notebooks/community/model_garden/model_garden_pytorch_llama3_finetuning.ipynb

{
  "cells": [
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "cellView": "form",
        "id": "7d9bbf86da5e"
      },
      "outputs": [],
      "source": [
        "# Copyright 2024 Google LLC\n",
        "#\n",
        "# Licensed under the Apache License, Version 2.0 (the \"License\");\n",
        "# you may not use this file except in compliance with the License.\n",
        "# You may obtain a copy of the License at\n",
        "#\n",
        "#     https://www.apache.org/licenses/LICENSE-2.0\n",
        "#\n",
        "# Unless required by applicable law or agreed to in writing, software\n",
        "# distributed under the License is distributed on an \"AS IS\" BASIS,\n",
        "# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n",
        "# See the License for the specific language governing permissions and\n",
        "# limitations under the License."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "99c1c3fc2ca5"
      },
      "source": [
        "# Vertex AI Model Garden - Llama 3 Finetuning\n",
        "\n",
        "<table><tbody><tr>\n",
        "  <td style=\"text-align: center\">\n",
        "    <a href=\"https://console.cloud.google.com/vertex-ai/colab/import/https:%2F%2Fraw.githubusercontent.com%2FGoogleCloudPlatform%2Fvertex-ai-samples%2Fmain%2Fnotebooks%2Fcommunity%2Fmodel_garden%2Fmodel_garden_pytorch_llama3_finetuning.ipynb\">\n",
        "      <img alt=\"Google Cloud Colab Enterprise logo\" src=\"https://lh3.googleusercontent.com/JmcxdQi-qOpctIvWKgPtrzZdJJK-J3sWE1RsfjZNwshCFgE_9fULcNpuXYTilIR2hjwN\" width=\"32px\"><br> Run in Colab Enterprise\n",
        "    </a>\n",
        "  </td>\n",
        "  <td style=\"text-align: center\">\n",
        "    <a href=\"https://github.com/GoogleCloudPlatform/vertex-ai-samples/blob/main/notebooks/community/model_garden/model_garden_pytorch_llama3_finetuning.ipynb\">\n",
        "      <img alt=\"GitHub logo\" src=\"https://cloud.google.com/ml-engine/images/github-logo-32px.png\" width=\"32px\"><br> View on GitHub\n",
        "    </a>\n",
        "  </td>\n",
        "</tr></tbody></table>"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "3de7470326a2"
      },
      "source": [
        "## Overview\n",
        "\n",
        "This notebook demonstrates finetuning and deploying Llama 3 models with Vertex AI. All of the examples in this notebook use parameter efficient finetuning methods [PEFT (LoRA)](https://github.com/huggingface/peft) to reduce training and storage costs. LoRA (Low-Rank Adaptation) is one approach of Parameter Efficient FineTuning (PEFT), where pretrained model weights are frozen and rank decomposition matrices representing the change in model weights are trained during finetuning. Read more about LoRA in the following publication: [Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L. and Chen, W., 2021. Lora: Low-rank adaptation of large language models. *arXiv preprint arXiv:2106.09685*](https://arxiv.org/abs/2106.09685).\n",
        "\n",
        "After finetuning, we can deploy models on Vertex with GPU.\n",
        "\n",
        "\n",
        "### Objective\n",
        "\n",
        "- Finetune Llama 3 models with Vertex AI Custom Training Jobs.\n",
        "- Deploy finetuned Llama 3 models on Vertex AI Prediction.\n",
        "- Send prediction requests to your finetuned Llama 3 models.\n",
        "\n",
        "\n",
        "### Costs\n",
        "\n",
        "This tutorial uses billable components of Google Cloud:\n",
        "\n",
        "* Vertex AI\n",
        "* Cloud Storage\n",
        "\n",
        "Learn about [Vertex AI pricing](https://cloud.google.com/vertex-ai/pricing), [Cloud Storage pricing](https://cloud.google.com/storage/pricing), and use the [Pricing Calculator](https://cloud.google.com/products/calculator/) to generate a cost estimate based on your projected usage."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "264c07757582"
      },
      "source": [
        "## Before you begin"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "cellView": "form",
        "id": "855d6b96f291"
      },
      "outputs": [],
      "source": [
        "# @title Setup Google Cloud project\n",
        "\n",
        "# @markdown 1. [Make sure that billing is enabled for your project](https://cloud.google.com/billing/docs/how-to/modify-project).\n",
        "\n",
        "# @markdown 2. **[Optional]** [Create a Cloud Storage bucket](https://cloud.google.com/storage/docs/creating-buckets) for storing experiment outputs. Set the BUCKET_URI for the experiment environment. The specified Cloud Storage bucket (`BUCKET_URI`) should be located in the same region as where the notebook was launched. Note that a multi-region bucket (eg. \"us\") is not considered a match for a single region covered by the multi-region range (eg. \"us-central1\"). If not set, a unique GCS bucket will be created instead.\n",
        "\n",
        "BUCKET_URI = \"gs://\"  # @param {type:\"string\"}\n",
        "\n",
        "# @markdown 3. **[Optional]** Set region. If not set, the region will be set automatically according to Colab Enterprise environment.\n",
        "\n",
        "REGION = \"\"  # @param {type:\"string\"}\n",
        "\n",
        "# @markdown 4. If you want to run predictions with A100 80GB or H100 GPUs, we recommend using the regions listed below. **NOTE:** Make sure you have associated quota in selected regions. Click the links to see your current quota for each GPU type: [Nvidia A100 80GB](https://console.cloud.google.com/iam-admin/quotas?metric=aiplatform.googleapis.com%2Fcustom_model_serving_nvidia_a100_80gb_gpus), [Nvidia H100 80GB](https://console.cloud.google.com/iam-admin/quotas?metric=aiplatform.googleapis.com%2Fcustom_model_serving_nvidia_h100_gpus).\n",
        "\n",
        "# @markdown > | Machine Type | Accelerator Type | Recommended Regions |\n",
        "# @markdown | ----------- | ----------- | ----------- |\n",
        "# @markdown | a2-ultragpu-1g | 1 NVIDIA_A100_80GB | us-central1, us-east4, europe-west4, asia-southeast1, us-east4 |\n",
        "# @markdown | a3-highgpu-8g | 8 NVIDIA_H100_80GB | us-central1, us-west1, europe-west4, asia-southeast1 |\n",
        "\n",
        "# @markdown 5. [Make sure that you have GPU quota for Vertex Training (finetuing) and Vertex Prediction (serving)](https://cloud.google.com/docs/quotas/view-manage). The quota name for Vertex Training is \"Custom model training your-gpu-type per region\" and the quota name for Vertex Prediction is \"Custom model serving your-gpu-type per region\" such as `Custom model training Nvidia L4 GPUs per region` and `Custom model serving Nvidia L4 GPUs per region` for L4 GPUs. [Submit a quota increase request](https://cloud.google.com/docs/quotas/view-manage#requesting_higher_quota) if additional quota is needed. At minimum, running this notebook requires 4 L4s for finetuning and 1 L4 for serving. More GPUs may be needed for larger models and different finetuning configurations. To secure GPUs for larger models, ask your customer engineer to get you allowlisted for a Shared Reservation or a Dynamic Workload Scheduler.\n",
        "\n",
        "# Import the necessary packages\n",
        "# Upgrade Vertex AI SDK.\n",
        "! pip3 install --upgrade --quiet 'google-cloud-aiplatform>=1.64.0'\n",
        "! git clone https://github.com/GoogleCloudPlatform/vertex-ai-samples.git\n",
        "\n",
        "import datetime\n",
        "import importlib\n",
        "import os\n",
        "import uuid\n",
        "from typing import Tuple\n",
        "\n",
        "from google.cloud import aiplatform\n",
        "\n",
        "common_util = importlib.import_module(\n",
        "    \"vertex-ai-samples.community-content.vertex_model_garden.model_oss.notebook_util.common_util\"\n",
        ")\n",
        "\n",
        "models, endpoints = {}, {}\n",
        "# Dedicated endpoint not supported yet\n",
        "use_dedicated_endpoint = False\n",
        "\n",
        "# Get the default cloud project id.\n",
        "PROJECT_ID = os.environ[\"GOOGLE_CLOUD_PROJECT\"]\n",
        "\n",
        "# Get the default region for launching jobs.\n",
        "if not REGION:\n",
        "    REGION = os.environ[\"GOOGLE_CLOUD_REGION\"]\n",
        "\n",
        "# Enable the Vertex AI API and Compute Engine API, if not already.\n",
        "print(\"Enabling Vertex AI API and Compute Engine API.\")\n",
        "! gcloud services enable aiplatform.googleapis.com compute.googleapis.com\n",
        "\n",
        "# Cloud Storage bucket for storing the experiment artifacts.\n",
        "# A unique GCS bucket will be created for the purpose of this notebook. If you\n",
        "# prefer using your own GCS bucket, change the value yourself below.\n",
        "now = datetime.datetime.now().strftime(\"%Y%m%d%H%M%S\")\n",
        "BUCKET_NAME = \"/\".join(BUCKET_URI.split(\"/\")[:3])\n",
        "\n",
        "if BUCKET_URI is None or BUCKET_URI.strip() == \"\" or BUCKET_URI == \"gs://\":\n",
        "    BUCKET_URI = f\"gs://{PROJECT_ID}-tmp-{now}-{str(uuid.uuid4())[:4]}\"\n",
        "    BUCKET_NAME = \"/\".join(BUCKET_URI.split(\"/\")[:3])\n",
        "    ! gsutil mb -l {REGION} {BUCKET_URI}\n",
        "else:\n",
        "    assert BUCKET_URI.startswith(\"gs://\"), \"BUCKET_URI must start with `gs://`.\"\n",
        "    shell_output = ! gsutil ls -Lb {BUCKET_NAME} | grep \"Location constraint:\" | sed \"s/Location constraint://\"\n",
        "    bucket_region = shell_output[0].strip().lower()\n",
        "    if bucket_region != REGION:\n",
        "        raise ValueError(\n",
        "            \"Bucket region %s is different from notebook region %s\"\n",
        "            % (bucket_region, REGION)\n",
        "        )\n",
        "print(f\"Using this GCS Bucket: {BUCKET_URI}\")\n",
        "\n",
        "STAGING_BUCKET = os.path.join(BUCKET_URI, \"temporal\")\n",
        "MODEL_BUCKET = os.path.join(BUCKET_URI, \"llama3\")\n",
        "\n",
        "\n",
        "# Initialize Vertex AI API.\n",
        "print(\"Initializing Vertex AI API.\")\n",
        "aiplatform.init(project=PROJECT_ID, location=REGION, staging_bucket=STAGING_BUCKET)\n",
        "\n",
        "# Gets the default SERVICE_ACCOUNT.\n",
        "shell_output = ! gcloud projects describe $PROJECT_ID\n",
        "project_number = shell_output[-1].split(\":\")[1].strip().replace(\"'\", \"\")\n",
        "SERVICE_ACCOUNT = f\"{project_number}-compute@developer.gserviceaccount.com\"\n",
        "print(\"Using this default Service Account:\", SERVICE_ACCOUNT)\n",
        "\n",
        "\n",
        "# Provision permissions to the SERVICE_ACCOUNT with the GCS bucket\n",
        "! gsutil iam ch serviceAccount:{SERVICE_ACCOUNT}:roles/storage.admin $BUCKET_NAME\n",
        "\n",
        "! gcloud config set project $PROJECT_ID\n",
        "! gcloud projects add-iam-policy-binding --no-user-output-enabled {PROJECT_ID} --member=serviceAccount:{SERVICE_ACCOUNT} --role=\"roles/storage.admin\"\n",
        "! gcloud projects add-iam-policy-binding --no-user-output-enabled {PROJECT_ID} --member=serviceAccount:{SERVICE_ACCOUNT} --role=\"roles/aiplatform.user\""
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "cellView": "form",
        "id": "36c21f10355f"
      },
      "outputs": [],
      "source": [
        "# @title Access Llama 3 models\n",
        "\n",
        "# @markdown For GPU based finetuning and serving, choose between accessing Llama 3 models on [Hugging Face](https://huggingface.co/)\n",
        "# @markdown or Vertex AI as described below.\n",
        "\n",
        "# @markdown If you already obtained access to Llama 3 models on [Hugging Face](https://huggingface.co/), you can load models from there.\n",
        "# @markdown Alternatively, you can also load the original Llama 3 models for finetuning and serving from Vertex AI after accepting the agreement.\n",
        "\n",
        "# @markdown **Only select and fill one of the following sections.**\n",
        "# fmt: off\n",
        "LOAD_MODEL_FROM = \"Hugging Face\"  # @param [\"Hugging Face\", \"Google Cloud\"] {isTemplate:true}\n",
        "# fmt: on\n",
        "\n",
        "# @markdown ---\n",
        "\n",
        "# @markdown ### Access Llama 3 models on Hugging Face for GPU based finetuning and serving\n",
        "# @markdown You must provide a Hugging Face User Access Token (read) to access the Llama 3 models. You can follow the [Hugging Face documentation](https://huggingface.co/docs/hub/en/security-tokens) to create a **read** access token and put it in the `HF_TOKEN` field below.\n",
        "\n",
        "HF_TOKEN = \"\"  # @param {type:\"string\", isTemplate:true}\n",
        "if LOAD_MODEL_FROM == \"Hugging Face\":\n",
        "    assert (\n",
        "        HF_TOKEN\n",
        "    ), \"Provide a read HF_TOKEN to load models from Hugging Face, or select a different model source.\"\n",
        "\n",
        "# @markdown *--- Or ---*\n",
        "# @markdown ### Access Llama 3 models on Vertex AI for GPU based serving\n",
        "# @markdown The original models from Meta are converted into the Hugging Face format for serving in Vertex AI.\n",
        "# @markdown Accept the model agreement to access the models:\n",
        "# @markdown 1. Open the [Llama 3 model card](https://console.cloud.google.com/vertex-ai/publishers/meta/model-garden/llama3) from [Vertex AI Model Garden](https://cloud.google.com/model-garden).\n",
        "# @markdown 2. Review and accept the agreement in the pop-up window on the model card page. If you have previously accepted the model agreement, there will not be a pop-up window on the model card page and this step is not needed.\n",
        "# @markdown 3. After accepting the agreement of Llama 3, a `gs://` URI containing Llama 3 pretrained and finetuned models will be shared.\n",
        "# @markdown 4. Paste the URI in the `VERTEX_AI_MODEL_GARDEN_LLAMA3` field below.\n",
        "\n",
        "VERTEX_AI_MODEL_GARDEN_LLAMA3 = \"\"  # @param {type:\"string\", isTemplate:true}\n",
        "\n",
        "if LOAD_MODEL_FROM == \"Google Cloud\":\n",
        "    assert (\n",
        "        VERTEX_AI_MODEL_GARDEN_LLAMA3\n",
        "    ), \"Click the agreement of Llama 3 in Vertex AI Model Garden, and get the GCS path of Llama 3 model artifacts.\"\n",
        "    print(\n",
        "        \"Copying Llama 3 model artifacts from\",\n",
        "        VERTEX_AI_MODEL_GARDEN_LLAMA3,\n",
        "        \"to \",\n",
        "        MODEL_BUCKET,\n",
        "    )\n",
        "    HF_TOKEN = \"\"\n",
        "\n",
        "    ! gsutil -m cp -R $VERTEX_AI_MODEL_GARDEN_LLAMA3/* $MODEL_BUCKET\n",
        "\n",
        "# @markdown ---"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "cb56d402e84a"
      },
      "source": [
        "## Finetune with HuggingFace PEFT and deploy with vLLM on GPUs"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "cellView": "form",
        "id": "KwAW99YZHTdy"
      },
      "outputs": [],
      "source": [
        "# @title Set dataset\n",
        "\n",
        "# @markdown Use the Vertex AI SDK to create and run the custom training jobs.\n",
        "\n",
        "# @markdown This notebook uses [timdettmers/openassistant-guanaco](https://huggingface.co/datasets/timdettmers/openassistant-guanaco) dataset as an example.\n",
        "# @markdown You can set `dataset_name` to any existing [Hugging Face dataset](https://huggingface.co/datasets) name, and set `instruct_column_in_dataset` to the name of the dataset column containing training data. The [timdettmers/openassistant-guanaco](https://huggingface.co/datasets/timdettmers/openassistant-guanaco) has only one column `text`, and therefore we set `instruct_column_in_dataset` to `text` in this notebook.\n",
        "\n",
        "# @markdown ### (Optional) Prepare a custom JSONL dataset for finetuning\n",
        "\n",
        "# @markdown You can prepare a JSONL file where each line is a valid JSON string as your custom training dataset. For example, here is one line from the [timdettmers/openassistant-guanaco](https://huggingface.co/datasets/timdettmers/openassistant-guanaco) dataset:\n",
        "# @markdown ```\n",
        "# @markdown {\"text\": \"### Human: Hola### Assistant: \\u00a1Hola! \\u00bfEn qu\\u00e9 puedo ayudarte hoy?\"}\n",
        "# @markdown ```\n",
        "\n",
        "# @markdown The JSON object has a key `text`, which should match `instruct_column_in_dataset`; The value should be one training data point, i.e. a string. After you prepared your JSONL file, you can either upload it to [Hugging Face datasets](https://huggingface.co/datasets) or [Google Cloud Storage](https://cloud.google.com/storage).\n",
        "\n",
        "# @markdown - To upload a JSONL dataset to [Hugging Face datasets](https://huggingface.co/datasets), follow the instructions on [Uploading Datasets](https://huggingface.co/docs/hub/en/datasets-adding). Then, set `dataset_name` to the name of your newly created dataset on Hugging Face.\n",
        "\n",
        "# @markdown - To upload a JSONL dataset to [Google Cloud Storage](https://cloud.google.com/storage), follow the instructions on [Upload objects from a filesystem](https://cloud.google.com/storage/docs/uploading-objects). Then, set `dataset_name` to the `gs://` URI to your JSONL file. For example: `gs://cloud-samples-data/vertex-ai/model-evaluation/peft_train_sample.jsonl`.\n",
        "\n",
        "# @markdown Optionally update the `instruct_column_in_dataset` field below if your JSON objects use a key other than the default `text`.\n",
        "\n",
        "# @markdown ### (Optional) Format your data with custom JSON template\n",
        "\n",
        "# @markdown Sometimes, your dataset might have multiple text columns and you want to construct the training data with a template. You can prepare a JSON template in the following format:\n",
        "\n",
        "# @markdown ```\n",
        "# @markdown {\n",
        "# @markdown   \"description\": \"Template used by Llama 3, accepting text-bison format.\",\n",
        "# @markdown   \"source\": \"https://cloud.google.com/vertex-ai/generative-ai/docs/models/tune-text-models-supervised#dataset-format\",\n",
        "# @markdown   \"prompt_input\": \"<|start_header_id|>user<|end_header_id|>\\n\\n{input_text}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\\n\\n{output_text}<|eot_id|>\",\n",
        "# @markdown   \"instruction_separator\": \"<|start_header_id|>user<|end_header_id|>\\n\\n\",\n",
        "# @markdown   \"response_separator\": \"<|start_header_id|>assistant<|end_header_id|>\\n\\n\"\n",
        "# @markdown }\n",
        "# @markdown ```\n",
        "\n",
        "# @markdown As an example, the template above can be used to format the following training data (this line comes from `gs://cloud-samples-data/vertex-ai/model-evaluation/peft_train_sample.jsonl`):\n",
        "\n",
        "# @markdown ```\n",
        "# @markdown {\"input_text\":\"TRANSCRIPT: \\nREASON FOR EVALUATION:,\\n\\n LABEL:\",\"output_text\":\"Chiropractic\"}\n",
        "# @markdown ```\n",
        "\n",
        "# @markdown This example template simply concatenates `input_text` with `output_text` with some special tokens in between.\n",
        "# @markdown\n",
        "# @markdown To try such custom dataset, you can make the following changes:\n",
        "# @markdown 1. Set `template` to `llama3-text-bison`\n",
        "# @markdown 1. Set `train_dataset_name` to `gs://cloud-samples-data/vertex-ai/model-evaluation/peft_train_sample.jsonl`\n",
        "# @markdown 1. Set `train_split_name` to `train`\n",
        "# @markdown 1. Set `eval_dataset_name` to `gs://cloud-samples-data/vertex-ai/model-evaluation/peft_eval_sample.jsonl`\n",
        "# @markdown 1. Set `eval_split_name` to `train` (**NOT** `test`)\n",
        "# @markdown 1. Set `instruct_column_in_dataset` as `input_text`.\n",
        "\n",
        "# Template name or gs:// URI to a custom template.\n",
        "template = \"openassistant-guanaco\"  # @param {type:\"string\"}\n",
        "\n",
        "# Hugging Face dataset name or gs:// URI to a custom JSONL dataset.\n",
        "train_dataset_name = \"timdettmers/openassistant-guanaco\"  # @param {type:\"string\"}\n",
        "train_split_name = \"train\"  # @param {type:\"string\"}\n",
        "eval_dataset_name = \"timdettmers/openassistant-guanaco\"  # @param {type:\"string\"}\n",
        "eval_split_name = \"test\"  # @param {type:\"string\"}\n",
        "\n",
        "# Name of the dataset column containing training text input.\n",
        "instruct_column_in_dataset = \"text\"  # @param {type:\"string\"}"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "cellView": "form",
        "id": "ivVGS9dHXPOz"
      },
      "outputs": [],
      "source": [
        "# @title Finetune\n",
        "# @markdown Use the Vertex AI SDK to create and run the custom training jobs.\n",
        "\n",
        "# @markdown **Note**:\n",
        "# @markdown 1. We recommend setting `finetuning_precision_mode` to `4bit` because it enables using fewer hardware resources for finetuning.\n",
        "# @markdown 1. We recommend using NVIDIA_L4 for 8B models and NVIDIA_A100_80GB for 70B models.\n",
        "# @markdown 1. If `max_steps>0`, it will precedence over `epochs`. One can set a small `max_steps` value to quickly check the pipeline.\n",
        "# @markdown 1. With the default setting, training takes between 1.5 ~ 2 hours.\n",
        "\n",
        "TRAIN_DOCKER_URI = \"us-docker.pkg.dev/vertex-ai/vertex-vision-model-garden-dockers/pytorch-peft-train:20240625_0902_RC00\"\n",
        "\n",
        "\n",
        "# The Llama 3 base model.\n",
        "MODEL_ID = \"meta-llama/Meta-Llama-3-8B-Instruct\"  # @param [\"meta-llama/Meta-Llama-3-8B\", \"meta-llama/Meta-Llama-3-8B-Instruct\", \"meta-llama/Meta-Llama-3-70B\", \"meta-llama/Meta-Llama-3-70B-Instruct\"] {isTemplate:true}\n",
        "if LOAD_MODEL_FROM == \"Google Cloud\":\n",
        "    if MODEL_ID == \"meta-llama/Meta-Llama-3-8B\":\n",
        "        base_model_id = \"llama3-8b-hf\"\n",
        "    elif MODEL_ID == \"meta-llama/Meta-Llama-3-8B-Instruct\":\n",
        "        base_model_id = \"llama3-8b-chat-hf\"\n",
        "    elif MODEL_ID == \"meta-llama/Meta-Llama-3-70B\":\n",
        "        base_model_id = \"llama3-70b-hf\"\n",
        "    elif MODEL_ID == \"meta-llama/Meta-Llama-3-70B-Instruct\":\n",
        "        base_model_id = \"llama3-70b-chat-hf\"\n",
        "    else:\n",
        "        raise ValueError(f\"Undefined model ID: {MODEL_ID}.\")\n",
        "    base_model_id = os.path.join(MODEL_BUCKET, base_model_id)\n",
        "else:\n",
        "    base_model_id = MODEL_ID\n",
        "\n",
        "# The accelerator to use.\n",
        "accelerator_type = \"NVIDIA_L4\"  # @param [\"NVIDIA_L4\", \"NVIDIA_A100_80GB\"]\n",
        "\n",
        "# Batch size for finetuning.\n",
        "per_device_train_batch_size = 1  # @param{type:\"integer\"}\n",
        "gradient_accumulation_steps = 8  # @param{type:\"integer\"}\n",
        "# Maximum sequence length.\n",
        "max_seq_length = 4096  # @param{type:\"integer\"}\n",
        "# Setting a positive `max_steps` here will override `num_epochs`\n",
        "max_steps = -1  # @param{type:\"integer\"}\n",
        "num_epochs = 1.0  # @param{type:\"number\"}\n",
        "# Precision mode for finetuning.\n",
        "finetuning_precision_mode = \"4bit\"  # @param [\"4bit\", \"8bit\", \"float16\"]\n",
        "# Learning rate.\n",
        "learning_rate = 5e-5  # @param{type:\"number\"}\n",
        "lr_scheduler_type = \"cosine\"  # @param{type:\"string\"}\n",
        "# LoRA parameters.\n",
        "lora_rank = 16  # @param{type:\"integer\"}\n",
        "lora_alpha = 32  # @param{type:\"integer\"}\n",
        "lora_dropout = 0.05  # @param{type:\"number\"}\n",
        "enable_gradient_checkpointing = True\n",
        "attn_implementation = \"flash_attention_2\"\n",
        "optimizer = \"paged_adamw_32bit\"\n",
        "warmup_ratio = \"0.01\"\n",
        "report_to = \"tensorboard\"\n",
        "save_steps = 10\n",
        "logging_steps = save_steps\n",
        "\n",
        "# Worker pool spec.\n",
        "machine_type = None\n",
        "if \"8b\" in MODEL_ID.lower():\n",
        "    if accelerator_type == \"NVIDIA_L4\":\n",
        "        accelerator_count = 4\n",
        "        machine_type = \"g2-standard-48\"\n",
        "    else:\n",
        "        raise ValueError(\n",
        "            f\"Recommended machine settings not found for: {accelerator_type}. To use another accelerator, edit this code block to pass in an appropriate `machine_type`, `accelerator_type`, and `accelerator_count` to the deploy_model_vllm function by clicking `Show Code` and then modifying the code.\"\n",
        "        )\n",
        "elif \"70b\" in MODEL_ID.lower():\n",
        "    if accelerator_type == \"NVIDIA_A100_80GB\":\n",
        "        accelerator_count = 4\n",
        "        machine_type = \"a2-ultragpu-4g\"\n",
        "    else:\n",
        "        raise ValueError(\n",
        "            f\"Recommended machine settings not found for: {accelerator_type}. To use another accelerator, edit this code block to pass in an appropriate `machine_type`, `accelerator_type`, and `accelerator_count` to the deploy_model_vllm function by clicking `Show Code` and then modifying the code.\"\n",
        "        )\n",
        "else:\n",
        "    raise ValueError(f\"Unsupported model ID or GCS path: {MODEL_ID}.\")\n",
        "\n",
        "replica_count = 1\n",
        "\n",
        "common_util.check_quota(\n",
        "    project_id=PROJECT_ID,\n",
        "    region=REGION,\n",
        "    accelerator_type=accelerator_type,\n",
        "    accelerator_count=accelerator_count,\n",
        "    is_for_training=True,\n",
        ")\n",
        "\n",
        "job_name = common_util.get_job_name_with_datetime(\"llama3-lora-train\")\n",
        "\n",
        "base_output_dir = os.path.join(STAGING_BUCKET, job_name)\n",
        "# Create a GCS folder to store the LORA adapter.\n",
        "lora_output_dir = os.path.join(base_output_dir, \"adapter\")\n",
        "# Create a GCS folder to store the merged model with the base model and the\n",
        "# finetuned LORA adapter.\n",
        "merged_model_output_dir = os.path.join(base_output_dir, \"merged-model\")\n",
        "\n",
        "eval_args = [\n",
        "    f\"--eval_dataset_path={eval_dataset_name}\",\n",
        "    f\"--eval_column={instruct_column_in_dataset}\",\n",
        "    f\"--eval_template={template}\",\n",
        "    f\"--eval_split={eval_split_name}\",\n",
        "    f\"--eval_steps={save_steps}\",\n",
        "    \"--eval_tasks=builtin_eval\",\n",
        "    \"--eval_metric_name=loss\",\n",
        "]\n",
        "\n",
        "train_job_args = [\n",
        "    \"--config_file=vertex_vision_model_garden_peft/deepspeed_zero2_4gpu.yaml\",\n",
        "    \"--task=instruct-lora\",\n",
        "    \"--completion_only=True\",\n",
        "    f\"--pretrained_model_id={base_model_id}\",\n",
        "    f\"--dataset_name={train_dataset_name}\",\n",
        "    f\"--train_split_name={train_split_name}\",\n",
        "    f\"--instruct_column_in_dataset={instruct_column_in_dataset}\",\n",
        "    f\"--output_dir={lora_output_dir}\",\n",
        "    f\"--merge_base_and_lora_output_dir={merged_model_output_dir}\",\n",
        "    f\"--per_device_train_batch_size={per_device_train_batch_size}\",\n",
        "    f\"--gradient_accumulation_steps={gradient_accumulation_steps}\",\n",
        "    f\"--lora_rank={lora_rank}\",\n",
        "    f\"--lora_alpha={lora_alpha}\",\n",
        "    f\"--lora_dropout={lora_dropout}\",\n",
        "    f\"--max_steps={max_steps}\",\n",
        "    f\"--max_seq_length={max_seq_length}\",\n",
        "    f\"--learning_rate={learning_rate}\",\n",
        "    f\"--lr_scheduler_type={lr_scheduler_type}\",\n",
        "    f\"--precision_mode={finetuning_precision_mode}\",\n",
        "    f\"--enable_gradient_checkpointing={enable_gradient_checkpointing}\",\n",
        "    f\"--num_epochs={num_epochs}\",\n",
        "    f\"--attn_implementation={attn_implementation}\",\n",
        "    f\"--optimizer={optimizer}\",\n",
        "    f\"--warmup_ratio={warmup_ratio}\",\n",
        "    f\"--report_to={report_to}\",\n",
        "    f\"--logging_output_dir={base_output_dir}\",\n",
        "    f\"--save_steps={save_steps}\",\n",
        "    f\"--logging_steps={logging_steps}\",\n",
        "    f\"--template={template}\",\n",
        "    f\"--huggingface_access_token={HF_TOKEN}\",\n",
        "] + eval_args\n",
        "\n",
        "# Create TensorBoard\n",
        "tensorboard = aiplatform.Tensorboard.create(job_name)\n",
        "exp = aiplatform.TensorboardExperiment.create(\n",
        "    tensorboard_experiment_id=job_name, tensorboard_name=tensorboard.name\n",
        ")\n",
        "\n",
        "# Pass training arguments and launch job.\n",
        "train_job = aiplatform.CustomContainerTrainingJob(\n",
        "    display_name=job_name,\n",
        "    container_uri=TRAIN_DOCKER_URI,\n",
        ")\n",
        "\n",
        "train_job.run(\n",
        "    args=train_job_args,\n",
        "    environment_variables={\"WANDB_DISABLED\": True},\n",
        "    replica_count=replica_count,\n",
        "    machine_type=machine_type,\n",
        "    accelerator_type=accelerator_type,\n",
        "    accelerator_count=accelerator_count,\n",
        "    boot_disk_size_gb=500,\n",
        "    service_account=SERVICE_ACCOUNT,\n",
        "    tensorboard=tensorboard.resource_name,\n",
        "    base_output_dir=base_output_dir,\n",
        ")\n",
        "\n",
        "print(\"LoRA adapter was saved in: \", lora_output_dir)\n",
        "print(\"Trained and merged models were saved in: \", merged_model_output_dir)\n",
        "\n",
        "# @markdown Click \"Show Code\" to see more details."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "cellView": "form",
        "id": "qmHW6m8xG_4U"
      },
      "outputs": [],
      "source": [
        "# @title Deploy\n",
        "# @markdown This section uploads the model to Model Registry and deploys it on the Endpoint. It takes 15 minutes to 1 hour to finish.\n",
        "\n",
        "print(\"Deploying models in: \", merged_model_output_dir)\n",
        "\n",
        "# The pre-built serving docker image for vLLM.\n",
        "VLLM_DOCKER_URI = \"us-docker.pkg.dev/vertex-ai/vertex-vision-model-garden-dockers/pytorch-vllm-serve:20240721_0916_RC00\"\n",
        "\n",
        "# Find Vertex AI prediction supported accelerators and regions in [here](https://cloud.google.com/vertex-ai/docs/predictions/configure-compute).\n",
        "if \"8b\" in MODEL_ID.lower():\n",
        "    machine_type = \"g2-standard-12\"\n",
        "    accelerator_type = \"NVIDIA_L4\"\n",
        "    accelerator_count = 1\n",
        "else:\n",
        "    machine_type = \"g2-standard-96\"\n",
        "    accelerator_type = \"NVIDIA_L4\"\n",
        "    accelerator_count = 8\n",
        "\n",
        "common_util.check_quota(\n",
        "    project_id=PROJECT_ID,\n",
        "    region=REGION,\n",
        "    accelerator_type=accelerator_type,\n",
        "    accelerator_count=accelerator_count,\n",
        "    is_for_training=False,\n",
        ")\n",
        "\n",
        "gpu_memory_utilization = 0.85\n",
        "max_model_len = 8192  # Maximum context length.\n",
        "\n",
        "# Ensure max_model_len does not exceed the limit\n",
        "if max_model_len > 8192:\n",
        "    raise ValueError(\"max_model_len cannot exceed 8192\")\n",
        "\n",
        "\n",
        "def deploy_model_vllm(\n",
        "    model_name: str,\n",
        "    model_id: str,\n",
        "    service_account: str,\n",
        "    base_model_id: str = None,\n",
        "    machine_type: str = \"g2-standard-8\",\n",
        "    accelerator_type: str = \"NVIDIA_L4\",\n",
        "    accelerator_count: int = 1,\n",
        "    gpu_memory_utilization: float = 0.9,\n",
        "    max_model_len: int = 4096,\n",
        "    dtype: str = \"auto\",\n",
        "    enable_trust_remote_code: bool = False,\n",
        "    enforce_eager: bool = False,\n",
        "    enable_lora: bool = False,\n",
        "    max_loras: int = 1,\n",
        "    max_cpu_loras: int = 8,\n",
        "    use_dedicated_endpoint: bool = False,\n",
        "    max_num_seqs: int = 256,\n",
        "    model_type: str = None,\n",
        ") -> Tuple[aiplatform.Model, aiplatform.Endpoint]:\n",
        "    \"\"\"Deploys trained models with vLLM into Vertex AI.\"\"\"\n",
        "    endpoint = aiplatform.Endpoint.create(\n",
        "        display_name=f\"{model_name}-endpoint\",\n",
        "        dedicated_endpoint_enabled=use_dedicated_endpoint,\n",
        "    )\n",
        "\n",
        "    if not base_model_id:\n",
        "        base_model_id = model_id\n",
        "\n",
        "    # See https://docs.vllm.ai/en/latest/models/engine_args.html for a list of possible arguments with descriptions.\n",
        "    vllm_args = [\n",
        "        \"python\",\n",
        "        \"-m\",\n",
        "        \"vllm.entrypoints.api_server\",\n",
        "        \"--host=0.0.0.0\",\n",
        "        \"--port=8080\",\n",
        "        f\"--model={model_id}\",\n",
        "        f\"--tensor-parallel-size={accelerator_count}\",\n",
        "        \"--swap-space=16\",\n",
        "        f\"--gpu-memory-utilization={gpu_memory_utilization}\",\n",
        "        f\"--max-model-len={max_model_len}\",\n",
        "        f\"--dtype={dtype}\",\n",
        "        f\"--max-loras={max_loras}\",\n",
        "        f\"--max-cpu-loras={max_cpu_loras}\",\n",
        "        f\"--max-num-seqs={max_num_seqs}\",\n",
        "        \"--disable-log-stats\",\n",
        "    ]\n",
        "\n",
        "    if enable_trust_remote_code:\n",
        "        vllm_args.append(\"--trust-remote-code\")\n",
        "\n",
        "    if enforce_eager:\n",
        "        vllm_args.append(\"--enforce-eager\")\n",
        "\n",
        "    if enable_lora:\n",
        "        vllm_args.append(\"--enable-lora\")\n",
        "\n",
        "    if model_type:\n",
        "        vllm_args.append(f\"--model-type={model_type}\")\n",
        "\n",
        "    env_vars = {\n",
        "        \"MODEL_ID\": base_model_id,\n",
        "        \"DEPLOY_SOURCE\": \"notebook\",\n",
        "    }\n",
        "\n",
        "    # HF_TOKEN is not a compulsory field and may not be defined.\n",
        "    try:\n",
        "        if HF_TOKEN:\n",
        "            env_vars[\"HF_TOKEN\"] = HF_TOKEN\n",
        "    except NameError:\n",
        "        pass\n",
        "\n",
        "    model = aiplatform.Model.upload(\n",
        "        display_name=model_name,\n",
        "        serving_container_image_uri=VLLM_DOCKER_URI,\n",
        "        serving_container_args=vllm_args,\n",
        "        serving_container_ports=[8080],\n",
        "        serving_container_predict_route=\"/generate\",\n",
        "        serving_container_health_route=\"/ping\",\n",
        "        serving_container_environment_variables=env_vars,\n",
        "        serving_container_shared_memory_size_mb=(16 * 1024),  # 16 GB\n",
        "        serving_container_deployment_timeout=7200,\n",
        "    )\n",
        "    print(\n",
        "        f\"Deploying {model_name} on {machine_type} with {accelerator_count} {accelerator_type} GPU(s).\"\n",
        "    )\n",
        "    model.deploy(\n",
        "        endpoint=endpoint,\n",
        "        machine_type=machine_type,\n",
        "        accelerator_type=accelerator_type,\n",
        "        accelerator_count=accelerator_count,\n",
        "        deploy_request_timeout=1800,\n",
        "        service_account=service_account,\n",
        "    )\n",
        "    print(\"endpoint_name:\", endpoint.name)\n",
        "\n",
        "    return model, endpoint\n",
        "\n",
        "\n",
        "models[\"vllm_gpu\"], endpoints[\"vllm_gpu\"] = deploy_model_vllm(\n",
        "    model_name=common_util.get_job_name_with_datetime(prefix=\"llama3-vllm-serve\"),\n",
        "    model_id=merged_model_output_dir,\n",
        "    service_account=SERVICE_ACCOUNT,\n",
        "    machine_type=machine_type,\n",
        "    accelerator_type=accelerator_type,\n",
        "    accelerator_count=accelerator_count,\n",
        "    gpu_memory_utilization=gpu_memory_utilization,\n",
        "    max_model_len=max_model_len,\n",
        "    use_dedicated_endpoint=use_dedicated_endpoint,\n",
        ")\n",
        "\n",
        "# @markdown Click \"Show Code\" to see more details."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "cellView": "form",
        "id": "2UYUNn60G_4U"
      },
      "outputs": [],
      "source": [
        "# @title Predict\n",
        "\n",
        "# @markdown Once deployment succeeds, you can send requests to the endpoint with text prompts. Sampling parameters supported by vLLM can be found [here](https://docs.vllm.ai/en/latest/dev/sampling_params.html).\n",
        "\n",
        "# @markdown Example:\n",
        "\n",
        "# @markdown ```\n",
        "# @markdown Human: What is a car?\n",
        "# @markdown Assistant:  A car, or a motor car, is a road-connected human-transportation system used to move people or goods from one place to another. The term also encompasses a wide range of vehicles, including motorboats, trains, and aircrafts. Cars typically have four wheels, a cabin for passengers, and an engine or motor. They have been around since the early 19th century and are now one of the most popular forms of transportation, used for daily commuting, shopping, and other purposes.\n",
        "# @markdown ```\n",
        "# @markdown Additionally, you can moderate the generated text with Vertex AI. See [Moderate text documentation](https://cloud.google.com/natural-language/docs/moderating-text) for more details.\n",
        "\n",
        "# Loads an existing endpoint instance using the endpoint name:\n",
        "# - Using `endpoint_name = endpoint.name` allows us to get the\n",
        "#   endpoint name of the endpoint `endpoint` created in the cell\n",
        "#   above.\n",
        "# - Alternatively, you can set `endpoint_name = \"1234567890123456789\"` to load\n",
        "#   an existing endpoint with the ID 1234567890123456789.\n",
        "# You may uncomment the code below to load an existing endpoint.\n",
        "\n",
        "# endpoint_name = \"\"  # @param {type:\"string\"}\n",
        "# aip_endpoint_name = (\n",
        "#     f\"projects/{PROJECT_ID}/locations/{REGION}/endpoints/{endpoint_name}\"\n",
        "# )\n",
        "# endpoint = aiplatform.Endpoint(aip_endpoint_name)\n",
        "\n",
        "prompt = \"What is a car?\"  # @param {type: \"string\"}\n",
        "# @markdown If you encounter an issue like `ServiceUnavailable: 503 Took too long to respond when processing`, you can reduce the maximum number of output tokens, by lowering `max_tokens`.\n",
        "max_tokens = 50  # @param {type:\"integer\"}\n",
        "temperature = 1.0  # @param {type:\"number\"}\n",
        "top_p = 1.0  # @param {type:\"number\"}\n",
        "top_k = 1  # @param {type:\"integer\"}\n",
        "# @markdown Set `raw_response` to `True` to obtain the raw model output. Set `raw_response` to `False` to apply additional formatting in the structure of `\"Prompt:\\n{prompt.strip()}\\nOutput:\\n{output}\"`.\n",
        "raw_response = False  # @param {type:\"boolean\"}\n",
        "\n",
        "# Overrides parameters for inferences.\n",
        "instances = [\n",
        "    {\n",
        "        \"prompt\": prompt,\n",
        "        \"max_tokens\": max_tokens,\n",
        "        \"temperature\": temperature,\n",
        "        \"top_p\": top_p,\n",
        "        \"top_k\": top_k,\n",
        "        \"raw_response\": raw_response,\n",
        "    },\n",
        "]\n",
        "response = endpoints[\"vllm_gpu\"].predict(\n",
        "    instances=instances, use_dedicated_endpoint=use_dedicated_endpoint\n",
        ")\n",
        "\n",
        "for prediction in response.predictions:\n",
        "    print(prediction)\n",
        "\n",
        "# @markdown Click \"Show Code\" to see more details."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "af21a3cff1e0"
      },
      "source": [
        "## Clean up resources"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "cellView": "form",
        "id": "911406c1561e"
      },
      "outputs": [],
      "source": [
        "# @title Delete the model and endpoint\n",
        "\n",
        "train_job.delete()\n",
        "\n",
        "# @markdown  Delete the experiment models and endpoints to recycle the resources\n",
        "# @markdown  and avoid unnecessary continuous charges that may incur.\n",
        "\n",
        "# Undeploy model and delete endpoint.\n",
        "for endpoint in endpoints.values():\n",
        "    endpoint.delete(force=True)\n",
        "\n",
        "# Delete models.\n",
        "for model in models.values():\n",
        "    model.delete()\n",
        "\n",
        "delete_bucket = False  # @param {type:\"boolean\"}\n",
        "if delete_bucket:\n",
        "    ! gsutil -m rm -r $BUCKET_NAME"
      ]
    }
  ],
  "metadata": {
    "colab": {
      "name": "model_garden_pytorch_llama3_finetuning.ipynb",
      "toc_visible": true
    },
    "kernelspec": {
      "display_name": "Python 3",
      "name": "python3"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 0
}