# Global release test configuration file.
# All your release test configuration should go here. Adding release tests here
# will automatically enable them in the Buildkite release testing schedules
# (except they have frequency: manual).
# Here is an example configuration for reference:
#- name: example_test
#  # Tests with the same group will be grouped in the Buildkite UI
#  group: Example group
#  # Provide the working directory which will be uploaded to the cluster
#  working_dir: example_dir
#
#  # How often to run the tests.
#  # One of [manual, any, multi, nightly, nightly-3x, weekly].
#  # Descriptions of each frequency (that's not immediately obvious):
#  # - manual: Not run on a schedule, but can be manually run through the buildkite UI.
#  # - nightly-3x: Run 3 times a week (Monday, Wednesday, Friday).
#  frequency: weekly
#  # Owning team. This field will be persisted to the database
#  team: ml
#
#  # Python version. This optional field determines which Python version to run tests
#  # on. This must be a string!
#  python: "3.7"
#
#  # Cluster information
#  cluster:
#    # Location of cluster compute, relative to working_dir
#    cluster_compute: cluster_compute.yaml
#    # Autosuspend parameter passed to the cluster.
#    # The cluster will automatically terminate if inactive for this
#    # many minutes. Defaults to 10 if not set.
#    autosuspend_mins: 10
#    # Optional cloud_id to use instead of the default cloud
#    cloud_id: cld_12345678
#    # Alternatively, you can specify a cloud name
#    cloud_name: anyscale_default_cloud
#
#  # Run configuration for the test
#  run:
#    # If you want to wait for nodes to be ready, you can specify this here:
#    wait_for_nodes:
#      # Number of nodes
#      num_nodes: 16
#      # Timeout for waiting for nodes. If nodes are not up by then, the
#      # test will fail.
#      timeout: 600
#
#    # Optional prepare script to be run on the cluster before the test script
#    prepare: python prepare.py
#    # The prepare command can have a separate timeout
#    prepare_timeout: 300
#
#    # Main script to run as the test script
#    script: python workloads/train_small.py
#    # Timeout in seconds. After this time the test is considered as failed.
#    timeout: 600
#
#  # You can specify smoke test definitions here. If a smoke test is triggered,
#  # it will deep update the main test configuration with the values provided
#  # here. Smoke tests will automatically run with IS_SMOKE_TEST=1 as en
#  # environment variable and receive the --smoke-test flag as a parameter in the
#  # run script.
#  smoke_test:
#    # Smoke tests can have different frequencies. A smoke test is only triggered
#    # when the regular test is not matched.
#    frequency: nightly
#    # Here we adjust the run timeout down and run on less nodes. The test script
#    # remains the same.
#    run:
#      timeout: 300
#      wait_for_nodes:
#        num_nodes: 4
#        timeout: 600
#
#  # After the test finished, this handler (in alerts/) will process the results.
#  # It can then let the test fail, e.g. if a metric regression is observed.
#  alert: default

#######################
# Cluster scaling tests
#######################
- name: cluster_tune_scale_up_down
  group: Cluster tests
  working_dir: cluster_tests

  frequency: nightly
  team: ml

  cluster:
    byod: {}
    cluster_compute: cpt_autoscaling_1-3_aws.yaml

  run:
    timeout: 3600
    script: python workloads/tune_scale_up_down.py
    wait_for_nodes:
      num_nodes: 0

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: cpt_autoscaling_1-3_gce.yaml

  alert: default

############################
# Batch Inference Benchmarks
############################

# 10 GB image classification raw images with 1 GPU.
# 1 g4dn.4xlarge
- name: torch_batch_inference_1_gpu_10gb_raw
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data
  cluster:
    byod:
      type: gpu
    cluster_compute: compute_gpu_1_cpu_16_aws.yaml

  run:
    timeout: 500
    script: python gpu_batch_inference.py --data-directory=10G-image-data-synthetic-raw --data-format raw

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gpu_1_cpu_16_gce.yaml

# 10 GB image classification parquet with 1 GPU.
# 1 g4dn.4xlarge
- name: torch_batch_inference_1_gpu_10gb_parquet
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data
  cluster:
    byod:
      type: gpu
    cluster_compute: compute_gpu_1_cpu_16_aws.yaml

  run:
    timeout: 500
    script: python gpu_batch_inference.py --data-directory=10G-image-data-synthetic-raw-parquet --data-format parquet

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gpu_1_cpu_16_gce.yaml


# 300 GB image classification raw images with 16 GPUs
# 4 g4dn.12xlarge
- name: torch_batch_inference_16_gpu_300gb_raw
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data
  cluster:
    byod:
      type: gpu
    cluster_compute: compute_gpu_4x4_aws.yaml

  run:
    timeout: 1000
    script: python gpu_batch_inference.py --data-directory 300G-image-data-synthetic-raw --data-format raw

    wait_for_nodes:
        num_nodes: 4

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gpu_4x4_gce.yaml


- name: chaos_torch_batch_inference_16_gpu_300gb_raw
  group: data-tests
  working_dir: nightly_tests
  stable: false

  frequency: nightly
  team: data
  cluster:
    byod:
      type: gpu
    cluster_compute: dataset/compute_gpu_4x4_aws.yaml

  run:
    timeout: 1000
    prepare: python setup_chaos.py --max-to-kill 2 --kill-delay 30
    script: python dataset/gpu_batch_inference.py --data-directory 300G-image-data-synthetic-raw --data-format raw

    wait_for_nodes:
        num_nodes: 4

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: dataset/compute_gpu_4x4_gce.yaml


# 300 GB image classification parquet data with 16 GPUs
# 4 g4dn.12xlarge
- name: torch_batch_inference_16_gpu_300gb_parquet
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_gpu_4x4_aws.yaml

  run:
    timeout: 1000
    script: python gpu_batch_inference.py --data-directory 300G-image-data-synthetic-raw-parquet --data-format parquet

    wait_for_nodes:
        num_nodes: 4

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gpu_4x4_gce.yaml

# 10 TB image classification parquet data with heterogenous cluster
# 10 g4dn.12xlarge, 10 m5.16xlarge
- name: torch_batch_inference_hetero_10tb_parquet
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: weekly
  team: data

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_hetero_10x10_aws.yaml

  run:
    timeout: 2000
    script: python gpu_batch_inference.py --data-directory 10T-image-data-synthetic-raw-parquet --data-format parquet

    wait_for_nodes:
      num_nodes: 20

  alert: default


#########################
# AIR release tests
#########################
- name: tune_with_frequent_pausing
  group: AIR tests
  working_dir: air_tests

  frequency: nightly-3x
  team: ml

  cluster:
    byod:
      runtime_env:
        - RAY_memory_usage_threshold=0.5
        - automatic_object_spilling_enabled=0
    cluster_compute: frequent_pausing/compute_config_aws.yaml

  run:
    timeout: 600  # 10min
    long_running: true
    script: python frequent_pausing/script.py

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: frequent_pausing/compute_config_gce.yaml

  alert: default


- name: long_running_horovod_tune_test
  group: AIR tests
  working_dir: air_tests

  frequency: weekly
  team: ml

  cluster:
    byod:
      type: gpu
      post_build_script: byod_horovod_master_test.sh
    cluster_compute: horovod/compute_tpl_aws.yaml

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: horovod/compute_tpl_gce.yaml

  run:
    timeout: 36000
    script: python horovod/workloads/horovod_tune_test.py
    long_running: true
    wait_for_nodes:
      num_nodes: 2

  smoke_test:
    frequency: manual

    run:
      timeout: 3600

  alert: default


# Ray AIR distributed Torch benchmarks
- name: air_benchmark_torch_mnist_cpu_4x1
  group: AIR tests
  working_dir: air_tests/air_benchmarks

  frequency: nightly
  team: ml

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_cpu_4_aws.yaml

  run:
    timeout: 3600
    script: python workloads/torch_benchmark.py run --num-runs 3 --num-epochs 20 --num-workers 4 --cpus-per-worker 8

    wait_for_nodes:
      num_nodes: 4

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_cpu_4_gce.yaml

  alert: default

- name: air_benchmark_torch_mnist_gpu_4x4
  group: AIR tests
  working_dir: air_tests/air_benchmarks

  frequency: weekly
  team: ml

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_gpu_4x4_aws.yaml

  run:
    timeout: 4800
    script: python workloads/torch_benchmark.py run --num-runs 3 --num-epochs 120 --num-workers 16 --cpus-per-worker 4 --batch-size 1024 --use-gpu

    wait_for_nodes:
      num_nodes: 4

  smoke_test:
    frequency: nightly

    cluster:
      cluster_compute: compute_gpu_2x2_aws.yaml

    run:
      timeout: 3600
      script: python workloads/torch_benchmark.py run --num-runs 3 --num-epochs 60 --num-workers 4 --cpus-per-worker 4 --batch-size 512 --use-gpu

      wait_for_nodes:
        num_nodes: 2

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gpu_4x4_gce.yaml
      smoke_test:
        frequency: manual

  alert: default


- name: air_benchmark_torch_mnist_cpu_1x4
  group: AIR tests
  working_dir: air_tests/air_benchmarks

  frequency: nightly
  team: ml

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_cpu_1_aws.yaml

  run:
    timeout: 3600
    script: python workloads/torch_benchmark.py run --num-runs 3 --num-epochs 20 --num-workers 4 --cpus-per-worker 2

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_cpu_1_gce.yaml

  alert: default


- name: air_benchmark_torch_mnist_cpu_4x4
  group: AIR tests
  working_dir: air_tests/air_benchmarks

  frequency: nightly
  team: ml

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_cpu_4_aws.yaml

  run:
    timeout: 5400
    script: python workloads/torch_benchmark.py run --num-runs 3 --num-epochs 20 --num-workers 16 --cpus-per-worker 2

    wait_for_nodes:
      num_nodes: 4

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_cpu_4_gce.yaml

  alert: default

- name: air_benchmark_tune_torch_mnist
  group: AIR tests
  working_dir: air_tests/air_benchmarks

  frequency: nightly
  team: ml

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_cpu_8_aws.yaml

  run:
    timeout: 3600
    script: python workloads/tune_torch_benchmark.py --num-runs 3 --num-trials 8 --num-workers 4

    wait_for_nodes:
      num_nodes: 8

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_cpu_8_gce.yaml

  alert: default

- name: air_benchmark_tune_torch_mnist_gpu
  group: AIR tests
  working_dir: air_tests/air_benchmarks

  frequency: nightly
  team: ml

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_gpu_4x4_aws.yaml

  run:
    timeout: 3600
    script: python workloads/tune_torch_benchmark.py --num-runs 2 --num-trials 4 --num-workers 4 --use-gpu

    wait_for_nodes:
      num_nodes: 4

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gpu_4x4_gce.yaml

  alert: default

# Ray AIR distributed Tensorflow benchmarks
- name: air_benchmark_tensorflow_mnist_cpu_4x1
  group: AIR tests
  working_dir: air_tests/air_benchmarks

  frequency: nightly
  team: ml

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_cpu_4_aws.yaml

  run:
    timeout: 5400
    script: python workloads/tensorflow_benchmark.py run --num-runs 3 --num-epochs 20 --num-workers 4 --cpus-per-worker 8

    wait_for_nodes:
      num_nodes: 4

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_cpu_4_gce.yaml

  alert: default


- name: air_benchmark_tensorflow_mnist_cpu_1x4
  group: AIR tests
  working_dir: air_tests/air_benchmarks

  frequency: nightly
  team: ml

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_cpu_1_aws.yaml

  run:
    timeout: 5400
    script: python workloads/tensorflow_benchmark.py run --num-runs 3 --num-epochs 20 --num-workers 4 --cpus-per-worker 2

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_cpu_1_gce.yaml

  alert: default


- name: air_benchmark_tensorflow_mnist_cpu_4x4
  group: AIR tests
  working_dir: air_tests/air_benchmarks

  frequency: nightly
  team: ml

  stable: false

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_cpu_4_aws.yaml

  run:
    timeout: 5400
    script: python workloads/tensorflow_benchmark.py run --num-runs 3 --num-epochs 20 --num-workers 16 --cpus-per-worker 2

    wait_for_nodes:
      num_nodes: 4

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_cpu_4_gce.yaml

  alert: default


- name: air_benchmark_tensorflow_mnist_gpu_4x4
  group: AIR tests
  working_dir: air_tests/air_benchmarks

  frequency: weekly
  team: ml

  stable: false

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_gpu_4x4_aws.yaml

  run:
    timeout: 5400
    script: python workloads/tensorflow_benchmark.py run --num-runs 3 --num-epochs 200 --num-workers 16 --cpus-per-worker 4 --batch-size 1024 --use-gpu

    wait_for_nodes:
      num_nodes: 4

  smoke_test:
    frequency: nightly

    cluster:
      cluster_compute: compute_gpu_2x2_aws.yaml

    run:
      script: python workloads/tensorflow_benchmark.py run --num-runs 3 --num-epochs 60 --num-workers 4 --cpus-per-worker 4 --batch-size 512 --use-gpu

      wait_for_nodes:
        num_nodes: 2

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gpu_4x4_gce.yaml
      smoke_test:
        frequency: manual

  alert: default

- name: air_benchmark_pytorch_training_e2e_gpu_1x1_20gb
  group: AIR tests
  working_dir: air_tests/air_benchmarks

  frequency: nightly
  team: ml

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_gpu_1_aws.yaml

  run:
    timeout: 3600
    script: python workloads/pytorch_training_e2e.py --data-size-gb 20

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gpu_1_gce.yaml


- name: air_benchmark_pytorch_training_e2e_gpu_4x4_100gb
  group: AIR tests
  working_dir: air_tests/air_benchmarks

  frequency: nightly
  team: ml

  stable: false

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_gpu_4x4_aws.yaml

  run:
    timeout: 10800
    script: python workloads/pytorch_training_e2e.py --data-size-gb=100 --num-workers=16

    wait_for_nodes:
      num_nodes: 4

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gpu_4x4_gce.yaml

# Test tiny, and medium input files to check that performance stays about
# constant.
- name: ray-data-resnet50-ingest-file-size-benchmark
  group: AIR tests
  working_dir: air_tests/air_benchmarks/mlperf-train

  frequency: nightly

  team: data
  cluster:
    byod:
      type: gpu
      runtime_env:
        - RAY_task_oom_retries=50
        - RAY_min_memory_free_bytes=1000000000
    cluster_compute: compute_cpu_16.yaml

  run:
    timeout: 3600
    script: bash file_size_benchmark.sh

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gce_cpu_16.yaml

# Test huge files to check that we do not OOM.
- name: ray-data-resnet50-ingest-out-of-memory-benchmark
  group: AIR tests
  working_dir: air_tests/air_benchmarks/mlperf-train

  stable: false

  frequency: nightly

  team: data
  cluster:
    byod:
      type: gpu
      runtime_env:
        - RAY_task_oom_retries=50
        - RAY_min_memory_free_bytes=1000000000
    cluster_compute: compute_cpu_16.yaml

  run:
    timeout: 3600
    script: bash oom_benchmark.sh

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gce_cpu_16.yaml

#######################
# AIR examples
#######################


# Test additional CPU nodes for preprocessing.
- name: air_example_dreambooth_finetuning
  group: AIR examples
  working_dir: air_examples/dreambooth

  stable: false

  frequency: weekly
  team: ml

  cluster:
    byod:
      type: gpu
    cluster_compute: dreambooth_compute_aws.yaml

  run:
    timeout: 1800
    script: pip install -Ur dreambooth/requirements.txt && bash dreambooth_run.sh
    artifact_path: /tmp/artifacts/example_out.jpg

  # variations: A10G not available on GCE, yet.

- name: air_example_dreambooth_finetuning_lora
  group: AIR examples
  working_dir: air_examples/dreambooth

  stable: false

  frequency: weekly
  team: ml

  cluster:
    byod:
      type: gpu
    cluster_compute: dreambooth_compute_aws.yaml

  run:
    timeout: 1800
    script: pip install -Ur dreambooth/requirements.txt && bash dreambooth_run.sh --lora
    artifact_path: /tmp/artifacts/example_out.jpg

- name: air_example_gptj_deepspeed_fine_tuning
  group: AIR examples
  working_dir: air_examples/gptj_deepspeed_finetuning
  frequency: weekly
  team: ml
  cluster:
    byod:
      type: gpu
      post_build_script: byod_gptj_test.sh
    cluster_compute: gptj_deepspeed_compute_aws.yaml

  run:
    timeout: 4500
    script: python test_myst_doc.py --path gptj_deepspeed_fine_tuning.ipynb

- name: air_example_dolly_v2_lightning_fsdp_finetuning
  group: AIR examples
  working_dir: air_examples/dolly_v2_lightning_fsdp_finetuning
  frequency: weekly
  team: ml
  cluster:
    byod:
      type: gpu
      post_build_script: byod_dolly_test.sh
    cluster_compute: dolly_v2_fsdp_compute_aws.yaml

  run:
    timeout: 4700
    script: python test_myst_doc.py --path lightning-llm-finetuning-7b.ipynb

  # variations: TODO(jungong): add GCP variation.

- name: air_example_vicuna_13b_lightning_deepspeed_finetuning
  group: AIR examples
  working_dir: air_examples/vicuna_13b_lightning_deepspeed_finetuning
  frequency: weekly
  team: ml
  cluster:
    byod:
      type: gpu
      post_build_script: byod_vicuna_test.sh
    cluster_compute: vicuna_13b_deepspeed_compute_aws.yaml

  run:
    timeout: 4700
    script: python test_myst_doc.py --path vicuna_13b_lightning_deepspeed_finetune.ipynb

#####################################
# Workspace templates release tests #
#####################################

- name: workspace_template_batch_inference
  group: Workspace templates
  working_dir: workspace_templates/01_batch_inference
  frequency: nightly-3x
  team: data
  cluster:
    byod:
      type: gpu
    cluster_compute: ../testing/compute_configs/gpu/aws.yaml

  run:
    timeout: 600
    script: jupyter nbconvert --to script --output _test start.ipynb && ipython _test.py

  variations:
      - __suffix__: aws
      - __suffix__: gce
        env: gce
        frequency: manual
        cluster:
          cluster_compute: ../testing/compute_configs/gpu/gce.yaml


- name: workspace_template_many_model_training
  group: Workspace templates
  working_dir: workspace_templates/02_many_model_training
  frequency: nightly-3x
  team: ml

  cluster:
    byod:
      type: gpu
    cluster_compute: ../testing/compute_configs/cpu/aws.yaml

  run:
    timeout: 600
    script: jupyter nbconvert --to script --output _test start.ipynb && ipython _test.py

  variations:
      - __suffix__: aws
      - __suffix__: gce
        env: gce
        frequency: manual
        cluster:
          cluster_compute: ../testing/compute_configs/cpu/gce.yaml


- name: workspace_template_serving_stable_diffusion
  group: Workspace templates
  working_dir: workspace_templates/03_serving_stable_diffusion
  frequency: nightly-3x
  team: serve
  cluster:
    byod:
      type: gpu
      post_build_script: byod_stable_diffusion.sh
    cluster_compute: ../testing/compute_configs/gpu/aws.yaml

  run:
    timeout: 600
    script: jupyter nbconvert --to script --output _test start.ipynb && ipython _test.py && serve run app:entrypoint --non-blocking && python query.py

  variations:
      - __suffix__: aws
      - __suffix__: gce
        env: gce
        frequency: manual
        cluster:
          cluster_compute: ../testing/compute_configs/gpu/gce.yaml

- name: workspace_template_finetuning_llms_with_deepspeed_llama_2_7b
  group: Workspace templates
  working_dir: workspace_templates/04_finetuning_llms_with_deepspeed
  frequency: nightly-3x
  team: ml
  cluster:
    byod:
      type: cu123
      # This needs to be in sync with requirements under go/llm-forge.
      post_build_script: byod_finetune_llvms.sh
    cluster_compute: ../testing/compute_configs/04_finetuning_llms_with_deepspeed/aws_7b.yaml

  run:
    timeout: 1000
    script: chmod +x ./run_llama_ft.sh && ./run_llama_ft.sh --size=7b --as-test

  variations:
      - __suffix__: aws
      - __suffix__: gce
        env: gce
        frequency: manual
        cluster:
          cluster_compute: ../testing/compute_configs/04_finetuning_llms_with_deepspeed/gce_7b.yaml

- name: workspace_template_finetuning_llms_with_deepspeed_llama_2_7b_lora
  group: Workspace templates
  working_dir: workspace_templates/04_finetuning_llms_with_deepspeed
  frequency: nightly-3x
  team: ml
  cluster:
    byod:
      type: cu123
      # This needs to be in sync with requirements under go/llm-forge.
      post_build_script: byod_finetune_llvms.sh
    cluster_compute: ../testing/compute_configs/04_finetuning_llms_with_deepspeed/aws_7b.yaml

  run:
    timeout: 1000
    script: chmod +x ./run_llama_ft.sh && ./run_llama_ft.sh --size=7b --lora --as-test

  variations:
      - __suffix__: aws
      - __suffix__: gce
        env: gce
        frequency: manual
        cluster:
          cluster_compute: ../testing/compute_configs/04_finetuning_llms_with_deepspeed/gce_7b.yaml

#######################
# ML user tests
#######################
- name: ml_user_horovod_user_test_latest
  group: ML user tests
  working_dir: ml_user_tests

  frequency: nightly-3x
  team: ml

  cluster:
    byod:
      type: gpu
      post_build_script: byod_horovod_test.sh
    cluster_compute: horovod/compute_tpl_aws.yaml

  run:
    timeout: 1200
    script: python horovod/horovod_user_test.py
    wait_for_nodes:
      num_nodes: 4

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: horovod/compute_tpl_gce.yaml

  alert: default

- name: ml_user_horovod_user_test_master
  group: ML user tests
  working_dir: ml_user_tests

  frequency: nightly-3x
  team: ml

  cluster:
    byod:
      type: gpu
      post_build_script: byod_horovod_master_test.sh
    cluster_compute: horovod/compute_tpl_aws.yaml

  run:
    timeout: 1200
    script: python horovod/horovod_user_test.py
    wait_for_nodes:
      num_nodes: 4

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: horovod/compute_tpl_gce.yaml

  alert: default

- name: ml_user_train_tensorflow_mnist_test
  group: ML user tests
  working_dir: ml_user_tests

  frequency: nightly-3x
  team: ml

  cluster:
    byod:
      runtime_env:
        - TRAIN_PLACEMENT_GROUP_TIMEOUT_S=2000
      type: cu123
    cluster_compute: train/compute_tpl_aws.yaml

  run:
    timeout: 36000
    script: python train/train_tensorflow_mnist_test.py
    wait_for_nodes:
      num_nodes: 3

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: train/compute_tpl_gce.yaml

  alert: default

- name: ml_user_train_torch_linear_test
  group: ML user tests
  working_dir: ml_user_tests

  frequency: nightly-3x
  team: ml

  cluster:
    byod:
      runtime_env:
        - TRAIN_PLACEMENT_GROUP_TIMEOUT_S=2000
      type: gpu
    cluster_compute: train/compute_tpl_aws.yaml

  run:
    timeout: 36000
    script: python train/train_torch_linear_test.py
    wait_for_nodes:
      num_nodes: 3

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: train/compute_tpl_gce.yaml

  alert: default

- name: ml_user_tune_rllib_connect_test
  group: ML user tests
  working_dir: ml_user_tests

  frequency: nightly-3x
  team: ml

  cluster:
    byod:
      type: gpu
      post_build_script: byod_rllib_test.sh
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
        - LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/ray/.mujoco/mujoco210/bin
    cluster_compute: tune_rllib/compute_tpl_aws.yaml

  run:
    timeout: 2000
    script: python tune_rllib/run_connect_tests.py
    wait_for_nodes:
      num_nodes: 9

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: tune_rllib/compute_tpl_gce.yaml

  alert: default

#######################
# Tune cloud  tests
#######################
- name: tune_cloud_long_running_cloud_storage
  group: Tune cloud tests
  working_dir: tune_tests/cloud_tests
  frequency: weekly
  team: ml

  cluster:
    byod: {}
    cluster_compute: tpl_aws_1x4.yaml

  run:
    # 14 hours
    timeout: 50400
    long_running: true
    script: python workloads/long_running_cloud_storage.py s3://tune-cloud-tests/long_running_cloud_storage

  # NOTE: This smoke test is not useful to run because the point of the test
  # is to be long running. This is just for debugging updates to the test quickly.
  smoke_test:
    frequency: manual
    run:
      timeout: 600

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: tpl_gce_1x4.yaml
      run:
        # 14 hours
        timeout: 50400
        long_running: true
        script: python workloads/long_running_cloud_storage.py gs://tune-cloud-tests/long_running_cloud_storage
        wait_for_nodes:
          num_nodes: 1

  alert: long_running_tests

########################
# Tune scalability tests
########################

- name: tune_scalability_bookkeeping_overhead
  group: Tune scalability tests
  working_dir: tune_tests/scalability_tests

  frequency: nightly
  team: ml

  cluster:
    byod: {}
    cluster_compute: tpl_1x16.yaml

  run:
    timeout: 1200
    script: python workloads/test_bookkeeping_overhead.py

  alert: tune_tests

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: tpl_gce_1x16.yaml

- name: tune_scalability_durable_trainable
  group: Tune scalability tests
  working_dir: tune_tests/scalability_tests

  frequency: nightly
  team: ml

  cluster:
    byod: {}
    cluster_compute: tpl_16x2.yaml

  run:
    timeout: 900
    script: python workloads/test_durable_trainable.py --bucket s3://tune-cloud-tests/scalability_durable_trainable
    wait_for_nodes:
      num_nodes: 16

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      run:
        timeout: 900
        script: python workloads/test_durable_trainable.py --bucket gs://tune-cloud-tests/scalability_durable_trainable
        wait_for_nodes:
          num_nodes: 16
      cluster:
        cluster_compute: tpl_gce_16x2.yaml

  alert: tune_tests


- name: tune_scalability_durable_multifile_checkpoints
  group: Tune scalability tests
  working_dir: tune_tests/scalability_tests

  frequency: nightly
  team: ml

  cluster:
    byod: {}
    cluster_compute: tpl_16x2.yaml

  run:
    timeout: 900
    script: python workloads/test_durable_multifile_checkpoints.py --bucket s3://tune-cloud-tests/scalability_durable_multifile_checkpoints
    wait_for_nodes:
      num_nodes: 16

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      run:
        timeout: 900
        script: python workloads/test_durable_multifile_checkpoints.py --bucket gs://tune-cloud-tests/scalability_durable_multifile_checkpoints
        wait_for_nodes:
          num_nodes: 16
      cluster:
        cluster_compute: tpl_gce_16x2.yaml

  alert: tune_tests

- name: tune_scalability_long_running_large_checkpoints
  group: Tune scalability tests
  working_dir: tune_tests/scalability_tests

  frequency: weekly
  team: ml

  cluster:
    byod: {}
    cluster_compute: tpl_1x32_hd.yaml

  run:
    timeout: 86400
    script: python workloads/test_long_running_large_checkpoints.py
    long_running: true

  smoke_test:
    frequency: nightly

    run:
      timeout: 3600

  alert: tune_tests

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      smoke_test:
        frequency: manual
      cluster:
        cluster_compute: tpl_gce_1x32_hd.yaml

- name: tune_scalability_network_overhead
  group: Tune scalability tests
  working_dir: tune_tests/scalability_tests

  frequency: weekly
  team: ml

  cluster:
    byod: {}
    cluster_compute: tpl_100x2.yaml

  run:
    timeout: 750
    prepare_timeout: 1200
    script: python workloads/test_network_overhead.py
    wait_for_nodes:
      num_nodes: 100

  alert: tune_tests

  variations:
    - __suffix__: aws
    - __suffix__: smoke-test
      frequency: nightly
      cluster:
        cluster_compute: tpl_20x2.yaml
      run:
        timeout: 750
        prepare_timeout: 600
        script: python workloads/test_network_overhead.py --smoke-test
        wait_for_nodes:
          num_nodes: 20
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: tpl_gce_100x2.yaml

- name: tune_scalability_result_throughput_cluster
  group: Tune scalability tests
  working_dir: tune_tests/scalability_tests

  frequency: nightly-3x
  team: ml

  cluster:
    byod: {}
    cluster_compute: tpl_16x64.yaml

  run:
    timeout: 600
    script: python workloads/test_result_throughput_cluster.py

    wait_for_nodes:
      num_nodes: 16

  alert: tune_tests

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: tpl_gce_16x64.yaml

- name: tune_scalability_result_throughput_single_node
  group: Tune scalability tests
  working_dir: tune_tests/scalability_tests

  frequency: nightly
  team: ml

  cluster:
    byod: {}
    cluster_compute: tpl_1x96.yaml

  run:
    timeout: 600
    script: python workloads/test_result_throughput_single_node.py

  alert: tune_tests

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: tpl_gce_1x96.yaml


############################
# Tune fault tolerance tests
############################
- name: tune_worker_fault_tolerance
  group: Tune fault tolerance tests
  working_dir: tune_tests/fault_tolerance_tests

  stable: true

  frequency: nightly-3x
  team: ml

  cluster:
    byod: {}
    cluster_compute: tpl_aws_16x1.yaml

  run:
    timeout: 5400
    script: python workloads/test_tune_worker_fault_tolerance.py --bucket s3://tune-cloud-tests/worker_fault_tolerance

    wait_for_nodes:
      num_nodes: 16

# Disabled until we can kill nodes in GCE
#  variations:
#    - __suffix__: aws
#    - __suffix__: gce
#      env: gce
#      frequency: manual
#      run:
#        timeout: 5400
#        script: python workloads/test_tune_worker_fault_tolerance.py --bucket gs://tune-cloud-tests/worker_fault_tolerance
#
#        wait_for_nodes:
#          num_nodes: 16
#      cluster:
#        cluster_compute: tpl_gce_16x1.yaml

########################
# Golden Notebook tests
########################
- name: golden_notebook_torch_tune_serve_test
  group: Golden Notebook tests
  working_dir: golden_notebook_tests

  frequency: nightly-3x
  team: ml

  cluster:
    byod:
      type: gpu
    cluster_compute: gpu_tpl_aws.yaml

  run:
    timeout: 600
    script: python workloads/torch_tune_serve_test.py
    wait_for_nodes:
      num_nodes: 2

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: gpu_tpl_gce.yaml

  alert: default


#######################
# Long running tests
#######################

- name: long_running_actor_deaths
  group: Long running tests
  working_dir: long_running_tests

  frequency: weekly

  team: core
  cluster:
    byod:
      pip:
        # TODO: https://github.com/Farama-Foundation/AutoROM/issues/48
        - https://ray-ci-deps-wheels.s3.us-west-2.amazonaws.com/AutoROM.accept_rom_license-0.5.4-py3-none-any.whl
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
        - LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/ray/.mujoco/mujoco210/bin
    cluster_compute: tpl_cpu_1.yaml

  run:
    timeout: 86400
    script: python workloads/actor_deaths.py
    long_running: true

  smoke_test:
    frequency: nightly

    run:
      timeout: 3600

  alert: long_running_tests

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      smoke_test:
        frequency: manual
      cluster:
        cluster_compute: tpl_cpu_1_gce.yaml

- name: long_running_apex
  group: Long running tests
  working_dir: long_running_tests
  stable: false

  frequency: weekly
  team: rllib

  cluster:
    byod:
      type: gpu
      post_build_script: byod_rllib_test.sh
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
        - LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/ray/.mujoco/mujoco210/bin
    cluster_compute: tpl_cpu_3.yaml

  run:
    timeout: 86400
    script: python workloads/apex.py
    long_running: true
    wait_for_nodes:
      num_nodes: 3

  smoke_test:
    frequency: nightly

    run:
      timeout: 3600


  alert: long_running_tests

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      smoke_test:
        frequency: manual
        run:
          timeout: 3600
      cluster:
        cluster_compute: tpl_cpu_3_gce.yaml

- name: long_running_impala
  group: Long running tests
  working_dir: long_running_tests

  frequency: weekly
  team: rllib

  cluster:
    byod:
      type: gpu
      post_build_script: byod_rllib_test.sh
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
        - LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/ray/.mujoco/mujoco210/bin
    cluster_compute: tpl_cpu_1_large.yaml

  run:
    timeout: 86400
    script: python workloads/impala.py
    long_running: true

  smoke_test:
    frequency: nightly

    run:
      timeout: 3600

  alert: long_running_tests

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      smoke_test:
        frequency: manual
        run:
          timeout: 3600
      cluster:
        cluster_compute: tpl_cpu_1_large_gce.yaml

- name: long_running_many_actor_tasks
  group: Long running tests
  working_dir: long_running_tests

  frequency: weekly

  team: core
  cluster:
    byod:
      pip:
        # TODO: https://github.com/Farama-Foundation/AutoROM/issues/48
        - https://ray-ci-deps-wheels.s3.us-west-2.amazonaws.com/AutoROM.accept_rom_license-0.5.4-py3-none-any.whl
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
    cluster_compute: tpl_cpu_1.yaml

  run:
    timeout: 86400
    script: python workloads/many_actor_tasks.py
    long_running: true

  smoke_test:
    frequency: nightly

    run:
      timeout: 3600

  alert: long_running_tests

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      smoke_test:
        frequency: manual
        run:
          timeout: 3600
      cluster:
        cluster_compute: tpl_cpu_1_gce.yaml

- name: long_running_many_drivers
  group: Long running tests
  working_dir: long_running_tests

  frequency: weekly

  team: core
  cluster:
    byod:
      pip:
        # TODO: https://github.com/Farama-Foundation/AutoROM/issues/48
        - https://ray-ci-deps-wheels.s3.us-west-2.amazonaws.com/AutoROM.accept_rom_license-0.5.4-py3-none-any.whl
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
    cluster_compute: tpl_cpu_4.yaml

  run:
    timeout: 86400
    script: python workloads/many_drivers.py --iteration-num=4000
    long_running: true
    wait_for_nodes:
      num_nodes: 4

  smoke_test:
    frequency: nightly

    run:
      timeout: 3600

  alert: long_running_tests

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      smoke_test:
        frequency: manual
        run:
          timeout: 3600
      cluster:
        cluster_compute: tpl_cpu_4_gce.yaml

- name: long_running_many_ppo
  group: Long running tests
  working_dir: long_running_tests

  stable: false

  frequency: weekly
  team: ml

  cluster:
    byod:
      type: gpu
      post_build_script: byod_rllib_test.sh
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
        - LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/ray/.mujoco/mujoco210/bin
    cluster_compute: many_ppo.yaml

  run:
    timeout: 86400
    script: python workloads/many_ppo.py
    long_running: true
    wait_for_nodes:
      num_nodes: 1


  smoke_test:
    frequency: nightly

    run:
      timeout: 3600

  alert: long_running_tests

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      smoke_test:
        frequency: manual
        run:
          timeout: 3600
      cluster:
        cluster_compute: many_ppo_gce.yaml

- name: long_running_many_tasks
  group: Long running tests
  working_dir: long_running_tests

  frequency: weekly

  team: core
  cluster:
    byod:
      pip:
        # TODO: https://github.com/Farama-Foundation/AutoROM/issues/48
        - https://ray-ci-deps-wheels.s3.us-west-2.amazonaws.com/AutoROM.accept_rom_license-0.5.4-py3-none-any.whl
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
    cluster_compute: tpl_cpu_1.yaml

  run:
    timeout: 86400
    script: python workloads/many_tasks.py
    long_running: true

  smoke_test:
    frequency: nightly

    run:
      timeout: 3600

  alert: long_running_tests

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      smoke_test:
        frequency: manual
        run:
          timeout: 3600
      cluster:
        cluster_compute: tpl_cpu_1_gce.yaml

- name: long_running_many_tasks_serialized_ids
  group: Long running tests
  working_dir: long_running_tests

  frequency: weekly

  team: core
  cluster:
    byod:
      pip:
        # TODO: https://github.com/Farama-Foundation/AutoROM/issues/48
        - https://ray-ci-deps-wheels.s3.us-west-2.amazonaws.com/AutoROM.accept_rom_license-0.5.4-py3-none-any.whl
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
    cluster_compute: tpl_cpu_1.yaml

  run:
    timeout: 86400
    script: python workloads/many_tasks_serialized_ids.py
    long_running: true

  smoke_test:
    frequency: nightly

    run:
      timeout: 3600

  alert: long_running_tests

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      smoke_test:
        frequency: manual
        run:
          timeout: 3600
      cluster:
        cluster_compute: tpl_cpu_1_gce.yaml

- name: long_running_node_failures
  group: Long running tests
  working_dir: long_running_tests

  frequency: weekly

  team: core
  cluster:
    byod:
      pip:
        # TODO: https://github.com/Farama-Foundation/AutoROM/issues/48
        - https://ray-ci-deps-wheels.s3.us-west-2.amazonaws.com/AutoROM.accept_rom_license-0.5.4-py3-none-any.whl
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
    cluster_compute: tpl_cpu_1.yaml

  run:
    timeout: 86400
    script: python workloads/node_failures.py
    long_running: true

  smoke_test:
    frequency: nightly

    run:
      timeout: 3600

  alert: long_running_tests

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      smoke_test:
        frequency: manual
        run:
          timeout: 3600
      cluster:
        cluster_compute: tpl_cpu_1_gce.yaml

- name: long_running_serve
  group: Long running tests
  working_dir: long_running_tests

  frequency: weekly
  team: serve

  cluster:
    byod:
      pip:
        # TODO: https://github.com/Farama-Foundation/AutoROM/issues/48
        - https://ray-ci-deps-wheels.s3.us-west-2.amazonaws.com/AutoROM.accept_rom_license-0.5.4-py3-none-any.whl
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
    cluster_compute: tpl_cpu_1_large.yaml

  run:
    timeout: 86400
    script: python workloads/serve.py
    long_running: true

  smoke_test:
    frequency: nightly

    run:
      timeout: 3600

  alert: long_running_tests

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      smoke_test:
        frequency: manual
        run:
          timeout: 3600
      cluster:
        cluster_compute: tpl_cpu_1_gce.yaml

- name: long_running_serve_failure
  group: Long running tests
  working_dir: long_running_tests

  stable: true

  frequency: weekly
  team: serve

  cluster:
    byod:
      pip:
        # TODO: https://github.com/Farama-Foundation/AutoROM/issues/48
        - https://ray-ci-deps-wheels.s3.us-west-2.amazonaws.com/AutoROM.accept_rom_license-0.5.4-py3-none-any.whl
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
    cluster_compute: tpl_cpu_1_c5.yaml

  run:
    timeout: 86400
    script: python workloads/serve_failure.py
    long_running: true

  smoke_test:
    frequency: nightly

    run:
      timeout: 600

  alert: long_running_tests

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      smoke_test:
        frequency: manual
        run:
          timeout: 86400
      cluster:
        cluster_compute: tpl_cpu_1_c5_gce.yaml

- name: long_running_many_jobs
  group: Long running tests
  working_dir: long_running_tests

  stable: true

  frequency: weekly
  team: serve

  cluster:
    byod:
      pip:
        # TODO: https://github.com/Farama-Foundation/AutoROM/issues/48
        - https://ray-ci-deps-wheels.s3.us-west-2.amazonaws.com/AutoROM.accept_rom_license-0.5.4-py3-none-any.whl
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
    cluster_compute: tpl_cpu_1.yaml

  run:
    timeout: 86400
    script: python workloads/long_running_many_jobs.py --num-clients=1
    long_running: true

  smoke_test:
    frequency: nightly

    run:
      timeout: 1800

  alert: long_running_tests

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      smoke_test:
        frequency: manual
        run:
          timeout: 3600
      cluster:
        cluster_compute: tpl_cpu_1_gce.yaml

- name: long_running_distributed_pytorch_pbt_failure
  group: Long running tests
  working_dir: long_running_distributed_tests

  frequency: weekly
  team: ml

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_tpl.yaml

  run:
    timeout: 86400
    script: python workloads/pytorch_pbt_failure.py
    long_running: true

  smoke_test:
    frequency: manual
    run:
      timeout: 3600

  alert: long_running_tests

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      smoke_test:
        frequency: manual
        run:
          timeout: 3600
      cluster:
        cluster_compute: compute_tpl_gce.yaml

########################
# Jobs tests
########################

- name: jobs_basic_local_working_dir
  group: Jobs tests
  working_dir: jobs_tests

  frequency: nightly
  team: serve

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_tpl_4_xlarge.yaml

  run:
    timeout: 600
    script: python workloads/jobs_basic.py --working-dir "workloads"
    wait_for_nodes:
      num_nodes: 4


  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_tpl_gce_4_xlarge.yaml

- name: jobs_basic_remote_working_dir
  group: Jobs tests
  working_dir: jobs_tests

  frequency: nightly
  team: serve

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_tpl_4_xlarge.yaml

  run:
    timeout: 600
    script: python workloads/jobs_basic.py --working-dir "https://github.com/anyscale/job-services-cuj-examples/archive/refs/heads/main.zip"
    wait_for_nodes:
      num_nodes: 4


  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_tpl_gce_4_xlarge.yaml

- name: jobs_remote_multi_node
  group: Jobs tests
  team: serve
  frequency: nightly
  working_dir: jobs_tests

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_tpl_4_xlarge.yaml
  run:
    timeout: 600
    script: python workloads/jobs_remote_multi_node.py
    wait_for_nodes:
      num_nodes: 4

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_tpl_gce_4_xlarge.yaml

- name: jobs_check_cuda_available
  group: Jobs tests
  team: serve

  frequency: nightly
  working_dir: jobs_tests
  cluster:
    byod:
      type: gpu
    cluster_compute: compute_tpl_gpu_node.yaml
  run:
    timeout: 600
    script: python workloads/jobs_check_cuda_available.py
    wait_for_nodes:
      num_nodes: 2

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_tpl_gce_gpu_node.yaml

- name: jobs_specify_num_gpus
  group: Jobs tests
  team: serve

  frequency: nightly
  working_dir: jobs_tests
  cluster:
    byod:
      type: gpu
    cluster_compute: compute_tpl_gpu_worker.yaml
  run:
    timeout: 600
    script: python workloads/jobs_specify_num_gpus.py --working-dir "workloads"
    wait_for_nodes:
      num_nodes: 2

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_tpl_gce_gpu_worker.yaml

########################
# Runtime env tests
########################
- name: runtime_env_rte_many_tasks_actors
  group: Runtime env tests
  working_dir: runtime_env_tests

  frequency: nightly
  team: core

  cluster:
    byod: {}
    cluster_compute: rte_small.yaml

  run:
    timeout: 600
    script: python workloads/rte_many_tasks_actors.py
    wait_for_nodes:
      num_nodes: 4


  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: rte_gce_small.yaml


- name: runtime_env_wheel_urls
  group: Runtime env tests
  working_dir: runtime_env_tests

  frequency: nightly
  team: core

  cluster:
    byod: {}
    cluster_compute: rte_minimal.yaml

  run:
    timeout: 9000
    script: python workloads/wheel_urls.py
    wait_for_nodes:
      num_nodes: 1


  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: rte_gce_minimal.yaml

# It seems like the consensus is that this should be tested in CI, and not in a nightly test.

# - name: runtime_env_rte_ray_client
#   group: Runtime env tests
#   working_dir: runtime_env_tests

#   frequency: nightly
#   team: core

#   cluster:
#     cluster_compute: rte_minimal.yaml

#   run:
#     timeout: 600
#     script: python workloads/rte_ray_client.py
#     wait_for_nodes:
#       num_nodes: 1

#   alert: default


########################
# Serve tests
########################

- name: serve_scale_replicas
  group: Serve tests
  working_dir: serve_tests

  frequency: nightly
  team: serve

  cluster:
    byod: {}
    cluster_compute: compute_tpl_single_node_32_cpu.yaml
    cloud_id: cld_wy5a6nhazplvu32526ams61d98

  run:
    timeout: 7200
    long_running: false
    script: python workloads/replica_scalability.py

  alert: default

  variations:
    - __suffix__: aws

- name: serve_multi_deployment_1k_noop_replica
  group: Serve tests
  working_dir: serve_tests

  frequency: nightly
  team: serve

  cluster:
    byod: {}
    cluster_compute: compute_tpl_32_cpu.yaml
    cloud_id: cld_wy5a6nhazplvu32526ams61d98

  run:
    timeout: 7200
    long_running: false
    script: python workloads/multi_deployment_1k_noop_replica.py

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: aws.py312
      python: "3.12"
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_tpl_32_cpu_gce.yaml

- name: serve_autoscaling_load_test
  group: Serve tests
  working_dir: serve_tests

  frequency: nightly
  team: serve

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_tpl_single_node_32_cpu.yaml
    cloud_id: cld_wy5a6nhazplvu32526ams61d98

  run:
    timeout: 7200
    long_running: false
    script: python workloads/autoscaling_load_test.py

  alert: default

  variations:
    - __suffix__: aws

- name: serve_microbenchmarks
  group: Serve tests
  working_dir: serve_tests

  frequency: nightly
  team: serve

  cluster:
    byod: {}
    cluster_compute: compute_tpl_single_node_32_cpu.yaml
    cloud_id: cld_wy5a6nhazplvu32526ams61d98

  run:
    timeout: 7200
    long_running: false
    script: python workloads/microbenchmarks.py --run-all

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_tpl_single_node_gce.yaml

- name: serve_resnet_benchmark
  group: Serve tests
  working_dir: serve_tests

  frequency: nightly
  team: serve

  cluster:
    byod:
      type: gpu
    cluster_compute: compute_tpl_gpu_node.yaml
    cloud_id: cld_wy5a6nhazplvu32526ams61d98

  run:
    timeout: 7200
    long_running: false
    script: python workloads/serve_resnet_benchmark.py --gpu-env

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_tpl_gpu_node_gce.yaml

########################
# Train tests
########################

- name: train_horovod_multi_node_test
  group: Train tests
  working_dir: train_tests/horovod

  frequency: nightly
  team: ml

  cluster:
    byod:
      type: gpu
      post_build_script: byod_horovod_test.sh
    cluster_compute: compute_tpl_aws.yaml

  run:
    timeout: 3000
    script: python train_horovod_multi_node_test.py

    wait_for_nodes:
      num_nodes: 2

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_tpl_gce.yaml

  alert: default


- name: train_multinode_persistence
  group: Train tests
  working_dir: train_tests/multinode_persistence

  frequency: nightly
  team: ml

  cluster:
    byod:
      post_build_script: byod_train_persistence_test.sh
    cluster_compute: compute_aws.yaml

  run:
    timeout: 3000
    script: pytest -v test_persistence.py -s

    wait_for_nodes:
      num_nodes: 4

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gce.yaml

  alert: default

- name: train_colocate_trainer
  group: Train tests
  working_dir: train_tests/colocate_trainer

  frequency: nightly
  team: ml

  cluster:
    byod: {}
    cluster_compute: compute_aws.yaml

  run:
    timeout: 3000
    script: pytest -v test_colocate_trainer.py -s

    wait_for_nodes:
      num_nodes: 4

  alert: default


- name: xgboost_train_batch_inference_benchmark_10G
  group: Train tests
  working_dir: train_tests/xgboost_lightgbm

  frequency: nightly
  team: ml
  cluster:
    byod:
      type: gpu
    cluster_compute: compute_aws_1worker.yaml

  run:
    timeout: 36000
    script: python train_batch_inference_benchmark.py "xgboost" --size=10G

    wait_for_nodes:
      num_nodes: 2

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gce_1worker.yaml

  smoke_test:
    frequency: manual

    run:
      timeout: 1800

  alert: default

- name: xgboost_train_batch_inference_benchmark_100G
  group: Train tests
  working_dir: train_tests/xgboost_lightgbm

  frequency: nightly-3x
  team: ml
  cluster:
    byod:
      type: gpu
    cluster_compute: compute_aws_10workers.yaml

  run:
    timeout: 36000
    script: python train_batch_inference_benchmark.py "xgboost" --size=100G

    wait_for_nodes:
      num_nodes: 11

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gce_10workers.yaml

  smoke_test:
    frequency: manual

    run:
      timeout: 1800

  alert: default


- name: lightgbm_train_batch_inference_benchmark_10G
  group: Train tests
  working_dir: train_tests/xgboost_lightgbm

  frequency: nightly
  team: ml
  cluster:
    byod:
      type: gpu
    cluster_compute: compute_aws_1worker.yaml

  run:
    timeout: 36000
    script: python train_batch_inference_benchmark.py "lightgbm" --size=10G

    wait_for_nodes:
      num_nodes: 2

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gce_1worker.yaml

  smoke_test:
    frequency: manual

    run:
      timeout: 1800

  alert: default


- name: lightgbm_train_batch_inference_benchmark_100G
  group: Train tests
  working_dir: train_tests/xgboost_lightgbm

  frequency: nightly-3x
  team: ml
  cluster:
    byod:
      type: gpu
    cluster_compute: compute_aws_10workers.yaml

  run:
    timeout: 36000
    script: python train_batch_inference_benchmark.py "lightgbm" --size=100G

    wait_for_nodes:
      num_nodes: 11

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gce_10workers.yaml

  smoke_test:
    frequency: manual

    run:
      timeout: 1800

  alert: default


########################
# RLlib tests
########################

# ----------------------------------------------------------
# Checkpointing with RLModule and Learner APIs
# ----------------------------------------------------------

- name: rllib_learner_group_checkpointing_multinode
  group: RLlib tests
  working_dir: rllib_tests

  frequency: nightly
  team: rllib
  stable: False

  cluster:
    byod:
      type: gpu
      post_build_script: byod_rllib_test.sh
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
        - LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/ray/.mujoco/mujoco210/bin
    cluster_compute: multi_node_checkpointing_compute_config.yaml

  run:
    timeout: 3600
    script: pytest checkpointing_tests/test_learner_group_checkpointing.py

    wait_for_nodes:
      num_nodes: 2

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: multi_node_checkpointing_compute_config_gce.yaml

- name: rllib_learner_e2e_module_loading
  group: RLlib tests
  working_dir: rllib_tests
  stable: false

  frequency: nightly
  team: rllib

  cluster:
    byod:
      type: gpu
      post_build_script: byod_rllib_test.sh
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
        - LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/ray/.mujoco/mujoco210/bin
    cluster_compute: multi_node_checkpointing_compute_config.yaml

  run:
    timeout: 3600
    script: pytest checkpointing_tests/test_e2e_rl_module_restore.py

    wait_for_nodes:
      num_nodes: 2

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: multi_node_checkpointing_compute_config_gce.yaml


# ----------------------------------------------------------
# Learning and benchmarking tests
# ----------------------------------------------------------

# --------------------------
# DreamerV3
# --------------------------
# TODO (sven): Move algo and this test to pytorch
- name: rllib_learning_tests_pong_dreamerv3_tf2
  group: RLlib tests
  working_dir: rllib_tests

  stable: false

  frequency: weekly
  team: rllib
  cluster:
    byod:
      type: gpu
      post_build_script: byod_rllib_dreamerv3_test.sh
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
        - LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/ray/.mujoco/mujoco210/bin
    cluster_compute: 1gpu_4cpus.yaml

  run:
    timeout: 43200  # 12h
    script: python learning_tests/tuned_examples/dreamerv3/atari_100k.py --framework=tf2 --env=ALE/Pong-v5 --num-gpus=1 --stop-reward=15.0 --as-release-test

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: 1gpu_4cpus_gce.yaml


# --------------------------
# PPO
# --------------------------
- name: rllib_learning_tests_pong_ppo_torch
  group: RLlib tests
  working_dir: rllib_tests

  stable: true

  frequency: nightly
  team: rllib
  cluster:
    byod:
      type: gpu
      post_build_script: byod_rllib_test.sh
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
        - LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/ray/.mujoco/mujoco210/bin
    cluster_compute: 8gpus_96cpus.yaml

  run:
    timeout: 1200
    script: python learning_tests/tuned_examples/ppo/atari_ppo.py --enable-new-api-stack --env=ALE/Pong-v5 --num-gpus=4 --num-env-runners=95 --stop-reward=20.0 --as-release-test

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: 8gpus_96cpus_gce.yaml


# --------------------------
# SAC
# --------------------------
- name: rllib_learning_tests_halfcheetah_sac_torch
  group: RLlib tests
  working_dir: rllib_tests

  stable: true

  frequency: nightly
  team: rllib
  cluster:
    byod:
      type: gpu
      post_build_script: byod_rllib_test.sh
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
        - LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/ray/.mujoco/mujoco210/bin
    cluster_compute: 4gpus_64cpus.yaml

  run:
    timeout: 7200
    script: python learning_tests/tuned_examples/sac/halfcheetah_sac.py --enable-new-api-stack --num-gpus=4 --num-env-runners=8 --stop-reward=1000.0 --as-release-test

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: 4gpus_64cpus_gce.yaml


########################
# Core Nightly Tests
########################

- name: shuffle_100gb
  group: core-multi-test
  working_dir: nightly_tests

  frequency: nightly

  team: core
  cluster:
    byod:
      runtime_env:
        - RAY_worker_killing_policy=retriable_lifo
    cluster_compute: shuffle/shuffle_compute_multi.yaml

  run:
    timeout: 3000
    script: python shuffle/shuffle_test.py --num-partitions=200 --partition-size=500e6
    wait_for_nodes:
      num_nodes: 4

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: shuffle/shuffle_compute_multi_gce.yaml


- name: stress_test_placement_group
  group: core-multi-test
  working_dir: nightly_tests
  env: aws_perf

  frequency: nightly

  team: core
  cluster:
    byod: {}
    cluster_compute: stress_tests/placement_group_tests_compute.yaml

  run:
    timeout: 7200
    script: python stress_tests/test_placement_group.py

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: stress_tests/placement_group_tests_compute_gce.yaml

- name: decision_tree_autoscaling_20_runs
  group: core-multi-test
  working_dir: nightly_tests

  frequency: nightly

  team: core
  cluster:
    byod: {}
    cluster_compute: decision_tree/autoscaling_compute.yaml

  run:
    timeout: 9600
    script: python decision_tree/cart_with_tree.py --concurrency=20

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: decision_tree/autoscaling_compute_gce.yaml

- name: autoscaling_shuffle_1tb_1000_partitions
  group: core-multi-test
  working_dir: nightly_tests

  frequency: nightly

  team: core
  cluster:
    byod:
      runtime_env:
        - RAY_worker_killing_policy=retriable_lifo
    cluster_compute: shuffle/shuffle_compute_autoscaling.yaml

  run:
    timeout: 4000
    script: python shuffle/shuffle_test.py --num-partitions=1000 --partition-size=1e9
      --no-streaming

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: shuffle/shuffle_compute_autoscaling_gce.yaml

- name: microbenchmark
  group: core-daily-test
  team: core
  frequency: nightly
  env: aws_perf
  working_dir: microbenchmark

  cluster:
    byod: {}
    cluster_compute: tpl_64.yaml

  run:
    timeout: 1800
    script: OMP_NUM_THREADS=64 RAY_ADDRESS=local python run_microbenchmark.py

  variations:
    - __suffix__: aws
      repeated_run: 5
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: tpl_64_gce.yaml
    - __suffix__: aws.py312
      frequency: weekly
      python: "3.12"

- name: microbenchmark_unstable
  group: core-daily-test
  team: core
  frequency: nightly
  working_dir: microbenchmark

  stable: false

  cluster:
    byod: {}
    cluster_compute: tpl_64.yaml

  run:
    timeout: 1800
    script: OMP_NUM_THREADS=64 RAY_ADDRESS=local python run_microbenchmark.py --experimental

- name: microbenchmark_gpu_unstable
  group: core-daily-test
  team: core
  frequency: nightly
  working_dir: microbenchmark

  stable: false

  cluster:
    byod:
      type: gpu
    cluster_compute: experimental/compute_gpu_2_aws.yaml

  run:
    timeout: 1800
    script: python experimental/accelerated_dag_gpu_microbenchmark.py

- name: microbenchmark_gpu_multinode_unstable
  group: core-daily-test
  team: core
  frequency: nightly
  working_dir: microbenchmark

  stable: false

  cluster:
    byod:
      type: gpu
    cluster_compute: experimental/compute_gpu_2x1_aws.yaml

  run:
    timeout: 1800
    script: python experimental/accelerated_dag_gpu_microbenchmark.py --distributed

- name: benchmark_worker_startup
  group: core-daily-test
  team: core
  frequency: nightly
  working_dir: benchmark-worker-startup
  stable: false


  cluster:
    byod:
      type: gpu
    cluster_compute: only_head_node_1gpu_64cpu.yaml

  run:
    timeout: 7200
    script: python benchmark_worker_startup.py
        --num_cpus_in_cluster 64
        --num_gpus_in_cluster 64
        --num_tasks_or_actors_per_run 64
        --num_measurements_per_configuration 5

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: only_head_node_1gpu_64cpu_gce.yaml

- name: dask_on_ray_100gb_sort
  group: core-daily-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  # https://github.com/ray-project/ray/issues/39165
  stable: false

  cluster:
    byod:
      runtime_env:
        - RAY_worker_killing_policy=retriable_lifo
    cluster_compute: dask_on_ray/dask_on_ray_sort_compute_template.yaml

  run:
    timeout: 7200
    script: python dask_on_ray/dask_on_ray_sort.py --nbytes 100_000_000_000 --npartitions
      200 --num-nodes 1 --ray --data-dir /tmp/ray --file-path /tmp/ray

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: dask_on_ray/dask_on_ray_sort_compute_template_gce.yaml


- name: dask_on_ray_large_scale_test_spilling
  group: core-daily-test
  working_dir: nightly_tests

  frequency: nightly
  team: data

  cluster:
    byod:
      runtime_env:
        - RAY_worker_killing_policy=retriable_lifo
    cluster_compute: dask_on_ray/dask_on_ray_stress_compute.yaml

  run:
    timeout: 7200
    script: python dask_on_ray/large_scale_test.py --num_workers 150 --worker_obj_store_size_in_gb
      70 --error_rate 0  --data_save_path /tmp/ray

    wait_for_nodes:
      num_nodes: 21


  smoke_test:
    frequency: nightly
    cluster:
      cluster_compute: dask_on_ray/large_scale_dask_on_ray_compute_template.yaml

    run:
      timeout: 7200
      script: python dask_on_ray/large_scale_test.py --num_workers 32 --worker_obj_store_size_in_gb
        70 --error_rate 0  --data_save_path /tmp/ray

      wait_for_nodes:
        num_nodes: 5

- name: stress_test_state_api_scale
  group: core-daily-test
  working_dir: nightly_tests

  stable: false

  frequency: nightly
  team: core
  cluster:
    byod:
      runtime_env:
        - RAY_MAX_LIMIT_FROM_API_SERVER=1000000000
        - RAY_MAX_LIMIT_FROM_DATA_SOURCE=1000000000
    cluster_compute: stress_tests/stress_tests_compute_large.yaml

  run:
    timeout: 4200
    script: python stress_tests/test_state_api_scale.py
    wait_for_nodes:
      num_nodes: 7

  smoke_test:
    frequency: nightly
    cluster:
      app_config: stress_tests/state_api_app_config.yaml
      cluster_compute: stress_tests/smoke_test_compute.yaml

    run:
      timeout: 3600
      wait_for_nodes:
        num_nodes: 5
      script: python stress_tests/test_state_api_scale.py --smoke-test

  variations:
    - __suffix__: aws
    - __suffix__: aws.py312
      frequency: manual
      python: "3.12"
      smoke_test:
        frequency: nightly-3x
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: stress_tests/stress_tests_compute_large_gce.yaml
      smoke_test:
        frequency: manual


- name: shuffle_20gb_with_state_api
  group: core-daily-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    byod:
      runtime_env:
        - RAY_MAX_LIMIT_FROM_API_SERVER=1000000000
        - RAY_MAX_LIMIT_FROM_DATA_SOURCE=1000000000
    cluster_compute: shuffle/shuffle_compute_single.yaml

  run:
    timeout: 1000
    script: python stress_tests/test_state_api_with_other_tests.py
      nightly_tests/shuffle/shuffle_test.py --test-args="--num-partitions=100 --partition-size=200e6"

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: shuffle/shuffle_compute_single_gce.yaml

- name: stress_test_many_tasks
  group: core-daily-test
  working_dir: nightly_tests
  env: aws_perf

  frequency: nightly
  team: core
  cluster:
    byod: {}
    cluster_compute: stress_tests/stress_tests_compute.yaml

  run:
    timeout: 14400
    wait_for_nodes:
      num_nodes: 101

    script: python stress_tests/test_many_tasks.py

  smoke_test:
    frequency: nightly
    cluster:
      app_config: stress_tests/stress_tests_app_config.yaml
      cluster_compute: stress_tests/smoke_test_compute.yaml

    run:
      timeout: 3600
      wait_for_nodes:
        num_nodes: 5
      script: python stress_tests/test_many_tasks.py --num-nodes=4 --smoke-test

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: stress_tests/stress_tests_compute_gce.yaml
      smoke_test:
        frequency: manual

- name: stress_test_dead_actors
  group: core-daily-test
  working_dir: nightly_tests
  env: aws_perf

  frequency: nightly
  team: core
  cluster:
    byod: {}
    cluster_compute: stress_tests/stress_tests_compute.yaml

  run:
    timeout: 7200
    wait_for_nodes:
      num_nodes: 101

    script: python stress_tests/test_dead_actors.py

  smoke_test:
    frequency: nightly
    cluster:
      app_config: stress_tests/stress_tests_app_config.yaml
      cluster_compute: stress_tests/smoke_test_compute.yaml

    run:
      timeout: 3600
      wait_for_nodes:
        num_nodes: 5
      script: python stress_tests/test_dead_actors.py --num-nodes=4 --num-parents=3
        --num-children=3

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: stress_tests/stress_tests_compute_gce.yaml
      smoke_test:
        frequency: manual

# The full test is not stable, so run the smoke test only.
# See https://github.com/ray-project/ray/issues/23244.
- name: threaded_actors_stress_test
  group: core-daily-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    byod: {}
    cluster_compute: stress_tests/smoke_test_compute.yaml

  run:
    timeout: 3600
    script: python stress_tests/test_threaded_actors.py --test-runtime 1800 --kill-interval_s
      30

    wait_for_nodes:
      num_nodes: 5

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: stress_tests/smoke_test_compute_gce.yaml

# - name: threaded_actors_stress_test
#   group: core-daily-test
#   working_dir: nightly_tests
#
#   frequency: nightly
#   team: core
#   cluster:
#     cluster_compute: stress_tests/stress_test_threaded_actor_compute.yaml
#
#   run:
#     timeout: 7200
#     script: python stress_tests/test_threaded_actors.py --test-runtime 3600 --kill-interval_s
#       60
#
#     wait_for_nodes:
#       num_nodes: 201
#       timeout: 600
#
#   smoke_test:
#     frequency: nightly
#     cluster:
#       app_config: stress_tests/stress_tests_app_config.yaml
#       cluster_compute: stress_tests/smoke_test_compute.yaml
#
#     run:
#       timeout: 3600
#       script: python stress_tests/test_threaded_actors.py --test-runtime 1800 --kill-interval_s
#         30
#
#       wait_for_nodes:
#         num_nodes: 5
#         timeout: 600

- name: stress_test_many_runtime_envs
  group: core-daily-test
  working_dir: nightly_tests

  frequency: nightly
  team: core

  cluster:
    byod: {}
    cluster_compute: stress_tests/smoke_test_compute.yaml

  run:
    timeout: 14400
    wait_for_nodes:
      num_nodes: 5
    script: python stress_tests/test_many_runtime_envs.py --num_runtime_envs=100 --num_tasks=10000

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: stress_tests/smoke_test_compute_gce.yaml
      smoke_test:
        frequency: manual

- name: single_node_oom
  group: core-daily-test
  working_dir: nightly_tests

  # TODO: https://github.com/ray-project/ray/issues/47596
  stable: false

  frequency: nightly
  team: core
  env: aws_perf
  cluster:
    byod: {}
    cluster_compute: stress_tests/stress_tests_single_node_oom_compute.yaml

  run:
    timeout: 1000
    script: python stress_tests/test_parallel_tasks_memory_pressure.py --num-tasks 20

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: stress_tests/stress_tests_single_node_oom_compute_gce.yaml


- name: tune_air_oom
  group: core-daily-test
  working_dir: air_tests

  stable: false

  frequency: nightly
  team: core

  cluster:
    byod:
      runtime_env:
        - RAY_memory_usage_threshold=0.7
        - RAY_task_oom_retries=-1
    cluster_compute: oom/stress_tests_tune_air_oom_compute.yaml

  run:
    timeout: 3600
    script: bash oom/tune_air_oom.sh


- name: dask_on_ray_1tb_sort
  group: core-daily-test
  working_dir: nightly_tests

  frequency: nightly-3x
  team: core

  cluster:
    byod:
      runtime_env:
        - RAY_worker_killing_policy=retriable_lifo
    cluster_compute: dask_on_ray/1tb_sort_compute.yaml

  run:
    timeout: 7200
    script: python dask_on_ray/dask_on_ray_sort.py --nbytes 1_000_000_000_000 --npartitions
      1000 --num-nodes 31 --ray --data-dir /tmp/ray --s3-bucket core-nightly-test

    wait_for_nodes:
      num_nodes: 32


- name: many_nodes_actor_test_on_v2
  group: core-daily-test
  working_dir: benchmarks

  frequency: nightly-3x
  team: core
  cluster:
    byod: {}
    cluster_compute: distributed/many_nodes_tests/compute_config.yaml

  run:
    timeout: 3600
    # 2cpus per node x 1000 nodes / 0.2 cpus per actor = 10k
    # 2cpus per node x 2000 nodes / 0.2 cpus per actor = 20k
    script: python distributed/many_nodes_tests/actor_test.py --no-wait --cpus-per-actor=0.2 --total-actors 10000 20000
    wait_for_nodes:
      num_nodes: 500

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: distributed/many_nodes_tests/compute_config_gce.yaml

#- name: many_nodes_multi_master_test
#  group: core-daily-test
#  working_dir: nightly_tests
#
#  frequency: nightly-3x
#  team: core
#  cluster:
#    cluster_compute: many_nodes_tests/compute_config.yaml
#
#  run:
#    timeout: 7200
#    script: python many_nodes_tests/multi_master_test.py
#    wait_for_nodes:
#      num_nodes: 251
#

- name: pg_autoscaling_regression_test
  group: core-daily-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    byod: {}
    cluster_compute: placement_group_tests/compute.yaml

  run:
    timeout: 1200
    script: python placement_group_tests/pg_run.py

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: placement_group_tests/compute_gce.yaml

- name: placement_group_performance_test
  group: core-daily-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    byod: {}
    cluster_compute: placement_group_tests/pg_perf_test_compute.yaml

  run:
    timeout: 1200
    script: python placement_group_tests/placement_group_performance_test.py
    wait_for_nodes:
      num_nodes: 5

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: placement_group_tests/pg_perf_test_compute_gce.yaml


#########################
# Core Scalability Tests
#########################

- name: single_node
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly
  team: core
  env: aws_perf
  cluster:
    byod:
      type: gpu
      runtime_env:
        - LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so
    cluster_compute: single_node.yaml

  run:
    timeout: 12000
    prepare: sleep 0
    script: python single_node/test_single_node.py

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: single_node_gce.yaml

- name: object_store
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly
  team: core
  env: aws_perf
  cluster:
    byod:
      type: gpu
      runtime_env:
        - LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so
    cluster_compute: object_store.yaml

  run:
    timeout: 3600
    script: python object_store/test_object_store.py
    wait_for_nodes:
      num_nodes: 50

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: object_store_gce.yaml

- name: many_actors
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly-3x
  team: core
  env: aws_perf
  cluster:
    byod:
      type: gpu
      runtime_env:
        - LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so
    cluster_compute: distributed.yaml

  run:
    timeout: 3600
    script: python distributed/test_many_actors.py
    wait_for_nodes:
      num_nodes: 65

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: distributed_gce.yaml

- name: many_actors_smoke_test
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly
  team: core
  cluster:
    byod:
      type: gpu
      runtime_env:
        - LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so
    cluster_compute: distributed_smoke_test.yaml

  run:
    timeout: 3600
    script: SMOKE_TEST=1 python distributed/test_many_actors.py
    wait_for_nodes:
      num_nodes: 2


- name: many_tasks
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly
  team: core
  env: aws_perf
  cluster:
    byod:
      type: gpu
      runtime_env:
        - LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so
    cluster_compute: distributed.yaml

  run:
    timeout: 3600
    script: python distributed/test_many_tasks.py --num-tasks=10000
    wait_for_nodes:
      num_nodes: 65

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: distributed_gce.yaml

- name: many_pgs
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly-3x
  team: core
  env: aws_perf
  cluster:
    byod:
      type: gpu
      runtime_env:
        - LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so
    cluster_compute: distributed.yaml

  run:
    timeout: 3600
    script: python distributed/test_many_pgs.py
    wait_for_nodes:
      num_nodes: 65

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: distributed_gce.yaml


- name: many_pgs_smoke_test
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly
  team: core
  cluster:
    byod:
      type: gpu
      runtime_env:
        - LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so
    cluster_compute: distributed_smoke_test.yaml

  run:
    timeout: 3600
    script: SMOKE_TEST=1 python distributed/test_many_pgs.py
    wait_for_nodes:
      num_nodes: 2


- name: many_nodes
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly-3x
  team: core
  env: aws_perf
  cluster:
    byod:
      type: gpu
      runtime_env:
        - LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so
    cluster_compute: many_nodes.yaml

  run:
    timeout: 3600
    script: python distributed/test_many_tasks.py --num-tasks=1000
    wait_for_nodes:
      num_nodes: 250

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: many_nodes_gce.yaml

- name: scheduling_test_many_0s_tasks_many_nodes
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly
  team: core
  cluster:
    byod:
      type: gpu
      runtime_env:
        - LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so
    cluster_compute: scheduling.yaml

  run:
    timeout: 3600
    script: python distributed/test_scheduling.py --total-num-task=1984000 --num-cpu-per-task=1
      --task-duration-s=0 --total-num-actors=32 --num-actors-per-nodes=1

    wait_for_nodes:
      num_nodes: 32

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: scheduling_gce.yaml


# - name: scheduling_test_many_5s_tasks_single_node
#   group: core-scalability-test
#   working_dir: benchmarks

#   frequency: nightly
#   team: core
#   cluster:
#     cluster_compute: scheduling.yaml

#   run:
#     timeout: 3600
#     script: python distributed/test_scheduling.py --total-num-task=1984000 --num-cpu-per-task=1
#       --task-duration-s=5 --total-num-actors=1 --num-actors-per-nodes=1

#     wait_for_nodes:
#       num_nodes: 32
#       timeout: 600

#   stable: false

# - name: scheduling_test_many_5s_tasks_many_nodes
#   group: core-scalability-test
#   working_dir: benchmarks

#   frequency: nightly
#   team: core
#   cluster:
#     cluster_compute: scheduling.yaml

#   run:
#     timeout: 3600
#     script: python distributed/test_scheduling.py --total-num-task=1984000 --num-cpu-per-task=1
#       --task-duration-s=5 --total-num-actors=32 --num-actors-per-nodes=1

#     wait_for_nodes:
#       num_nodes: 32
#       timeout: 600

#   stable: false

###############
# Dataset tests
###############

- name: parquet_metadata_resolution
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data

  cluster:
    byod:
      type: gpu
    cluster_compute: single_node_benchmark_compute.yaml

  run:
    # Expect the test to finish around 40 seconds.
    timeout: 100
    script: python parquet_metadata_resolution.py --num-files 915 --cloud aws

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: single_node_benchmark_compute_gce.yaml
      run:
        script: python parquet_metadata_resolution.py --num-files 915 --cloud gcp

- name: dataset_random_access
  group: data-tests
  working_dir: nightly_tests/dataset
  stable: false

  frequency: manual
  team: data

  cluster:
    byod:
      type: gpu
      pip:
        - git+https://github.com/ray-project/ray_shuffling_data_loader.git@add-embedding-model
    cluster_compute: pipelined_training_compute.yaml

  run:
    timeout: 1200
    script: python dataset_random_access.py
    wait_for_nodes:
      num_nodes: 15

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: pipelined_training_compute_gce.yaml

- name: stable_diffusion_benchmark
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data

  cluster:
    byod:
      type: gpu
      post_build_script: byod_stable_diffusion.sh
    cluster_compute: stable_diffusion_benchmark_compute.yaml

  run:
    timeout: 1800
    script: python stable_diffusion_benchmark.py

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: stable_diffusion_benchmark_compute_gce.yaml

- name: streaming_data_ingest_benchmark_1tb
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data

  cluster:
    byod:
      type: gpu
    cluster_compute: data_ingest_benchmark_compute.yaml

  run:
    timeout: 300
    script: python data_ingest_benchmark.py --dataset-size-gb=1000 --num-workers=20 --streaming
    wait_for_nodes:
      num_nodes: 20

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: data_ingest_benchmark_compute_gce.yaml

- name: streaming_data_ingest_benchmark_100gb_gpu
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data

  cluster:
    byod:
      type: gpu
    cluster_compute: data_ingest_benchmark_compute_gpu.yaml

  run:
    timeout: 300
    script: python data_ingest_benchmark.py --dataset-size-gb=100 --num-workers=4 --streaming --use-gpu
    wait_for_nodes:
      num_nodes: 3

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: data_ingest_benchmark_compute_gpu_gce.yaml

# This test case will early stop the data ingestion iteration on the GPU actors.
# This is a common usage in PyTorch Lightning
# (https://lightning.ai/docs/pytorch/stable/common/trainer.html#limit-train-batches).
# There was a bug in Ray Data that caused GPU memoy leak (see #3.919).
# We add this test case to cover this scenario.
- name: streaming_data_ingest_benchmark_100gb_gpu_early_stop
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data

  cluster:
    byod:
      type: gpu
    cluster_compute: data_ingest_benchmark_compute_gpu.yaml

  run:
    timeout: 300
    script: python data_ingest_benchmark.py --dataset-size-gb=100 --num-workers=4 --streaming --use-gpu --early-stop
    wait_for_nodes:
      num_nodes: 3

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: data_ingest_benchmark_compute_gpu_gce.yaml

- name: aggregate_benchmark
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data

  cluster:
    byod:
      type: gpu
    cluster_compute: single_node_benchmark_compute.yaml

  run:
    timeout: 1800
    script: python aggregate_benchmark.py

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: single_node_benchmark_compute_gce.yaml

- name: read_parquet_benchmark_single_node
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data

  cluster:
    byod:
      type: gpu
      post_build_script: byod_install_mosaicml.sh
    cluster_compute: single_node_benchmark_compute.yaml

  run:
    # Expect the benchmark to finish in 400 seconds.
    timeout: 400
    script: python read_parquet_benchmark.py

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: single_node_benchmark_compute_gce.yaml

- name: read_images_benchmark_single_node
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data

  cluster:
    byod:
      type: gpu
      post_build_script: byod_install_mosaicml.sh
    cluster_compute: single_node_benchmark_compute.yaml

  run:
    timeout: 1800
    script: python read_images_benchmark.py --single-node

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: single_node_benchmark_compute_gce.yaml

# TODO: Re-enable this test once we fix https://github.com/ray-project/ray/issues/40686.
# - name: read_images_benchmark_multi_node
#   group: data-tests
#   working_dir: nightly_tests/dataset

#   frequency: nightly-3x
#   team: data

#   cluster:
#     byod:
#       type: gpu
#     cluster_compute: multi_node_read_images_benchmark_compute.yaml

#   run:
#     timeout: 28800
#     script: python read_images_benchmark.py --multi-node

#   variations:
#     - __suffix__: aws
#     - __suffix__: gce
#       env: gce
#       frequency: manual
#       cluster:
#         cluster_compute: multi_node_read_images_benchmark_compute_gce.yaml

- name: read_images_comparison_microbenchmark_single_node
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data
  cluster:
    byod:
      type: gpu
      post_build_script: byod_install_mosaicml.sh
    cluster_compute: single_worker_node_0_head_node_benchmark_compute.yaml

  run:
    timeout: 1800
    script: bash run_image_loader_microbenchmark.sh

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: single_node_benchmark_compute_gce.yaml

- name: read_images_train_4_gpu
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data
  cluster:
    byod:
      type: gpu
      post_build_script: byod_install_mosaicml.sh
    cluster_compute: multi_node_train_4_workers.yaml

  run:
    timeout: 18000
    script: python multi_node_train_benchmark.py --num-workers 4 --file-type image --use-gpu --num-epochs 2

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: ../../air_tests/air_benchmarks/compute_gpu_2x2_gce.yaml

- name: read_images_train_4_gpu_worker_chaos
  group: data-tests
  working_dir: nightly_tests

  frequency: nightly
  team: data
  cluster:
    byod:
      type: gpu
      post_build_script: byod_install_mosaicml.sh
    cluster_compute: dataset/multi_node_train_4_workers.yaml

  run:
    timeout: 18000
    prepare: python setup_chaos.py --kill-workers --kill-interval 100 --max-to-kill 3 --task-names "ReadImage->Map(wnid_to_index)->Map(crop_and_flip_image)"
    script: python dataset/multi_node_train_benchmark.py --num-workers 4 --file-type image --use-gpu --num-epochs 1

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: ../air_tests/air_benchmarks/compute_gpu_2x2_gce.yaml

- name: read_images_train_4_gpu_node_chaos
  group: data-tests
  working_dir: nightly_tests

  frequency: nightly
  team: data
  cluster:
    byod:
      type: gpu
      post_build_script: byod_install_mosaicml.sh
    cluster_compute: dataset/multi_node_train_4_workers.yaml

  run:
    timeout: 18000
    prepare: python setup_chaos.py --kill-interval 200 --max-to-kill 1 --task-names "_RayTrainWorker__execute.get_next"
    script: python dataset/multi_node_train_benchmark.py --num-workers 4 --file-type image --use-gpu --num-epochs 1

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: ../air_tests/air_benchmarks/compute_gpu_2x2_gce.yaml

- name: read_images_train_16_gpu
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data
  cluster:
    byod:
      type: gpu
      post_build_script: byod_install_mosaicml.sh
    cluster_compute: multi_node_train_16_workers.yaml

  run:
    timeout: 18000
    script: python multi_node_train_benchmark.py --num-workers 16 --file-type image --use-gpu --num-epochs 2

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: ../../air_tests/air_benchmarks/compute_gpu_4x4_gce.yaml

- name: read_images_train_16_gpu_preserve_order
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data
  cluster:
    byod:
      type: gpu
      post_build_script: byod_install_mosaicml.sh
    cluster_compute: multi_node_train_16_workers.yaml

  run:
    timeout: 18000
    script: python multi_node_train_benchmark.py --num-workers 16 --file-type image --preserve-order --use-gpu --num-epochs 2

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: ../../air_tests/air_benchmarks/compute_gpu_4x4_gce.yaml

- name: read_parquet_train_4_gpu
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data
  cluster:
    byod:
      type: gpu
      post_build_script: byod_install_mosaicml.sh
    cluster_compute: multi_node_train_4_workers.yaml

  run:
    timeout: 3600
    script: python multi_node_train_benchmark.py --num-workers 4 --file-type parquet --target-worker-gb 50 --use-gpu

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: ../../air_tests/air_benchmarks/compute_gpu_2x2_gce.yaml

- name: read_parquet_train_16_gpu
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data
  cluster:
    byod:
      type: gpu
      post_build_script: byod_install_mosaicml.sh
    cluster_compute: multi_node_train_16_workers.yaml

  run:
    timeout: 3600
    script: python multi_node_train_benchmark.py --num-workers 16 --file-type parquet --target-worker-gb 50 --use-gpu

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: ../../air_tests/air_benchmarks/compute_gpu_4x4_gce.yaml

- name: read_images_train_1_gpu_5_cpu
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data
  cluster:
    byod:
      type: gpu
      post_build_script: byod_install_mosaicml.sh
    cluster_compute: multi_node_train_1g5c.yaml

  run:
    timeout: 2400
    script: python multi_node_train_benchmark.py --num-workers 1 --file-type image --use-gpu --num-epochs 2 --skip-train-model --prefetch-batches 16 --batch-size -1 --disable-locality-with-output

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gpu_1g5c_gce.yaml

- name: read_tfrecords_benchmark_single_node
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data

  cluster:
    byod:
      type: gpu
      post_build_script: byod_install_mosaicml.sh
    cluster_compute: single_node_benchmark_compute.yaml

  run:
    # Expect the benchmark to finish around 22 minutes.
    timeout: 1800
    script: python read_tfrecords_benchmark.py

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: single_node_benchmark_compute_gce.yaml

- name: map_batches_benchmark_single_node
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data

  cluster:
    byod:
      type: gpu
    cluster_compute: single_node_benchmark_compute.yaml

  run:
    # Expect the benchmark to finish around 30 minutes.
    timeout: 2400
    script: python map_batches_benchmark.py

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: single_node_benchmark_compute_gce.yaml

- name: iter_tensor_batches_benchmark_single_node
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data

  cluster:
    byod:
      type: gpu
    cluster_compute: single_node_benchmark_compute.yaml

  run:
    # Expect the benchmark to finish around 30 minutes.
    timeout: 2400
    script: python iter_tensor_batches_benchmark.py

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: single_node_benchmark_compute_gce.yaml

- name: iter_tensor_batches_benchmark_multi_node
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data

  cluster:
    byod:
      type: gpu
    cluster_compute: multi_node_benchmark_compute.yaml

  run:
    # Expect the benchmark to finish within 90 minutes.
    timeout: 5400
    script: python iter_tensor_batches_benchmark.py --data-size-gb=10

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: multi_node_benchmark_compute_gce.yaml

- name: iter_batches_benchmark_single_node
  group: data-tests
  working_dir: nightly_tests/dataset

  frequency: nightly
  team: data

  cluster:
    byod:
      type: gpu
    cluster_compute: single_node_benchmark_compute.yaml

  run:
    # Expect the benchmark to finish around 12 minutes.
    timeout: 1080
    script: python iter_batches_benchmark.py

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: single_node_benchmark_compute_gce.yaml

- name: dataset_shuffle_random_shuffle_1tb
  group: data-tests
  working_dir: nightly_tests

  frequency: nightly
  team: data

  cluster:
    byod:
      runtime_env:
        - RAY_worker_killing_policy=retriable_lifo
      pip:
        - ray[default]
    cluster_compute: shuffle/datasets_large_scale_compute_small_instances.yaml

  run:
    timeout: 7200
    script: python dataset/sort.py --num-partitions=1000 --partition-size=1e9 --shuffle
    wait_for_nodes:
      num_nodes: 20

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: shuffle/datasets_large_scale_compute_small_instances_gce.yaml

- name: dataset_shuffle_sort_1tb
  group: data-tests
  working_dir: nightly_tests

  frequency: nightly
  team: data
  stable: False

  cluster:
    byod:
      runtime_env:
        - RAY_worker_killing_policy=retriable_lifo
      pip:
        - ray[default]
    cluster_compute: shuffle/datasets_large_scale_compute_small_instances.yaml

  run:
    timeout: 7200
    script: python dataset/sort.py --num-partitions=1000 --partition-size=1e9
    wait_for_nodes:
      num_nodes: 20

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: shuffle/datasets_large_scale_compute_small_instances_gce.yaml

- name: dataset_shuffle_push_based_random_shuffle_1tb
  group: data-tests
  working_dir: nightly_tests

  stable: false

  frequency: nightly
  team: data

  cluster:
    byod:
      runtime_env:
        - RAY_worker_killing_policy=retriable_lifo
      pip:
        - ray[default]
    cluster_compute: shuffle/datasets_large_scale_compute_small_instances.yaml

  run:
    timeout: 7200
    script: RAY_DATA_PUSH_BASED_SHUFFLE=1 python dataset/sort.py --num-partitions=1000 --partition-size=1e9 --shuffle
    wait_for_nodes:
      num_nodes: 20

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: shuffle/datasets_large_scale_compute_small_instances_gce.yaml

- name: dataset_shuffle_push_based_sort_1tb
  group: data-tests
  working_dir: nightly_tests

  frequency: nightly
  team: data
  stable: False

  cluster:
    byod:
      runtime_env:
        - RAY_worker_killing_policy=retriable_lifo
      pip:
        - ray[default]
    cluster_compute: shuffle/datasets_large_scale_compute_small_instances.yaml

  run:
    timeout: 7200
    script: RAY_DATA_PUSH_BASED_SHUFFLE=1 python dataset/sort.py --num-partitions=1000 --partition-size=1e9
    wait_for_nodes:
      num_nodes: 20

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: shuffle/datasets_large_scale_compute_small_instances_gce.yaml

- name: dataset_shuffle_push_based_random_shuffle_100tb
  group: data-tests
  working_dir: nightly_tests
  stable: false

  frequency: weekly
  team: data
  cluster:
    byod:
      runtime_env:
        - RAY_object_spilling_config={"type":"filesystem","params":{"directory_path":["/tmp/data0","/tmp/data1"]}}
      post_build_script: byod_dataset_shuffle.sh
    cluster_compute: shuffle/100tb_shuffle_compute.yaml

  run:
    timeout: 28800
    script: RAY_DATA_PUSH_BASED_SHUFFLE=1 python dataset/sort.py --num-partitions=100000 --partition-size=1e9 --shuffle
    wait_for_nodes:
      num_nodes: 100

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: shuffle/100tb_shuffle_compute_gce.yaml
      run:
        timeout: 28800
        script: RAY_DATA_PUSH_BASED_SHUFFLE=1 python dataset/sort.py --num-partitions=40000 --partition-size=1e9 --shuffle
        wait_for_nodes:
          num_nodes: 100

##################
# Core Chaos tests
##################

- name: chaos_many_tasks_kill_raylet
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    byod: {}
    cluster_compute: chaos_test/compute_template.yaml

  run:
    timeout: 3600
    wait_for_nodes:
      num_nodes: 10
    prepare: python setup_chaos.py --no-start
    script: python chaos_test/test_chaos_basic.py --workload=tasks

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: chaos_test/compute_template_gce.yaml

- name: chaos_many_tasks_terminate_instance
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    byod: {}
    cluster_compute: chaos_test/compute_template.yaml

  run:
    timeout: 3600
    wait_for_nodes:
      num_nodes: 10
    prepare: python setup_chaos.py --no-start --chaos TerminateEC2Instance
    script: python chaos_test/test_chaos_basic.py --workload=tasks

  variations:
    - __suffix__: aws

- name: chaos_many_actors_kill_raylet
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    byod: {}
    cluster_compute: chaos_test/compute_template.yaml

  run:
    timeout: 4200
    wait_for_nodes:
      num_nodes: 10
    prepare: python setup_chaos.py --no-start
    script: python chaos_test/test_chaos_basic.py --workload=actors

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: chaos_test/compute_template_gce.yaml

- name: chaos_many_actors_terminate_instance
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    byod: {}
    cluster_compute: chaos_test/compute_template.yaml

  run:
    timeout: 4200
    wait_for_nodes:
      num_nodes: 10
    prepare: python setup_chaos.py --no-start --chaos TerminateEC2Instance
    script: python chaos_test/test_chaos_basic.py --workload=actors

  variations:
    - __suffix__: aws

- name: chaos_dask_on_ray_large_scale_test_no_spilling
  group: data-tests
  working_dir: nightly_tests

  frequency: nightly
  team: data

  cluster:
    byod:
      runtime_env:
        - RAY_lineage_pinning_enabled=1
    cluster_compute: dask_on_ray/chaos_dask_on_ray_stress_compute.yaml

  run:
    timeout: 7200
    wait_for_nodes:
      num_nodes: 21
    prepare: python setup_chaos.py --kill-interval 100
    script: python dask_on_ray/large_scale_test.py --num_workers 20 --worker_obj_store_size_in_gb
      20 --error_rate 0  --data_save_path /tmp/ray

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: dask_on_ray/dask_on_ray_stress_compute_gce.yaml

- name: chaos_dask_on_ray_large_scale_test_spilling
  group: data-tests
  working_dir: nightly_tests

  frequency: nightly
  team: data

  cluster:
    byod:
      runtime_env:
        - RAY_lineage_pinning_enabled=1
    cluster_compute: dask_on_ray/dask_on_ray_stress_compute.yaml

  run:
    timeout: 7200
    wait_for_nodes:
      num_nodes: 21
    prepare: python setup_chaos.py --kill-interval 100
    script: python dask_on_ray/large_scale_test.py --num_workers 150 --worker_obj_store_size_in_gb
      70 --error_rate 0  --data_save_path /tmp/ray

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: dask_on_ray/dask_on_ray_stress_compute_gce.yaml

- name: chaos_dataset_shuffle_push_based_sort_1tb
  group: data-tests
  working_dir: nightly_tests

  stable: false

  frequency: nightly
  team: data

  cluster:
    byod:
      runtime_env:
        - RAY_worker_killing_policy=retriable_lifo
      pip:
        - ray[default]
    cluster_compute: shuffle/datasets_large_scale_compute_small_instances.yaml

  run:
    timeout: 7200
    prepare: 'python setup_chaos.py --kill-interval 1200 --max-to-kill 3'
    script: RAY_DATA_PUSH_BASED_SHUFFLE=1 python dataset/sort.py --num-partitions=1000 --partition-size=1e9
    wait_for_nodes:
      num_nodes: 20

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: shuffle/datasets_large_scale_compute_small_instances_gce.yaml

- name: chaos_dataset_shuffle_sort_1tb
  group: data-tests
  working_dir: nightly_tests

  stable: false

  frequency: nightly
  team: data

  cluster:
    byod:
      runtime_env:
        - RAY_memory_monitor_refresh_ms=0
    cluster_compute: shuffle/datasets_large_scale_compute_small_instances.yaml

  run:
    timeout: 7200
    prepare: 'python setup_chaos.py --kill-interval 900 --max-to-kill 3'
    script: python dataset/sort.py --num-partitions=1000 --partition-size=1e9
    wait_for_nodes:
      num_nodes: 20

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: shuffle/datasets_large_scale_compute_small_instances_gce.yaml

- name: chaos_dataset_shuffle_random_shuffle_1tb
  group: data-tests
  working_dir: nightly_tests

  stable: false

  frequency: nightly
  team: data

  cluster:
    # leave oom disabled as test is marked unstable at the moment.
    byod:
      runtime_env:
        - RAY_memory_monitor_refresh_ms=0
      pip:
        - ray[default]
    cluster_compute: shuffle/datasets_large_scale_compute_small_instances.yaml

  run:
    timeout: 7200
    prepare: ' python setup_chaos.py --kill-interval 600 --max-to-kill 2'
    script: python dataset/sort.py --num-partitions=1000 --partition-size=1e9 --shuffle
    wait_for_nodes:
      num_nodes: 20

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: shuffle/datasets_large_scale_compute_small_instances_gce.yaml

- name: chaos_dataset_shuffle_push_based_random_shuffle_1tb
  group: data-tests
  working_dir: nightly_tests

  stable: false

  frequency: nightly
  team: data

  cluster:
    # leave oom disabled as test is marked unstable at the moment.
    byod:
      runtime_env:
        - RAY_memory_monitor_refresh_ms=0
    cluster_compute: shuffle/datasets_large_scale_compute_small_instances.yaml

  run:
    timeout: 7200
    prepare: ' python setup_chaos.py --kill-interval 600 --max-to-kill 2'
    script: RAY_DATA_PUSH_BASED_SHUFFLE=1 python dataset/sort.py --num-partitions=1000 --partition-size=1e9 --shuffle
    wait_for_nodes:
      num_nodes: 20

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: shuffle/datasets_large_scale_compute_small_instances_gce.yaml

#####################
# Observability tests
#####################
- name: agent_stress_test
  group: core-observability-test
  working_dir: dashboard

  frequency: nightly
  team: core
  cluster:
    byod:
      type: gpu
      runtime_env:
        - RAY_INTERNAL_MEM_PROFILE_COMPONENTS=dashboard_agent
      post_build_script: byod_agent_stress_test.sh
    cluster_compute: agent_stress_compute.yaml

  run:
    timeout: 14400
    script: python mem_check.py --working-dir .

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: agent_stress_compute_gce.yaml

- name: k8s_serve_ha_test
  group: k8s-test
  working_dir: k8s_tests

  stable: false

  frequency: nightly
  team: serve
  cluster:
    byod: {}
    cluster_compute: compute_tpl.yaml

  run:
    timeout: 28800 # 8h
    prepare: bash prepare.sh
    script: python run_gcs_ft_on_k8s.py

- name: aws_cluster_launcher
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  frequency: nightly
  team: clusters

  cluster:
    byod: {}
    cluster_compute: aws/tests/aws_compute.yaml

  run:
    timeout: 2400
    script: python launch_and_verify_cluster.py aws/tests/aws_cluster.yaml --num-expected-nodes 2 --retries 10


- name: aws_cluster_launcher_nightly_image
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  frequency: nightly
  team: clusters
  cluster:
    byod: {}
    cluster_compute: aws/tests/aws_compute.yaml

  run:
    timeout: 2400
    script: python launch_and_verify_cluster.py aws/tests/aws_cluster.yaml --num-expected-nodes 2 --retries 10 --docker-override nightly


- name: aws_cluster_launcher_latest_image
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  frequency: nightly
  team: clusters
  cluster:
    byod: {}
    cluster_compute: aws/tests/aws_compute.yaml

  run:
    timeout: 2400
    script: python launch_and_verify_cluster.py aws/tests/aws_cluster.yaml --num-expected-nodes 2 --retries 10 --docker-override latest


- name: aws_cluster_launcher_release_image
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  frequency: manual
  team: clusters
  cluster:
    byod: {}
    cluster_compute: aws/tests/aws_compute.yaml

  run:
    timeout: 2400
    script: python launch_and_verify_cluster.py aws/tests/aws_cluster.yaml --num-expected-nodes 2 --retries 10 --docker-override commit



- name: aws_cluster_launcher_minimal
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  frequency: nightly
  team: clusters
  cluster:
    byod: {}
    cluster_compute: aws/tests/aws_compute.yaml

  run:
    timeout: 1200
    script: python launch_and_verify_cluster.py aws/example-minimal.yaml

- name: aws_cluster_launcher_full
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  frequency: nightly
  team: clusters
  cluster:
    byod: {}
    cluster_compute: aws/tests/aws_compute.yaml

  run:
    timeout: 3000
    script: python launch_and_verify_cluster.py aws/example-full.yaml --num-expected-nodes 2 --retries 20 --docker-override latest

- name: gcp_cluster_launcher_minimal
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  stable: true

  env: gce
  frequency: nightly
  team: clusters
  cluster:
    byod: {}
    cluster_compute: gcp/tests/single_node_32_cpu_gce.yaml

  run:
    timeout: 1200
    script: python launch_and_verify_cluster.py gcp/example-minimal.yaml

- name: gcp_cluster_launcher_full
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  stable: true

  env: gce
  frequency: nightly
  team: clusters
  cluster:
    byod: {}
    cluster_compute: gcp/tests/single_node_32_cpu_gce.yaml

  run:
    timeout: 4800
    script: python launch_and_verify_cluster.py gcp/example-full.yaml --num-expected-nodes 2 --retries 30

- name: gcp_cluster_launcher_latest_image
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  stable: true

  env: gce
  frequency: nightly
  team: clusters
  cluster:
    byod: {}
    cluster_compute: gcp/tests/single_node_32_cpu_gce.yaml

  run:
    timeout: 3600
    script: python launch_and_verify_cluster.py gcp/example-full.yaml --num-expected-nodes 2 --retries 20  --docker-override latest

- name: gcp_cluster_launcher_nightly_image
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  stable: true

  env: gce
  frequency: nightly
  team: clusters
  cluster:
    byod: {}
    cluster_compute: gcp/tests/single_node_32_cpu_gce.yaml

  run:
    timeout: 3600
    script: python launch_and_verify_cluster.py gcp/example-full.yaml --num-expected-nodes 2 --retries 20  --docker-override nightly


- name: gcp_cluster_launcher_release_image
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  stable: true

  env: gce
  frequency: manual
  team: clusters
  cluster:
    byod: {}
    cluster_compute: gcp/tests/single_node_32_cpu_gce.yaml

  run:
    timeout: 3600
    script: python launch_and_verify_cluster.py gcp/example-full.yaml --num-expected-nodes 2 --retries 20  --docker-override commit

- name: gcp_cluster_launcher_gpu_docker
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  stable: true

  env: gce
  frequency: weekly
  team: clusters
  cluster:
    byod: {}
    cluster_compute: gcp/tests/single_node_32_cpu_gce.yaml

  run:
    timeout: 1200
    script: python launch_and_verify_cluster.py gcp/example-gpu-docker.yaml

- name: autoscaler_aws
  group: autoscaler-test
  working_dir: autoscaling_tests

  stable: False

  frequency: nightly
  team: core

  cluster:
    # leave oom disabled as test is marked unstable at the moment.
    byod:
      runtime_env:
        - RAY_memory_monitor_refresh_ms=0
      pip:
        - ray[default]
    cluster_compute: aws.yaml

  run:
    timeout: 1800
    script: python run.py