur_step % training_args.save_steps == 0 and cur_step > 0:
+                # save checkpoint after each epoch and push checkpoint to the hub
+                if jax.process_index() == 0:
+                    params = jax.device_get(unreplicate(state.params))
+                    model.save_pretrained(
+                        training_args.output_dir,
+                        params=params,
+                        push_to_hub=training_args.push_to_hub,
+                        commit_message=f"Saving weights and logs of step {cur_step}",
+                    )
+
+
+if __name__ == "__main__":
+    main()