codestral-22b-fsdp-qlora-remote-decorator_qa.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Fine-tune Codestral 22B with PyTorch FSDP, QLoRA and SageMaker remote decorator\n",
    "\n",
    "## Question & Answering"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "---\n",
    "\n",
    "In this demo notebook, we demonstrate how to fine-tune the Codestral 22B model using PyTorch FSDP, QLoRA, Hugging Face PEFT, and bitsandbytes.\n",
    "\n",
    "We are using SageMaker remote decorator for runinng the fine-tuning job on Amazon SageMaker Training job\n",
    "---\n",
    "\n",
    "JupyterLab Instance Type: ml.t3.medium\n",
    "\n",
    "Python version: 3.11\n",
    "\n",
    "Fine-Tuning:\n",
    "* Instance Type: ml.g5.24xlarge"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Install the required libriaries, including the Hugging Face libraries, and restart the kernel."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-07-20T12:38:06.851473Z",
     "start_time": "2023-07-20T12:38:04.440644Z"
    },
    "collapsed": false,
    "jupyter": {
     "outputs_hidden": false
    },
    "scrolled": true,
    "tags": []
   },
   "outputs": [],
   "source": [
    "%pip install -r requirements.txt --upgrade\n",
    "%pip install -q -U python-dotenv"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "\n",
    "## Setup Configuration file path\n",
    "\n",
    "We are setting the directory in which the config.yaml file resides so that remote decorator can make use of the settings through [SageMaker Defaults](https://sagemaker.readthedocs.io/en/stable/overview.html#configuring-and-using-defaults-with-the-sagemaker-python-sdk).\n",
    "\n",
    "This notebook is using the Hugging Face container for the `us-east-1` region. Make sure you are using the right image for your AWS region, otherwise edit [config.yaml](./config.yaml). Container Images are available [here](https://github.com/aws/deep-learning-containers/blob/master/available_images.md)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-11-15T09:24:51.291677Z",
     "start_time": "2023-11-15T09:24:51.282905Z"
    },
    "collapsed": false,
    "jupyter": {
     "outputs_hidden": false
    },
    "tags": []
   },
   "outputs": [],
   "source": [
    "from dotenv import load_dotenv\n",
    "import os\n",
    "\n",
    "# Use .env in case of hidden variables\n",
    "load_dotenv()\n",
    "\n",
    "# Set path to config file\n",
    "os.environ[\"SAGEMAKER_USER_CONFIG_OVERRIDE\"] = os.getcwd()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Visualize and upload the dataset\n",
    "\n",
    "We are going to load [ed001/ds-coder-instruct-v1](https://huggingface.co/datasets/ed001/ds-coder-instruct-v1) dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-11-15T09:24:55.572481Z",
     "start_time": "2023-11-15T09:24:52.575954Z"
    },
    "collapsed": false,
    "jupyter": {
     "outputs_hidden": false
    },
    "tags": []
   },
   "outputs": [],
   "source": [
    "from datasets import load_dataset\n",
    "import pandas as pd\n",
    "\n",
    "dataset = load_dataset(\"ed001/ds-coder-instruct-v1\")\n",
    "df = pd.DataFrame(dataset['train'])\n",
    "df = df.iloc[0:1000]\n",
    "df = df[['instruction', 'output']].rename(columns={\n",
    "    'instruction': 'question',\n",
    "    'output': 'answer'\n",
    "})\n",
    "\n",
    "df.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-09-03T00:02:00.270970Z",
     "start_time": "2023-09-03T00:01:59.205060Z"
    },
    "collapsed": false,
    "jupyter": {
     "outputs_hidden": false
    },
    "tags": []
   },
   "outputs": [],
   "source": [
    "from sklearn.model_selection import train_test_split\n",
    "\n",
    "train, test = train_test_split(df, test_size=0.1, random_state=42)\n",
    "\n",
    "print(\"Number of train elements: \", len(train))\n",
    "print(\"Number of test elements: \", len(test))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Create a prompt template and load the dataset with a random sample to try summarization."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-09-03T00:02:01.435195Z",
     "start_time": "2023-09-03T00:02:01.429794Z"
    },
    "collapsed": false,
    "jupyter": {
     "outputs_hidden": false
    },
    "tags": []
   },
   "outputs": [],
   "source": [
    "from random import randint\n",
    "\n",
    "# custom instruct prompt start\n",
    "prompt_template = f\"<s>[INST] {{question}} [/INST] {{answer}} </s>\"\n",
    "\n",
    "# template dataset to add prompt to each sample\n",
    "def template_dataset(sample):\n",
    "    sample[\"text\"] = prompt_template.format(question=sample[\"question\"],\n",
    "                                            answer=sample[\"answer\"])\n",
    "    return sample"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Use the Hugging Face Trainer class to fine-tune the model. Define the hyperparameters we want to use. We also create a DataCollator that will take care of padding our inputs and labels."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-09-03T00:02:10.364459Z",
     "start_time": "2023-09-03T00:02:09.672705Z"
    },
    "tags": []
   },
   "outputs": [],
   "source": [
    "from datasets import Dataset, DatasetDict\n",
    "\n",
    "train_dataset = Dataset.from_pandas(train)\n",
    "test_dataset = Dataset.from_pandas(test)\n",
    "\n",
    "dataset = DatasetDict({\"train\": train_dataset, \"test\": test_dataset})\n",
    "\n",
    "train_dataset = dataset[\"train\"].map(template_dataset, remove_columns=list(dataset[\"train\"].features))\n",
    "\n",
    "print(train_dataset[randint(0, len(dataset))][\"text\"])\n",
    "\n",
    "test_dataset = dataset[\"test\"].map(template_dataset, remove_columns=list(dataset[\"test\"].features))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "\n",
    "\n",
    "To train our model, we need to convert our inputs (text) to token IDs. This is done by a Hugging Face Transformers Tokenizer. In addition to QLoRA, we will use bitsanbytes 4-bit precision to quantize out frozen LLM to 4-bit and attach LoRA adapters on it.\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Get PyTorch image_uri\n",
    "\n",
    "We are going to use the native PyTorch container image, pre-built for Amazon SageMaker"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import sagemaker\n",
    "from sagemaker import image_uris\n",
    "from sagemaker.config import load_sagemaker_config"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "sagemaker_session = sagemaker.Session()\n",
    "configs = load_sagemaker_config()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "instance_type = \"ml.g5.24xlarge\" # Override the instance type if you want to get a different container version\n",
    "\n",
    "if instance_type is None:\n",
    "    if \"InstanceType\" in configs[\"SageMaker\"][\"PythonSDK\"][\"Modules\"][\"RemoteFunction\"]:\n",
    "        instance_type = configs[\"SageMaker\"][\"PythonSDK\"][\"Modules\"][\"RemoteFunction\"][\"InstanceType\"]\n",
    "\n",
    "instance_type"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "image_uri = image_uris.retrieve(\n",
    "    framework=\"pytorch\",\n",
    "    image_scope=\"training\",\n",
    "    instance_type=instance_type,\n",
    "    region=sagemaker_session.boto_session.region_name, \n",
    "    version=\"2.4\"\n",
    ")\n",
    "\n",
    "image_uri"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Define the train function"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model_id = \"mistralai/Codestral-22B-v0.1\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-09-03T00:02:21.382486Z",
     "start_time": "2023-09-03T00:02:20.962208Z"
    },
    "collapsed": false,
    "jupyter": {
     "outputs_hidden": false
    },
    "tags": []
   },
   "outputs": [],
   "source": [
    "from accelerate import Accelerator\n",
    "import datetime\n",
    "from huggingface_hub import snapshot_download\n",
    "import os\n",
    "from peft import AutoPeftModelForCausalLM, LoraConfig, get_peft_model, prepare_model_for_kbit_training\n",
    "from sagemaker.remote_function import remote\n",
    "import torch\n",
    "from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, set_seed\n",
    "import transformers\n",
    "\n",
    "# Start training\n",
    "@remote(\n",
    "    keep_alive_period_in_seconds=0,\n",
    "    image_uri=image_uri,\n",
    "    instance_type=instance_type,\n",
    "    volume_size=100,\n",
    "    job_name_prefix=f\"train-{model_id.split('/')[-1].replace('.', '-')}\", \n",
    "    use_torchrun=True, #For distribution\n",
    "    nproc_per_node=4 #GPU number\n",
    ")\n",
    "def train_fn(\n",
    "        model_name,\n",
    "        train_ds,\n",
    "        test_ds=None,\n",
    "        lora_r=8,\n",
    "        lora_alpha=16,\n",
    "        lora_dropout=0.1,\n",
    "        per_device_train_batch_size=8,\n",
    "        per_device_eval_batch_size=8,\n",
    "        gradient_accumulation_steps=1,\n",
    "        learning_rate=2e-4,\n",
    "        num_train_epochs=1,\n",
    "        fsdp=\"\",\n",
    "        fsdp_config=None,\n",
    "        gradient_checkpointing=False,\n",
    "        merge_weights=False,\n",
    "        seed=42,\n",
    "        token=None\n",
    "):\n",
    "    def init_distributed():\n",
    "        # Initialize the process group\n",
    "        torch.distributed.init_process_group(\n",
    "            backend=\"nccl\", # Use \"gloo\" backend for CPU\n",
    "            timeout=datetime.timedelta(seconds=5400)\n",
    "        )\n",
    "        local_rank = int(os.environ[\"LOCAL_RANK\"])\n",
    "        torch.cuda.set_device(local_rank)\n",
    "\n",
    "        return local_rank\n",
    "\n",
    "    if torch.cuda.is_available() and (torch.cuda.device_count() > 1 or int(os.environ.get(\"SM_HOST_COUNT\", 1)) > 1):\n",
    "        # Call this function at the beginning of your script\n",
    "        local_rank = init_distributed()\n",
    "\n",
    "        # Now you can use distributed functionalities\n",
    "        torch.distributed.barrier(device_ids=[local_rank])\n",
    "\n",
    "    os.environ.update({\"HF_HUB_ENABLE_HF_TRANSFER\": \"1\"})\n",
    "\n",
    "    set_seed(seed)\n",
    "\n",
    "    accelerator = Accelerator()\n",
    "\n",
    "    if token is not None:\n",
    "        os.environ.update({\"HF_TOKEN\": token})\n",
    "        accelerator.wait_for_everyone()\n",
    "\n",
    "    print(\"Downloading model \", model_name)\n",
    "\n",
    "    os.makedirs(\"/tmp/tmp_folder\", exist_ok=True)\n",
    "\n",
    "    snapshot_download(repo_id=model_name, local_dir=\"/tmp/tmp_folder\")\n",
    "\n",
    "    print(f\"Downloading model {model_name} under /tmp/tmp_folder\")\n",
    "\n",
    "    model_name = \"/tmp/tmp_folder\"\n",
    "\n",
    "    tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
    "\n",
    "    # Set Tokenizer pad Token\n",
    "    tokenizer.pad_token = tokenizer.eos_token\n",
    "\n",
    "    with accelerator.main_process_first():\n",
    "        # tokenize and chunk dataset\n",
    "        lm_train_dataset = train_ds.map(\n",
    "            lambda sample: tokenizer(sample[\"text\"]), remove_columns=list(train_ds.features)\n",
    "        )\n",
    "\n",
    "        print(f\"Total number of train samples: {len(lm_train_dataset)}\")\n",
    "\n",
    "        if test_ds is not None:\n",
    "            lm_test_dataset = test_ds.map(\n",
    "                lambda sample: tokenizer(sample[\"text\"]), remove_columns=list(test_ds.features)\n",
    "            )\n",
    "\n",
    "            print(f\"Total number of test samples: {len(lm_test_dataset)}\")\n",
    "        else:\n",
    "            lm_test_dataset = None\n",
    "\n",
    "    torch_dtype = torch.bfloat16\n",
    "\n",
    "    # Defining additional configs for FSDP\n",
    "    if fsdp != \"\" and fsdp_config is not None:\n",
    "        bnb_config_params = {\n",
    "            \"bnb_4bit_quant_storage\": torch_dtype\n",
    "        }\n",
    "\n",
    "        model_configs = {\n",
    "            \"torch_dtype\": torch_dtype\n",
    "        }\n",
    "\n",
    "        fsdp_configurations = {\n",
    "            \"fsdp\": fsdp,\n",
    "            \"fsdp_config\": fsdp_config,\n",
    "            \"gradient_checkpointing_kwargs\": {\n",
    "                \"use_reentrant\": False\n",
    "            },\n",
    "            \"tf32\": True\n",
    "        }\n",
    "    else:\n",
    "        bnb_config_params = dict()\n",
    "        model_configs = dict()\n",
    "        fsdp_configurations = dict()\n",
    "\n",
    "    bnb_config = BitsAndBytesConfig(\n",
    "        load_in_4bit=True,\n",
    "        bnb_4bit_use_double_quant=True,\n",
    "        bnb_4bit_quant_type=\"nf4\",\n",
    "        bnb_4bit_compute_dtype=torch_dtype,\n",
    "        **bnb_config_params\n",
    "    )\n",
    "\n",
    "    model = AutoModelForCausalLM.from_pretrained(\n",
    "        model_name,\n",
    "        trust_remote_code=True,\n",
    "        quantization_config=bnb_config,\n",
    "        attn_implementation=\"flash_attention_2\",\n",
    "        use_cache=not gradient_checkpointing,\n",
    "        cache_dir=\"/tmp/.cache\",\n",
    "        **model_configs\n",
    "    )\n",
    "\n",
    "    if fsdp == \"\" and fsdp_config is None:\n",
    "        model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=gradient_checkpointing)\n",
    "\n",
    "    if gradient_checkpointing:\n",
    "        model.gradient_checkpointing_enable()\n",
    "\n",
    "    config = LoraConfig(\n",
    "        r=lora_r,\n",
    "        lora_alpha=lora_alpha,\n",
    "        target_modules=\"all-linear\",\n",
    "        lora_dropout=lora_dropout,\n",
    "        bias=\"none\",\n",
    "        task_type=\"CAUSAL_LM\"\n",
    "    )\n",
    "\n",
    "    model = get_peft_model(model, config)\n",
    "\n",
    "    trainer = transformers.Trainer(\n",
    "        model=model,\n",
    "        train_dataset=lm_train_dataset,\n",
    "        eval_dataset=lm_test_dataset if lm_test_dataset is not None else None,\n",
    "        args=transformers.TrainingArguments(\n",
    "            per_device_train_batch_size=per_device_train_batch_size,\n",
    "            per_device_eval_batch_size=per_device_eval_batch_size,\n",
    "            gradient_accumulation_steps=gradient_accumulation_steps,\n",
    "            gradient_checkpointing=gradient_checkpointing,\n",
    "            logging_strategy=\"steps\",\n",
    "            logging_steps=1,\n",
    "            log_on_each_node=False,\n",
    "            num_train_epochs=num_train_epochs,\n",
    "            learning_rate=learning_rate,\n",
    "            bf16=True,\n",
    "            ddp_find_unused_parameters=False,\n",
    "            save_strategy=\"no\",\n",
    "            output_dir=\"outputs\",\n",
    "            **fsdp_configurations\n",
    "        ),\n",
    "        data_collator=transformers.DataCollatorForLanguageModeling(tokenizer, mlm=False),\n",
    "    )\n",
    "\n",
    "    if trainer.accelerator.is_main_process:\n",
    "        trainer.model.print_trainable_parameters()\n",
    "\n",
    "    if trainer.is_fsdp_enabled:\n",
    "        trainer.accelerator.state.fsdp_plugin.set_state_dict_type(\"FULL_STATE_DICT\")\n",
    "    \n",
    "    trainer.train()\n",
    "\n",
    "    if merge_weights:\n",
    "        output_dir = \"/tmp/model\"\n",
    "\n",
    "        # merge adapter weights with base model and save\n",
    "        # save int 4 model\n",
    "        trainer.model.save_pretrained(output_dir, safe_serialization=False)\n",
    "\n",
    "        if accelerator.is_main_process:\n",
    "            # clear memory\n",
    "            del model\n",
    "            del trainer\n",
    "\n",
    "            torch.cuda.empty_cache()\n",
    "\n",
    "            # load PEFT model\n",
    "            model = AutoPeftModelForCausalLM.from_pretrained(\n",
    "                output_dir,\n",
    "                torch_dtype=torch.float16,\n",
    "                low_cpu_mem_usage=True,\n",
    "                trust_remote_code=True,\n",
    "            )\n",
    "\n",
    "            # Merge LoRA and base model and save\n",
    "            model = model.merge_and_unload()\n",
    "            model.save_pretrained(\n",
    "                \"/opt/ml/model\", safe_serialization=True, max_shard_size=\"2GB\"\n",
    "            )\n",
    "    else:\n",
    "        trainer.model.save_pretrained(\"/opt/ml/model\", safe_serialization=True)\n",
    "\n",
    "    if accelerator.is_main_process:\n",
    "        tokenizer.save_pretrained(\"/opt/ml/model\")\n",
    "\n",
    "    accelerator.wait_for_everyone()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "jupyter": {
     "outputs_hidden": false
    },
    "scrolled": true,
    "tags": []
   },
   "outputs": [],
   "source": [
    "train_fn(\n",
    "    model_id,\n",
    "    train_ds=train_dataset,\n",
    "    test_ds=test_dataset,\n",
    "    per_device_train_batch_size=2,\n",
    "    per_device_eval_batch_size=1,\n",
    "    gradient_accumulation_steps=2,\n",
    "    gradient_checkpointing=True,\n",
    "    num_train_epochs=2,\n",
    "    fsdp=\"full_shard auto_wrap offload\",\n",
    "    fsdp_config={\n",
    "        'backward_prefetch': 'backward_pre',\n",
    "        'cpu_ram_efficient_loading': True,\n",
    "        'offload_params': True,\n",
    "        'forward_prefetch': False,\n",
    "        'use_orig_params': False\n",
    "    },\n",
    "    merge_weights=True,\n",
    "    token=os.environ.get(\"HF_TOKEN\", None) # Change None to your HF_TOKEN in case you don't use .env\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "tags": []
   },
   "source": [
    "## Load Fine-Tuned model\n",
    "\n",
    "Note: Run `train_fn` with `merge_weights=True` for merging the trained adapter"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Download model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-11-20T18:41:08.515277Z",
     "start_time": "2023-11-20T18:41:08.503555Z"
    },
    "collapsed": false,
    "jupyter": {
     "outputs_hidden": false
    }
   },
   "outputs": [],
   "source": [
    "import sagemaker"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "sagemaker_session = sagemaker.Session()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-09-03T13:45:11.757861Z",
     "start_time": "2023-09-03T13:45:11.747993Z"
    },
    "collapsed": false,
    "jupyter": {
     "outputs_hidden": false
    },
    "tags": []
   },
   "outputs": [],
   "source": [
    "model_id = \"mistralai/Codestral-22B-v0.1\"\n",
    "\n",
    "bucket_name = sagemaker_session.default_bucket()\n",
    "default_prefix = sagemaker_session.default_bucket_prefix\n",
    "job_prefix = f\"train-{model_id.split('/')[-1].replace('.', '-')}\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_last_job_name(job_name_prefix):\n",
    "    import boto3\n",
    "    sagemaker_client = boto3.client('sagemaker')\n",
    "\n",
    "    search_response = sagemaker_client.search(\n",
    "        Resource='TrainingJob',\n",
    "        SearchExpression={\n",
    "            'Filters': [\n",
    "                {\n",
    "                    'Name': 'TrainingJobName',\n",
    "                    'Operator': 'Contains',\n",
    "                    'Value': job_name_prefix\n",
    "                },\n",
    "                {\n",
    "                    'Name': 'TrainingJobStatus',\n",
    "                    'Operator': 'Equals',\n",
    "                    'Value': \"Completed\"\n",
    "                }\n",
    "            ]\n",
    "        },\n",
    "        SortBy='CreationTime',\n",
    "        SortOrder='Descending',\n",
    "        MaxResults=1)\n",
    "\n",
    "    return search_response['Results'][0]['TrainingJob']['TrainingJobName']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "job_name = get_last_job_name(job_prefix)\n",
    "\n",
    "job_name"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false,
    "jupyter": {
     "outputs_hidden": false
    }
   },
   "source": [
    "#### Inference configurations"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import sagemaker\n",
    "from sagemaker import get_execution_role\n",
    "from sagemaker import Model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-11-20T18:41:10.492877Z",
     "start_time": "2023-11-20T18:41:10.488495Z"
    },
    "collapsed": false,
    "jupyter": {
     "outputs_hidden": false
    }
   },
   "outputs": [],
   "source": [
    "instance_count = 1\n",
    "instance_type = \"ml.g5.12xlarge\"\n",
    "number_of_gpu = 4\n",
    "health_check_timeout = 700"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-11-20T18:41:11.162811Z",
     "start_time": "2023-11-20T18:41:10.787936Z"
    },
    "collapsed": false,
    "jupyter": {
     "outputs_hidden": false
    },
    "tags": []
   },
   "outputs": [],
   "source": [
    "image_uri = sagemaker.image_uris.retrieve(\n",
    "    framework=\"djl-lmi\",\n",
    "    region=sagemaker_session.boto_session.region_name,\n",
    "    version=\"latest\"\n",
    ")\n",
    "\n",
    "image_uri"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-11-20T18:41:12.433399Z",
     "start_time": "2023-11-20T18:41:11.963091Z"
    },
    "collapsed": false,
    "jupyter": {
     "outputs_hidden": false
    },
    "tags": []
   },
   "outputs": [],
   "source": [
    "if default_prefix:\n",
    "    model_data = f\"s3://{bucket_name}/{default_prefix}/{job_name}/{job_name}/output/model.tar.gz\"\n",
    "else:\n",
    "    model_data = f\"s3://{bucket_name}/{job_name}/{job_name}/output/model.tar.gz\"\n",
    "\n",
    "model = Model(\n",
    "    image_uri=image_uri,\n",
    "    model_data=model_data,\n",
    "    role=get_execution_role(),\n",
    "    env={\n",
    "        'HF_MODEL_ID': \"/opt/ml/model\", # path to where sagemaker stores the model\n",
    "        'OPTION_TRUST_REMOTE_CODE': 'true',\n",
    "        'OPTION_ROLLING_BATCH': \"vllm\",\n",
    "        'OPTION_DTYPE': 'bf16',\n",
    "        'OPTION_TENSOR_PARALLEL_DEGREE': 'max',\n",
    "        'OPTION_MAX_ROLLING_BATCH_SIZE': '1',\n",
    "        'OPTION_MODEL_LOADING_TIMEOUT': '3600'\n",
    "    }\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-11-20T18:45:49.265298Z",
     "start_time": "2023-11-20T18:41:14.621743Z"
    },
    "collapsed": false,
    "jupyter": {
     "outputs_hidden": false
    },
    "tags": []
   },
   "outputs": [],
   "source": [
    "model.deploy(\n",
    "    initial_instance_count=instance_count,\n",
    "    instance_type=instance_type,\n",
    "    container_startup_health_check_timeout=health_check_timeout,\n",
    "    model_data_download_timeout=3600\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Predict"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import sagemaker"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "endpoint_name = \"<ENDPOINT_NAME>\" #Required if you want to create a predictor without running the previous code"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "predictor = sagemaker.Predictor(\n",
    "    endpoint_name=endpoint_name,\n",
    "    sagemaker_session=sagemaker_session,\n",
    "    serializer=sagemaker.serializers.JSONSerializer(),\n",
    "    deserializer=sagemaker.deserializers.JSONDeserializer(),\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "base_prompt = f\"<s>[INST] <<SYS>> {{system}} <<SYS>> {{question}} [/INST] \""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "question = \"Return the code for a ML model that recognizes handwirtten text.\"\n",
    "\n",
    "prompt = base_prompt.format(\n",
    "    question=question\n",
    ")\n",
    "\n",
    "predictor.predict({\n",
    "\t\"inputs\": prompt,\n",
    "    \"parameters\": {\n",
    "        \"max_new_tokens\": 300,\n",
    "        \"temperature\": 0.2,\n",
    "        \"top_p\": 0.9,\n",
    "        \"return_full_text\": False,\n",
    "        \"stop\": ['</s>']\n",
    "    }\n",
    "})"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false,
    "jupyter": {
     "outputs_hidden": false
    }
   },
   "source": [
    "#### Delete Endpoint"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-11-20T18:48:55.153276Z",
     "start_time": "2023-11-20T18:48:54.165351Z"
    },
    "collapsed": false,
    "jupyter": {
     "outputs_hidden": false
    },
    "tags": []
   },
   "outputs": [],
   "source": [
    "predictor.delete_model()\n",
    "predictor.delete_endpoint(delete_endpoint_config=True)"
   ]
  }
 ],
 "metadata": {
  "availableInstances": [
   {
    "_defaultOrder": 0,
    "_isFastLaunch": true,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 4,
    "name": "ml.t3.medium",
    "vcpuNum": 2
   },
   {
    "_defaultOrder": 1,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 8,
    "name": "ml.t3.large",
    "vcpuNum": 2
   },
   {
    "_defaultOrder": 2,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 16,
    "name": "ml.t3.xlarge",
    "vcpuNum": 4
   },
   {
    "_defaultOrder": 3,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 32,
    "name": "ml.t3.2xlarge",
    "vcpuNum": 8
   },
   {
    "_defaultOrder": 4,
    "_isFastLaunch": true,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 8,
    "name": "ml.m5.large",
    "vcpuNum": 2
   },
   {
    "_defaultOrder": 5,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 16,
    "name": "ml.m5.xlarge",
    "vcpuNum": 4
   },
   {
    "_defaultOrder": 6,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 32,
    "name": "ml.m5.2xlarge",
    "vcpuNum": 8
   },
   {
    "_defaultOrder": 7,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 64,
    "name": "ml.m5.4xlarge",
    "vcpuNum": 16
   },
   {
    "_defaultOrder": 8,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 128,
    "name": "ml.m5.8xlarge",
    "vcpuNum": 32
   },
   {
    "_defaultOrder": 9,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 192,
    "name": "ml.m5.12xlarge",
    "vcpuNum": 48
   },
   {
    "_defaultOrder": 10,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 256,
    "name": "ml.m5.16xlarge",
    "vcpuNum": 64
   },
   {
    "_defaultOrder": 11,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 384,
    "name": "ml.m5.24xlarge",
    "vcpuNum": 96
   },
   {
    "_defaultOrder": 12,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 8,
    "name": "ml.m5d.large",
    "vcpuNum": 2
   },
   {
    "_defaultOrder": 13,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 16,
    "name": "ml.m5d.xlarge",
    "vcpuNum": 4
   },
   {
    "_defaultOrder": 14,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 32,
    "name": "ml.m5d.2xlarge",
    "vcpuNum": 8
   },
   {
    "_defaultOrder": 15,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 64,
    "name": "ml.m5d.4xlarge",
    "vcpuNum": 16
   },
   {
    "_defaultOrder": 16,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 128,
    "name": "ml.m5d.8xlarge",
    "vcpuNum": 32
   },
   {
    "_defaultOrder": 17,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 192,
    "name": "ml.m5d.12xlarge",
    "vcpuNum": 48
   },
   {
    "_defaultOrder": 18,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 256,
    "name": "ml.m5d.16xlarge",
    "vcpuNum": 64
   },
   {
    "_defaultOrder": 19,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 384,
    "name": "ml.m5d.24xlarge",
    "vcpuNum": 96
   },
   {
    "_defaultOrder": 20,
    "_isFastLaunch": false,
    "category": "General purpose",
    "gpuNum": 0,
    "hideHardwareSpecs": true,
    "memoryGiB": 0,
    "name": "ml.geospatial.interactive",
    "supportedImageNames": [
     "sagemaker-geospatial-v1-0"
    ],
    "vcpuNum": 0
   },
   {
    "_defaultOrder": 21,
    "_isFastLaunch": true,
    "category": "Compute optimized",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 4,
    "name": "ml.c5.large",
    "vcpuNum": 2
   },
   {
    "_defaultOrder": 22,
    "_isFastLaunch": false,
    "category": "Compute optimized",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 8,
    "name": "ml.c5.xlarge",
    "vcpuNum": 4
   },
   {
    "_defaultOrder": 23,
    "_isFastLaunch": false,
    "category": "Compute optimized",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 16,
    "name": "ml.c5.2xlarge",
    "vcpuNum": 8
   },
   {
    "_defaultOrder": 24,
    "_isFastLaunch": false,
    "category": "Compute optimized",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 32,
    "name": "ml.c5.4xlarge",
    "vcpuNum": 16
   },
   {
    "_defaultOrder": 25,
    "_isFastLaunch": false,
    "category": "Compute optimized",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 72,
    "name": "ml.c5.9xlarge",
    "vcpuNum": 36
   },
   {
    "_defaultOrder": 26,
    "_isFastLaunch": false,
    "category": "Compute optimized",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 96,
    "name": "ml.c5.12xlarge",
    "vcpuNum": 48
   },
   {
    "_defaultOrder": 27,
    "_isFastLaunch": false,
    "category": "Compute optimized",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 144,
    "name": "ml.c5.18xlarge",
    "vcpuNum": 72
   },
   {
    "_defaultOrder": 28,
    "_isFastLaunch": false,
    "category": "Compute optimized",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 192,
    "name": "ml.c5.24xlarge",
    "vcpuNum": 96
   },
   {
    "_defaultOrder": 29,
    "_isFastLaunch": true,
    "category": "Accelerated computing",
    "gpuNum": 1,
    "hideHardwareSpecs": false,
    "memoryGiB": 16,
    "name": "ml.g4dn.xlarge",
    "vcpuNum": 4
   },
   {
    "_defaultOrder": 30,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 1,
    "hideHardwareSpecs": false,
    "memoryGiB": 32,
    "name": "ml.g4dn.2xlarge",
    "vcpuNum": 8
   },
   {
    "_defaultOrder": 31,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 1,
    "hideHardwareSpecs": false,
    "memoryGiB": 64,
    "name": "ml.g4dn.4xlarge",
    "vcpuNum": 16
   },
   {
    "_defaultOrder": 32,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 1,
    "hideHardwareSpecs": false,
    "memoryGiB": 128,
    "name": "ml.g4dn.8xlarge",
    "vcpuNum": 32
   },
   {
    "_defaultOrder": 33,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 4,
    "hideHardwareSpecs": false,
    "memoryGiB": 192,
    "name": "ml.g4dn.12xlarge",
    "vcpuNum": 48
   },
   {
    "_defaultOrder": 34,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 1,
    "hideHardwareSpecs": false,
    "memoryGiB": 256,
    "name": "ml.g4dn.16xlarge",
    "vcpuNum": 64
   },
   {
    "_defaultOrder": 35,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 1,
    "hideHardwareSpecs": false,
    "memoryGiB": 61,
    "name": "ml.p3.2xlarge",
    "vcpuNum": 8
   },
   {
    "_defaultOrder": 36,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 4,
    "hideHardwareSpecs": false,
    "memoryGiB": 244,
    "name": "ml.p3.8xlarge",
    "vcpuNum": 32
   },
   {
    "_defaultOrder": 37,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 8,
    "hideHardwareSpecs": false,
    "memoryGiB": 488,
    "name": "ml.p3.16xlarge",
    "vcpuNum": 64
   },
   {
    "_defaultOrder": 38,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 8,
    "hideHardwareSpecs": false,
    "memoryGiB": 768,
    "name": "ml.p3dn.24xlarge",
    "vcpuNum": 96
   },
   {
    "_defaultOrder": 39,
    "_isFastLaunch": false,
    "category": "Memory Optimized",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 16,
    "name": "ml.r5.large",
    "vcpuNum": 2
   },
   {
    "_defaultOrder": 40,
    "_isFastLaunch": false,
    "category": "Memory Optimized",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 32,
    "name": "ml.r5.xlarge",
    "vcpuNum": 4
   },
   {
    "_defaultOrder": 41,
    "_isFastLaunch": false,
    "category": "Memory Optimized",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 64,
    "name": "ml.r5.2xlarge",
    "vcpuNum": 8
   },
   {
    "_defaultOrder": 42,
    "_isFastLaunch": false,
    "category": "Memory Optimized",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 128,
    "name": "ml.r5.4xlarge",
    "vcpuNum": 16
   },
   {
    "_defaultOrder": 43,
    "_isFastLaunch": false,
    "category": "Memory Optimized",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 256,
    "name": "ml.r5.8xlarge",
    "vcpuNum": 32
   },
   {
    "_defaultOrder": 44,
    "_isFastLaunch": false,
    "category": "Memory Optimized",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 384,
    "name": "ml.r5.12xlarge",
    "vcpuNum": 48
   },
   {
    "_defaultOrder": 45,
    "_isFastLaunch": false,
    "category": "Memory Optimized",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 512,
    "name": "ml.r5.16xlarge",
    "vcpuNum": 64
   },
   {
    "_defaultOrder": 46,
    "_isFastLaunch": false,
    "category": "Memory Optimized",
    "gpuNum": 0,
    "hideHardwareSpecs": false,
    "memoryGiB": 768,
    "name": "ml.r5.24xlarge",
    "vcpuNum": 96
   },
   {
    "_defaultOrder": 47,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 1,
    "hideHardwareSpecs": false,
    "memoryGiB": 16,
    "name": "ml.g5.xlarge",
    "vcpuNum": 4
   },
   {
    "_defaultOrder": 48,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 1,
    "hideHardwareSpecs": false,
    "memoryGiB": 32,
    "name": "ml.g5.2xlarge",
    "vcpuNum": 8
   },
   {
    "_defaultOrder": 49,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 1,
    "hideHardwareSpecs": false,
    "memoryGiB": 64,
    "name": "ml.g5.4xlarge",
    "vcpuNum": 16
   },
   {
    "_defaultOrder": 50,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 1,
    "hideHardwareSpecs": false,
    "memoryGiB": 128,
    "name": "ml.g5.8xlarge",
    "vcpuNum": 32
   },
   {
    "_defaultOrder": 51,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 1,
    "hideHardwareSpecs": false,
    "memoryGiB": 256,
    "name": "ml.g5.16xlarge",
    "vcpuNum": 64
   },
   {
    "_defaultOrder": 52,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 4,
    "hideHardwareSpecs": false,
    "memoryGiB": 192,
    "name": "ml.g5.12xlarge",
    "vcpuNum": 48
   },
   {
    "_defaultOrder": 53,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 4,
    "hideHardwareSpecs": false,
    "memoryGiB": 384,
    "name": "ml.g5.24xlarge",
    "vcpuNum": 96
   },
   {
    "_defaultOrder": 54,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 8,
    "hideHardwareSpecs": false,
    "memoryGiB": 768,
    "name": "ml.g5.48xlarge",
    "vcpuNum": 192
   },
   {
    "_defaultOrder": 55,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 8,
    "hideHardwareSpecs": false,
    "memoryGiB": 1152,
    "name": "ml.p4d.24xlarge",
    "vcpuNum": 96
   },
   {
    "_defaultOrder": 56,
    "_isFastLaunch": false,
    "category": "Accelerated computing",
    "gpuNum": 8,
    "hideHardwareSpecs": false,
    "memoryGiB": 1152,
    "name": "ml.p4de.24xlarge",
    "vcpuNum": 96
   }
  ],
  "instance_type": "ml.g5.24xlarge",
  "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": "3.11.11"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}