llama2-finetuning-slurm/hpc-slurm-llama2.yaml

# Copyright 2024 Google LLC
#
# 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.

---

blueprint_name: hpc-slurm-llama

vars:
  project_id:  ## Set GCP Project ID Here ##
  bucket_model: ## Set your bucket name prefix here ##
  deployment_name: hpc-slurm-llama2
  region: us-central1
  zone: us-central1-a
  zone_list: [us-central1-a, us-central1-b, us-central1-c]
  new_image_family: llama2-slurm-v6
  instance_image_custom: true
  disk_size_gb: 200


# Documentation for each of the modules used below can be found at
# https://github.com/GoogleCloudPlatform/hpc-toolkit/blob/main/modules/README.md

deployment_groups:
- group: enable_apis
  modules:

  - id: enable_apis
    source: community/modules/project/service-enablement
    settings:
      gcp_service_list: [
        "cloudresourcemanager.googleapis.com",
        "stackdriver.googleapis.com",
        "iam.googleapis.com",
        "logging.googleapis.com",
        "compute.googleapis.com"
      ]
- group: setup
  modules:

  ## Monitoring
  - id: hpc_dash
    source: modules/monitoring/dashboard
    settings:
      title: HPC
  - id: gpu_dash
    source: modules/monitoring/dashboard
    settings:
      title: GPU
      base_dashboard: Empty
      widgets:
      - |
          {
            "title": "GPU Memory Utilization",
            "xyChart": {
              "dataSets": [
                {
                  "timeSeriesQuery": {
                    "timeSeriesFilter": {
                      "filter": "metric.type=\"agent.googleapis.com/gpu/memory/bytes_used\" resource.type=\"gce_instance\"",
                      "aggregation": {
                        "perSeriesAligner": "ALIGN_MEAN",
                        "crossSeriesReducer": "REDUCE_NONE",
                        "groupByFields": []
                      }
                    }
                  },
                  "plotType": "LINE",
                  "targetAxis": "Y1",
                  "minAlignmentPeriod": "60s"
                }
              ],
              "chartOptions": {
                "mode": "COLOR",
                "displayHorizontal": false
              },
              "thresholds": [],
              "yAxis": {
                "scale": "LINEAR"
              }
            }
          }
      - |
          {
            "title": "GPU Utilization",
            "xyChart": {
              "dataSets": [
                {
                  "timeSeriesQuery": {
                    "prometheusQuery": "avg_over_time(agent_googleapis_com:gpu_utilization{monitored_resource=\"gce_instance\"}[${__interval}])"
                  },
                  "plotType": "LINE",
                  "targetAxis": "Y1"
                }
              ],
              "chartOptions": {
                "mode": "COLOR",
                "displayHorizontal": false
              },
              "thresholds": [],
              "yAxis": {
                "scale": "LINEAR"
              }
            }
          }

  ## network
  - id: network1
    source: modules/network/vpc

  ## Filesystems
  - id: homefs
    source: community/modules/file-system/nfs-server
    use: [network1]
    settings:
      local_mounts: [/home]
      disk_size: 2560
      instance_image:  
        project: "cloud-hpc-image-public"
        family:  "hpc-rocky-linux-8"

  - id: data_bucket
    source: community/modules/file-system/cloud-storage-bucket
    settings:
      name_prefix: $(vars.bucket_model)
      random_suffix: true
      force_destroy: true
      local_mount: /data_bucket
      mount_options: defaults,_netdev,implicit_dirs,allow_other,dir_mode=0777,file_mode=766

  - id: move_files
    source: ./files
    use: [data_bucket]


  ## Install Scripts

  - id: packer_script
    # configure conda environment for llama
    source: modules/scripts/startup-script
    settings:
      runners:
      - type: shell
        destination: install-ml-libraries.sh
        content: |
          #!/bin/bash
          # this script is designed to execute on Slurm images published by SchedMD that:
          # - are based on Debian 11 distribution of Linux
          # - have NVIDIA Drivers v530 pre-installed
          # - have CUDA Toolkit 12.1 pre-installed.

          set -e -o pipefail


          CONDA_BASE=/opt/conda

          if [ -d $CONDA_BASE ]; then
                  exit 0
          fi

          DL_DIR=\$(mktemp -d)
          cd $DL_DIR
          curl -O https://repo.anaconda.com/miniconda/Miniconda3-py310_23.3.1-0-Linux-x86_64.sh
          HOME=$DL_DIR bash Miniconda3-py310_23.3.1-0-Linux-x86_64.sh -b -p $CONDA_BASE
          cd -
          rm -rf $DL_DIR
          unset DL_DIR

          tee /tmp/llama2_env.yml << EOLL
          name: llama2
          channels:
            - conda-forge
            - nvidia
            - nvidia/label/cuda-12.4.0
          dependencies:
            - appdirs
            - loralib
            - black
            - black-jupyter
            - py7zr
            - scipy
            - optimum
            - datasets
            - accelerate
            - peft
            - fairscale
            - fire
            - sentencepiece
            - transformers
            - huggingface_hub
            - git
            - pip
            - pip:
              - bitsandbytes
              - nvidia-cudnn-cu12
              - dataclasses
              - nvidia-nccl-cu12
              - trl
              - torch
              - torchaudio 
              - torchvision
              - nvitop
          EOLL

          source $CONDA_BASE/bin/activate base
          conda env create -n llama2 --file /tmp/llama2_env.yml

  - id: startup_script
    source: modules/scripts/startup-script
    settings:
      install_cloud_ops_agent: false
      runners:
      - type: shell
        destination: startup-script.sh
        content: |
          #!/bin/bash
          CONDA_BASE=/opt/conda
          source $CONDA_BASE/bin/activate base
          conda init --system

          # UnInstall Stackdriver Agent

          sudo systemctl stop stackdriver-agent.service
          sudo systemctl disable stackdriver-agent.service
          curl -sSO https://dl.google.com/cloudagents/add-monitoring-agent-repo.sh
          sudo dpkg --configure -a
          sudo bash add-monitoring-agent-repo.sh --uninstall
          sudo bash add-monitoring-agent-repo.sh --remove-repo

          # Install ops-agent

          curl -sSO https://dl.google.com/cloudagents/add-google-cloud-ops-agent-repo.sh
          sudo bash add-google-cloud-ops-agent-repo.sh --also-install
          sudo service google-cloud-ops-agent start
- group: packer
  modules:
  - id: custom-image
    source: modules/packer/custom-image
    kind: packer
    use:
    - network1
    - packer_script
    settings:
      source_image_project_id: [schedmd-slurm-public]
      source_image_family: slurm-gcp-6-6-debian-11
      disk_size: $(vars.disk_size_gb)
      image_family: $(vars.new_image_family)
      machine_type: c2-standard-8 # building this image does not require a GPU-enabled VM
      state_timeout: 30m


- group: cluster
  modules:

  - id: n1t4_nodeset
    source: community/modules/compute/schedmd-slurm-gcp-v6-nodeset
    use: [network1]
    settings:
      zones: $(vars.zone_list)
      node_count_dynamic_max: 1
      bandwidth_tier: gvnic_enabled
      disk_size_gb: $(vars.disk_size_gb)
      enable_public_ips: true
      enable_placement: false
      advanced_machine_features:
        threads_per_core: 1
      machine_type: n1-standard-96
      guest_accelerator:
      - type: nvidia-tesla-t4
        count: 4

      on_host_maintenance: TERMINATE
      instance_image:
        family: $(vars.new_image_family)
        project: $(vars.project_id)
  - id: n1t4_partition
    source: community/modules/compute/schedmd-slurm-gcp-v6-partition
    use: [n1t4_nodeset]
    settings:
      partition_name: n1t4
      is_default: true
      exclusive: false


  - id: n2_nodeset
    source: community/modules/compute/schedmd-slurm-gcp-v6-nodeset
    use: [network1]
    settings:
      zones: $(vars.zone_list)
      node_count_dynamic_max: 1
      bandwidth_tier: gvnic_enabled
      disk_size_gb: $(vars.disk_size_gb)
      enable_public_ips: true
      advanced_machine_features:
        threads_per_core: 1
      machine_type: n2-standard-4
      on_host_maintenance: TERMINATE
      instance_image:
        family: $(vars.new_image_family)
        project: $(vars.project_id)
  - id: n2_partition
    source: community/modules/compute/schedmd-slurm-gcp-v6-partition
    use: [n2_nodeset]
    settings:
      partition_name: n2
      is_default: true

  - id: g2_nodeset
    source: community/modules/compute/schedmd-slurm-gcp-v6-nodeset
    use: [network1]
    settings:
      zones: $(vars.zone_list)
      node_count_dynamic_max: 1
      bandwidth_tier: gvnic_enabled
      disk_size_gb: $(vars.disk_size_gb)
      enable_public_ips: true
      advanced_machine_features:
        threads_per_core: 1
      machine_type: g2-standard-96
      on_host_maintenance: TERMINATE
      instance_image:
        family: $(vars.new_image_family)
        project: $(vars.project_id)
  - id: g2_partition
    source: community/modules/compute/schedmd-slurm-gcp-v6-partition
    use: [g2_nodeset]
    settings:
      partition_name: g2gpu8
      is_default: false

  - id: slurm_login
    source: community/modules/scheduler/schedmd-slurm-gcp-v6-login
    use: [network1]
    settings:
      name_prefix: login
      machine_type: n2-standard-4
      enable_login_public_ips: true
      instance_image: 
        family: $(vars.new_image_family)
        project: $(vars.project_id)
        

  - id: slurm_controller
    source: community/modules/scheduler/schedmd-slurm-gcp-v6-controller
    use:
    - network1
    - n1t4_partition
    - n2_partition
    - g2_partition
    - slurm_login
    - homefs
    - data_bucket
    settings:
      enable_controller_public_ips: true
      controller_startup_script: $(startup_script.startup_script)
      controller_startup_scripts_timeout: 21600
      login_startup_script: $(startup_script.startup_script)
      login_startup_scripts_timeout: 21600
      instance_image: 
        family: $(vars.new_image_family)
        project: $(vars.project_id)