# 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
# (unless 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
#
#  # 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

#######################
# Baseline test
#######################
- name: hello_world
  python: "3.10"
  team: reef
  group: hello_world
  frequency: nightly
  working_dir: hello_world_tests

  cluster:
    anyscale_sdk_2026: true
    cluster_compute: hello_world_compute_config.yaml

  run:
    timeout: 1800
    num_retries: 3
    script: python hello_world.py

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      cluster:
        anyscale_sdk_2026: true
        cluster_compute: hello_world_compute_config_gce.yaml
    - __suffix__: azure
      env: azure
      cluster:
        anyscale_sdk_2026: true
        cluster_compute: hello_world_compute_config_azure.yaml
    - __suffix__: released
      cluster:
        ray_version: 2.50.0

- name: hello_world_custom_byod
  python: "3.10"
  team: reef
  group: hello_world
  frequency: nightly
  working_dir: hello_world_tests

  cluster:
    anyscale_sdk_2026: true
    byod:
      post_build_script: byod_hello_world.sh
      python_depset: emoji.lock
    cluster_compute: hello_world_compute_config.yaml

  run:
    timeout: 1800
    script: python hello_world_emoji.py

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      cluster:
        anyscale_sdk_2026: true
        cluster_compute: hello_world_compute_config_gce.yaml
    - __suffix__: released
      cluster:
        anyscale_sdk_2026: true
        ray_version: 2.50.0

- name: hello_world_custom_runtime_env
  python: "3.10"
  team: reef
  group: hello_world
  frequency: nightly
  working_dir: hello_world_tests

  cluster:
    anyscale_sdk_2026: true
    byod:
      runtime_env:
        - HELLO_WORLD_TEST=1
    cluster_compute: hello_world_compute_config.yaml

  run:
    timeout: 1800
    script: python hello_world.py

  variations:
    - __suffix__: aws

- name: hello_world_custom_byod_depset_only
  python: "3.10"
  team: reef
  group: hello_world
  frequency: nightly
  working_dir: hello_world_tests

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu
      python_depset: emoji.lock
    cluster_compute: hello_world_compute_config.yaml

  run:
    timeout: 1800
    script: bash test.sh

  variations:
    - __suffix__: aws

- name: hello_world_py313
  python: "3.13"
  team: reef
  group: hello_world
  frequency: nightly
  working_dir: hello_world_tests

  cluster:
    anyscale_sdk_2026: true
    byod: {}
    cluster_compute: hello_world_compute_config.yaml

  run:
    timeout: 1800
    num_retries: 3
    script: python hello_world.py

  variations:
    - __suffix__: aws

- name: "hello_world_cu130_py{{python}}"
  python: "{{python}}"
  team: reef
  group: hello_world
  frequency: nightly
  working_dir: hello_world_tests
  matrix:
    setup:
      python: ["3.10", "3.11", "3.12", "3.13"]
  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu-cu130
    cluster_compute: hello_world_compute_config.yaml

  run:
    timeout: 1800
    num_retries: 3
    script: python hello_world.py

#######################
# SSH sanity checks
#######################

- name: ssh_sanity_check
  python: "3.11"
  team: reef
  group: hello_world
  frequency: nightly
  working_dir: hello_world_tests

  cluster:
    anyscale_sdk_2026: true
    cluster_compute: hello_world_compute_config.yaml

  run:
    timeout: 600
    script: bash sanity_check_ssh.sh

  variations:
    - __suffix__: aws

- name: ssh_sanity_check_llm
  python: "3.12"
  team: llm
  group: hello_world
  frequency: nightly
  working_dir: hello_world_tests

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: llm-cu130
    # Doesn't need GPUs to run ssh sanity check
    cluster_compute: hello_world_compute_config.yaml

  run:
    timeout: 600
    script: bash sanity_check_ssh.sh

  variations:
    - __suffix__: aws

#######################
# Cluster scaling tests
#######################
- name: cluster_tune_scale_up_down
  python: "3.10"
  group: Cluster tests
  working_dir: cluster_tests

  frequency: nightly
  team: ml

  cluster:
    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
      cluster:
        anyscale_sdk_2026: true
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        anyscale_sdk_2026: true
        cluster_compute: cpt_autoscaling_1-3_gce.yaml
    - __suffix__: kuberay
      env: kuberay
      frequency: nightly
      cluster:
        cluster_compute: cpt_autoscaling_1-3_kuberay.yaml

  alert: default


#########################
# AIR release tests
#########################

# Ray Train distributed Torch benchmarks
- name: air_benchmark_torch_mnist_gpu_4x4
  python: "3.10"
  group: AIR tests
  working_dir: air_tests/air_benchmarks

  frequency: weekly
  team: ml

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu
    cluster_compute: compute_gpu_4x4_aws.yaml

  run:
    timeout: 4800
    script: RAY_TRAIN_V2_ENABLED=1 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: weekly

    cluster:
      anyscale_sdk_2026: true
      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 2 --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:
        anyscale_sdk_2026: true
        cluster_compute: compute_gpu_4x4_gce.yaml
      smoke_test:
        frequency: manual

  alert: default

- name: air_benchmark_tune_torch_mnist
  python: "3.10"
  group: AIR tests
  working_dir: air_tests/air_benchmarks

  frequency: nightly
  team: ml

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu
    cluster_compute: compute_cpu_4_aws.yaml

  run:
    timeout: 3600
    script: RAY_TRAIN_V2_ENABLED=1 python workloads/tune_torch_benchmark.py --num-runs 1 --num-trials 2 --num-workers 2

    wait_for_nodes:
      num_nodes: 4

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        anyscale_sdk_2026: true
        cluster_compute: compute_cpu_4_gce.yaml

  alert: default

# Ray Train distributed Tensorflow benchmarks
- name: air_benchmark_tensorflow_mnist_gpu_4x4
  python: "3.10"
  group: AIR tests
  working_dir: air_tests/air_benchmarks

  # https://github.com/ray-project/ray/issues/46687
  frequency: manual
  team: ml

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu
    cluster_compute: compute_gpu_4x4_aws.yaml

  run:
    timeout: 5400
    script: RAY_TRAIN_V2_ENABLED=1 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: manual

    cluster:
      anyscale_sdk_2026: true
      cluster_compute: compute_gpu_2x2_aws.yaml

    run:
      script: python workloads/tensorflow_benchmark.py run --num-runs 3 --num-epochs 60 --num-workers 2 --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:
        anyscale_sdk_2026: true
        cluster_compute: compute_gpu_4x4_gce.yaml
      smoke_test:
        frequency: manual

  alert: default


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


# Test additional CPU nodes for preprocessing.
- name: air_example_dreambooth_finetuning
  python: "3.10"
  group: AIR examples
  working_dir: air_examples/dreambooth

  # https://github.com/ray-project/ray/issues/49847
  frequency: manual
  team: ml

  cluster:
    anyscale_sdk_2026: true
    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
  python: "3.10"
  group: AIR examples
  working_dir: air_examples/dreambooth

  # https://github.com/ray-project/ray/issues/49846
  frequency: manual
  team: ml

  cluster:
    anyscale_sdk_2026: true
    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
  python: "3.10"
  group: AIR examples
  working_dir: air_examples/gptj_deepspeed_finetuning
  frequency: manual
  team: ml
  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu
      post_build_script: byod_gptj_test.sh
    cluster_compute: gptj_deepspeed_compute_aws.yaml

  run:
    timeout: 4500
    script: RAY_TRAIN_V2_ENABLED=1 python test_myst_doc.py --path gptj_deepspeed_fine_tuning.ipynb

- name: air_example_dolly_v2_lightning_fsdp_finetuning
  python: "3.10"
  group: AIR examples
  working_dir: air_examples/dolly_v2_lightning_fsdp_finetuning
  frequency: manual
  team: ml
  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu
      post_build_script: byod_dolly_test.sh
    cluster_compute: dolly_v2_fsdp_compute_aws.yaml

  run:
    timeout: 4700
    script: RAY_TRAIN_V2_ENABLED=1 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
  python: "3.10"
  group: AIR examples
  working_dir: air_examples/vicuna_13b_lightning_deepspeed_finetuning
  frequency: manual
  team: ml
  cluster:
    anyscale_sdk_2026: true
    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

#######################
# Tune tests
#######################
- name: tune_with_frequent_pausing
  python: "3.10"
  group: Tune tests
  working_dir: tune_tests

  frequency: nightly-3x
  team: ml

  cluster:
    anyscale_sdk_2026: true
    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: RAY_TRAIN_V2_ENABLED=1 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: tune_cloud_long_running_cloud_storage
  python: "3.10"
  group: Tune cloud tests
  working_dir: tune_tests/cloud_tests
  frequency: weekly
  team: ml

  cluster:
    anyscale_sdk_2026: true
    cluster_compute: tpl_aws_1x4.yaml

  run:
    # 14 hours
    timeout: 50400
    long_running: true
    script: RAY_TRAIN_V2_ENABLED=1 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: RAY_TRAIN_V2_ENABLED=1 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
  python: "3.10"
  group: Tune scalability tests
  working_dir: tune_tests/scalability_tests

  frequency: nightly
  team: ml

  cluster:
    cluster_compute: tpl_1x16.yaml

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

  alert: tune_tests

  variations:
    - __suffix__: aws
      cluster:
        anyscale_sdk_2026: true
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        anyscale_sdk_2026: true
        cluster_compute: tpl_gce_1x16.yaml
    - __suffix__: kuberay
      env: kuberay
      frequency: nightly
      cluster:
        cluster_compute: kuberay.yaml

- name: tune_scalability_durable_trainable
  python: "3.10"
  group: Tune scalability tests
  working_dir: tune_tests/scalability_tests

  frequency: nightly
  team: ml

  cluster:
    anyscale_sdk_2026: true
    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
  python: "3.10"
  group: Tune scalability tests
  working_dir: tune_tests/scalability_tests

  frequency: nightly
  team: ml

  cluster:
    anyscale_sdk_2026: true
    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
  python: "3.10"
  group: Tune scalability tests
  working_dir: tune_tests/scalability_tests

  frequency: weekly
  team: ml

  cluster:
    anyscale_sdk_2026: true
    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
  python: "3.10"
  group: Tune scalability tests
  working_dir: tune_tests/scalability_tests

  frequency: weekly
  team: ml

  cluster:
    anyscale_sdk_2026: true
    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
  python: "3.10"
  group: Tune scalability tests
  working_dir: tune_tests/scalability_tests

  frequency: nightly-3x
  team: ml

  cluster:
    anyscale_sdk_2026: true
    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
  python: "3.10"
  group: Tune scalability tests
  working_dir: tune_tests/scalability_tests

  frequency: nightly
  team: ml

  cluster:
    anyscale_sdk_2026: true
    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
  python: "3.10"
  group: Tune fault tolerance tests
  working_dir: tune_tests/fault_tolerance_tests

  frequency: nightly-3x
  team: ml

  cluster:
    anyscale_sdk_2026: true
    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

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

- name: long_running_actor_deaths
  python: "3.10"
  group: Long running tests
  working_dir: long_running_tests

  frequency: weekly

  team: core
  cluster:
    anyscale_sdk_2026: true
    byod:
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
    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_many_actor_tasks
  python: "3.10"
  group: Long running tests
  working_dir: long_running_tests

  frequency: weekly

  team: core
  cluster:
    anyscale_sdk_2026: true
    byod:
      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
  python: "3.10"
  group: Long running tests
  working_dir: long_running_tests

  frequency: weekly

  team: core
  cluster:
    anyscale_sdk_2026: true
    byod:
      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_tasks
  python: "3.10"
  group: Long running tests
  working_dir: long_running_tests

  frequency: weekly

  team: core
  cluster:
    anyscale_sdk_2026: true
    byod:
      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
  python: "3.10"
  group: Long running tests
  working_dir: long_running_tests

  frequency: weekly

  team: core
  cluster:
    anyscale_sdk_2026: true
    byod:
      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
  python: "3.10"
  group: Long running tests
  working_dir: long_running_tests

  frequency: weekly

  team: core
  cluster:
    anyscale_sdk_2026: true
    byod:
      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
  python: "3.10"
  group: Long running tests
  working_dir: long_running_tests

  frequency: weekly
  team: serve

  cluster:
    anyscale_sdk_2026: true
    byod:
      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
  python: "3.10"
  group: Long running tests
  working_dir: long_running_tests

  frequency: weekly
  team: serve

  cluster:
    anyscale_sdk_2026: true
    byod:
      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
  python: "3.10"
  group: Long running tests
  working_dir: long_running_tests

  frequency: weekly
  team: serve

  cluster:
    anyscale_sdk_2026: true
    byod:
      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

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

- name: jobs_basic_local_working_dir
  python: "3.10"
  group: Jobs tests
  working_dir: jobs_tests

  frequency: nightly
  team: serve

  cluster:
    anyscale_sdk_2026: true
    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
  python: "3.10"
  group: Jobs tests
  working_dir: jobs_tests

  frequency: nightly
  team: serve

  cluster:
    anyscale_sdk_2026: true
    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
  python: "3.10"
  group: Jobs tests
  team: serve
  frequency: nightly
  working_dir: jobs_tests

  cluster:
    anyscale_sdk_2026: true
    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
  python: "3.10"
  group: Jobs tests
  team: serve

  frequency: nightly
  working_dir: jobs_tests
  cluster:
    anyscale_sdk_2026: true
    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
    - __suffix__: py310_cu130
      python: "3.10"
      cluster:
        byod:
          type: gpu-cu130
          python_depset: gpu_cu130_py3.10.lock
        cluster_compute: compute_tpl_gpu_node.yaml
    - __suffix__: py311_cu130
      python: "3.11"
      cluster:
        byod:
          type: gpu-cu130
          python_depset: gpu_cu130_py3.11.lock
        cluster_compute: compute_tpl_gpu_node.yaml
    - __suffix__: py312_cu130
      python: "3.12"
      cluster:
        byod:
          type: gpu-cu130
          python_depset: gpu_cu130_py3.12.lock
        cluster_compute: compute_tpl_gpu_node.yaml
    - __suffix__: py313_cu130
      python: "3.13"
      cluster:
        byod:
          type: gpu-cu130
          python_depset: gpu_cu130_py3.13.lock
        cluster_compute: compute_tpl_gpu_node.yaml

- name: "jobs_check_cuda_version_cu130_py{{python}}"
  python: "{{python}}"
  group: Jobs tests
  team: serve
  frequency: nightly
  working_dir: jobs_tests
  matrix:
    setup:
      python: ["3.10", "3.11", "3.12", "3.13"]
  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu-cu130
      python_depset: gpu_cu130_py{{python}}.lock
    cluster_compute: compute_tpl_gpu_node.yaml
  run:
    timeout: 600
    script: python workloads/jobs_check_cuda_version.py --expected-cuda-version 13.0
    wait_for_nodes:
      num_nodes: 2

- name: jobs_specify_num_gpus
  python: "3.10"
  group: Jobs tests
  team: serve

  frequency: nightly
  working_dir: jobs_tests
  cluster:
    anyscale_sdk_2026: true
    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
  python: "3.10"
  group: Runtime env tests
  working_dir: runtime_env_tests

  frequency: nightly
  team: core

  cluster:
    anyscale_sdk_2026: true
    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
  python: "3.10"
  group: Runtime env tests
  working_dir: runtime_env_tests

  frequency: nightly
  team: core

  cluster:
    anyscale_sdk_2026: true
    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: pytest_serve_scale_replicas
  python: "3.10"
  group: Serve tests
  working_dir: serve_tests

  frequency: nightly
  team: serve

  cluster:
    anyscale_sdk_2026: true
    cluster_compute: compute_tpl_single_node_32_cpu.yaml
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

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

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: aws.haproxy
      cluster:
        byod:
          runtime_env:
            - RAY_SERVE_ENABLE_HA_PROXY=1

- name: pytest_serve_multi_deployment_1k_noop_replica
  python: "3.10"
  group: Serve tests
  working_dir: serve_tests

  frequency: weekly
  team: serve

  cluster:
    anyscale_sdk_2026: true
    cluster_compute: compute_tpl_32_cpu.yaml
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  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__: aws.haproxy
      cluster:
        byod:
          runtime_env:
            - RAY_SERVE_ENABLE_HA_PROXY=1
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        project_id: prj_xqmpk8ps6civt438u1hp5pi88g
        cloud_id: cld_vy7xqacrvddvbuy95auinvuqmt
        cluster_compute: compute_tpl_32_cpu_gce.yaml

- name: pytest_serve_autoscaling_load_test
  python: "3.10"
  group: Serve tests
  working_dir: serve_tests

  frequency: nightly
  team: serve

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu
    cluster_compute: compute_tpl_single_node_32_cpu.yaml
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

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

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: aws.haproxy
      cluster:
        byod:
          runtime_env:
            - RAY_SERVE_ENABLE_HA_PROXY=1

- name: serve_controller_benchmark_haproxy
  python: "3.10"
  group: Serve tests
  working_dir: serve_tests

  frequency: weekly
  team: serve

  cluster:
    anyscale_sdk_2026: true
    byod:
      runtime_env:
        # specifically testing haproxy here because with serve
        # proxy the benchmark profile changes significantly.
        # Given that haproxy is going to be the path forward, make sense
        # to use that here.
        - RAY_SERVE_ENABLE_HA_PROXY=1
        - RAY_SERVE_DIRECT_INGRESS_MIN_DRAINING_PERIOD_S=0.01
    cluster_compute: compute_tpl_8_cpu_autoscaling.yaml
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 14400
    long_running: false
    script: python workloads/microbenchmarks.py --run-controller

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        project_id: prj_xqmpk8ps6civt438u1hp5pi88g
        cloud_id: cld_vy7xqacrvddvbuy95auinvuqmt
        cluster_compute: compute_tpl_single_node_gce.yaml

- name: pytest_serve_microbenchmarks
  python: "3.10"
  group: Serve tests
  working_dir: serve_tests

  frequency: nightly
  team: serve

  cluster:
    anyscale_sdk_2026: true
    cluster_compute: compute_tpl_single_node_16_cpu.yaml
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  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:
        project_id: prj_xqmpk8ps6civt438u1hp5pi88g
        cloud_id: cld_vy7xqacrvddvbuy95auinvuqmt
        cluster_compute: compute_tpl_single_node_gce.yaml

- name: pytest_serve_throughput_optimized_microbenchmarks
  python: "3.10"
  group: Serve tests
  working_dir: serve_tests

  frequency: nightly
  team: serve

  cluster:
    anyscale_sdk_2026: true
    byod:
      runtime_env:
        - RAY_SERVE_THROUGHPUT_OPTIMIZED=1
        - RAY_SERVE_DIRECT_INGRESS_MIN_DRAINING_PERIOD_S=0.01
        - RAY_SERVE_REPLICA_GRPC_MAX_MESSAGE_LENGTH=10500000
    cluster_compute: compute_tpl_single_node_16_cpu.yaml
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

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

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: aws.haproxy
      cluster:
        byod:
          runtime_env:
            - RAY_SERVE_THROUGHPUT_OPTIMIZED=1
            - RAY_SERVE_DIRECT_INGRESS_MIN_DRAINING_PERIOD_S=0.01
            - RAY_SERVE_REPLICA_GRPC_MAX_MESSAGE_LENGTH=10500000
            - RAY_SERVE_ENABLE_HA_PROXY=1
      run:
        timeout: 7200
        long_running: false
        # HAProxy is HTTP-only and does not proxy the gRPC server bound by
        # gRPCOptions(port=9000); --run-grpc would crash with connection refused
        # under RAY_SERVE_ENABLE_HA_PROXY=1. Run the HAProxy-proxied transports
        # only.
        script: >-
          python workloads/microbenchmarks.py
          --run-http --run-handle --run-latency --run-throughput
          --run-streaming
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        project_id: prj_xqmpk8ps6civt438u1hp5pi88g
        cloud_id: cld_vy7xqacrvddvbuy95auinvuqmt
        cluster_compute: compute_tpl_single_node_gce.yaml

- name: pytest_serve_router_benchmark
  python: "3.10"
  group: Serve tests
  working_dir: serve_tests

  frequency: weekly
  team: serve

  cluster:
    anyscale_sdk_2026: true
    cluster_compute: compute_tpl_8_cpu_autoscaling.yaml
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 7200
    long_running: false
    # Every variation fully overrides this
    script: "false"

  alert: default

  variations:
    # Low-latency / high-throughput workload
    - __suffix__: aws.standard
      cluster:
        byod:
          runtime_env:
            - RAY_SERVE_RUN_USER_CODE_IN_SEPARATE_THREAD=0
            - RAY_SERVE_RUN_ROUTER_IN_SEPARATE_LOOP=0
      run:
        timeout: 7200
        long_running: false
        script: >-
          python workloads/router_microbenchmark.py
          --simulated-latency-mean-s 0.01
          --simulated-latency-cap-s 0.1
          --max-ongoing-requests-child 5
          --max-ongoing-requests-parent 5
          --max-users-per-task 24
          --router-type pow2
    # Supply-constrained workload
    - __suffix__: aws.supply_constraint
      run:
        timeout: 7200
        long_running: false
        script: >-
          python workloads/router_microbenchmark.py
          --simulated-latency-mean-s 1
          --simulated-latency-cap-s 10
          --max-ongoing-requests-child 2
          --max-ongoing-requests-parent 5
          --router-type pow2
          --router-type capacity_queue

- name: pytest_serve_resnet_benchmark
  python: "3.10"
  group: Serve tests
  working_dir: serve_tests

  frequency: nightly
  team: serve

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu
    cluster_compute: compute_tpl_gpu_node.yaml
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

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

  alert: default

  variations:
    - __suffix__: aws
    - __suffix__: aws.haproxy
      cluster:
        byod:
          runtime_env:
            - RAY_SERVE_ENABLE_HA_PROXY=1
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        project_id: prj_xqmpk8ps6civt438u1hp5pi88g
        cloud_id: cld_vy7xqacrvddvbuy95auinvuqmt
        cluster_compute: compute_tpl_gpu_node_gce.yaml

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

- name: training_ingest_benchmark-task=image_classification
  python: "3.10"
  group: Train tests
  working_dir: train_tests/benchmark

  frequency: nightly
  team: ml

  cluster:
    byod:
      runtime_env:
        # Enable verbose stats for resource manager
        - RAY_DATA_DEBUG_RESOURCE_MANAGER=1
        # Fail the test if a worker OOMs
        - RAYTEST_FAIL_ON_WORKER_OOM=1

      # 'type: gpu' means: use the 'ray-ml' image.
      type: gpu
    anyscale_sdk_2026: true
    cluster_compute: compute_configs/compute_gpu_4x4_aws.yaml

  variations:
    - __suffix__: full_training.parquet
      run:
        timeout: 2000
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --dataloader_type=ray_data --num_workers=16 --image_classification_data_format=parquet

    - __suffix__: full_training.parquet.preserve_order
      frequency: weekly
      run:
        timeout: 2000
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --dataloader_type=ray_data --num_workers=16 --image_classification_data_format=parquet --preserve_order=True

    - __suffix__: full_training.parquet.torch_dataloader
      frequency: manual
      run:
        timeout: 2000
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --dataloader_type=torch --num_workers=16 --image_classification_data_format=parquet

    - __suffix__: skip_training.parquet
      run:
        timeout: 1200
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --dataloader_type=ray_data --num_workers=16 --skip_train_step=True --skip_validation_at_epoch_end=True --image_classification_data_format=parquet

    - __suffix__: skip_training.parquet.preserve_order
      frequency: weekly
      run:
        timeout: 1200
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --dataloader_type=ray_data --num_workers=16 --skip_train_step=True --skip_validation_at_epoch_end=True --image_classification_data_format=parquet --preserve_order=True

    - __suffix__: skip_training.parquet.torch_dataloader
      frequency: manual
      run:
        timeout: 1200
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --dataloader_type=torch --num_workers=16 --skip_train_step=True --skip_validation_at_epoch_end=True --image_classification_data_format=parquet

    - __suffix__: skip_training.parquet.slow_consumer
      run:
        timeout: 3000
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --dataloader_type=ray_data --num_workers=16 --skip_train_step=True --skip_validation_at_epoch_end=True --image_classification_data_format=parquet --train_step_sleep_s=5.0 --max_train_batches=500 --limit_training_rows=-1 --image_classification_use_1t_dataset=True

    - __suffix__: skip_training.parquet.fault_tolerance
      frequency: weekly
      run:
        timeout: 2700
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --dataloader_type=ray_data --num_workers=16 --mock_gpu=True --skip_train_step=True --skip_validation_step=True --num_epochs=5 --max_failures=4 --image_classification_data_format=parquet
        prepare: python setup_chaos.py --kill-interval 480 --max-to-kill 2 --kill-delay 360
      cluster:
        byod:
          runtime_env:
            - RAY_DATA_DEBUG_RESOURCE_MANAGER=1
            - RAYTEST_FAIL_ON_WORKER_OOM=0  # Workers are intentionally killed
        cluster_compute: compute_configs/compute_mock_gpu_4x4_aws.yaml

    - __suffix__: full_training.s3_url
      run:
        timeout: 2000
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --dataloader_type=ray_data --num_workers=16 --image_classification_data_format=s3_url

    - __suffix__: skip_training.s3_url
      run:
        timeout: 1200
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --dataloader_type=ray_data --num_workers=16 --skip_train_step True --skip_validation_at_epoch_end True --image_classification_data_format=s3_url

    - __suffix__: full_training.jpeg
      run:
        timeout: 2000
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --image_classification_data_format=jpeg --dataloader_type=ray_data --num_workers=16

    - __suffix__: full_training.jpeg.preserve_order
      frequency: weekly
      run:
        timeout: 2000
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --image_classification_data_format=jpeg --dataloader_type=ray_data --num_workers=16 --preserve_order=True

    - __suffix__: full_training.jpeg.torch_dataloader
      frequency: manual
      run:
        timeout: 2000
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --image_classification_data_format=jpeg --dataloader_type=torch --num_workers=16 --num_torch_workers=16

    - __suffix__: skip_training.jpeg
      run:
        timeout: 1200
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --image_classification_data_format=jpeg --dataloader_type=ray_data --num_workers=16 --skip_train_step=True --skip_validation_at_epoch_end=True

    - __suffix__: skip_training.jpeg.preserve_order
      frequency: weekly
      run:
        timeout: 2400
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --image_classification_data_format=jpeg --dataloader_type=ray_data --num_workers=16 --skip_train_step=True --skip_validation_at_epoch_end=True --preserve_order=True

    - __suffix__: skip_training.jpeg.torch_dataloader
      frequency: manual
      run:
        timeout: 1200
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --image_classification_data_format=jpeg --dataloader_type=torch --num_workers=16 --num_torch_workers=16 --skip_train_step=True --skip_validation_at_epoch_end=True

    - __suffix__: skip_training.jpeg.local_fs
      run:
        timeout: 1200
        script: bash image_classification/jpeg/download_input_data_from_s3.sh && RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --image_classification_data_format=jpeg --image_classification_local_dataset=True --dataloader_type=ray_data --num_workers=1 --skip_train_step=True --skip_validation_at_epoch_end=True
      cluster:
        cluster_compute: compute_configs/compute_gpu_1x1_aws.yaml

    - __suffix__: skip_training.jpeg.local_fs.preserve_order
      frequency: manual
      run:
        timeout: 1200
        script: bash image_classification/jpeg/download_input_data_from_s3.sh && RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --image_classification_data_format=jpeg --image_classification_local_dataset=True --dataloader_type=ray_data --num_workers=1 --skip_train_step=True --skip_validation_at_epoch_end=True --preserve_order=True
      cluster:
        cluster_compute: compute_configs/compute_gpu_1x1_aws.yaml

    - __suffix__: skip_training.jpeg.local_fs.torch_dataloader
      frequency: manual
      run:
        timeout: 1200
        script: bash image_classification/jpeg/download_input_data_from_s3.sh && RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --image_classification_data_format=jpeg --image_classification_local_dataset=True --dataloader_type=torch --num_workers=1 --num_torch_workers=32 --skip_train_step=True --skip_validation_at_epoch_end=True
      cluster:
        cluster_compute: compute_configs/compute_gpu_1x1_aws.yaml

    - __suffix__: skip_training.jpeg.local_fs_multi_gpus
      run:
        timeout: 1200
        script: bash image_classification/jpeg/download_input_data_from_s3.sh && RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --image_classification_data_format=jpeg --image_classification_local_dataset=True --dataloader_type=ray_data --num_workers=4 --skip_train_step=True --skip_validation_at_epoch_end=True
      cluster:
        cluster_compute: compute_configs/compute_gpu_1x4_aws.yaml

    - __suffix__: skip_training.jpeg.local_fs_multi_gpus.preserve_order
      frequency: manual
      run:
        timeout: 1200
        script: bash image_classification/jpeg/download_input_data_from_s3.sh && RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --image_classification_data_format=jpeg --image_classification_local_dataset=True --dataloader_type=ray_data --num_workers=4 --skip_train_step=True --skip_validation_at_epoch_end=True --preserve_order=True
      cluster:
        cluster_compute: compute_configs/compute_gpu_1x4_aws.yaml

    - __suffix__: skip_training.jpeg.local_fs_multi_gpus.torch_dataloader
      frequency: manual
      run:
        timeout: 1200
        script: bash image_classification/jpeg/download_input_data_from_s3.sh && RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --image_classification_data_format=jpeg --image_classification_local_dataset=True --dataloader_type=torch --num_workers=4 --num_torch_workers=9 --skip_train_step=True --skip_validation_at_epoch_end=True
      cluster:
        cluster_compute: compute_configs/compute_gpu_1x4_aws.yaml

- name: training_ingest_benchmark-task=image_classification-heterogenous_fixed_size
  python: "3.10"
  group: Train tests
  working_dir: train_tests/benchmark

  frequency: nightly
  team: ml

  cluster:
    byod:
      runtime_env:
        # Enable verbose stats for resource manager
        - RAY_DATA_DEBUG_RESOURCE_MANAGER=1
        # Fail the test if a worker OOMs
        - RAYTEST_FAIL_ON_WORKER_OOM=1
        # Fail the test if a node dies
        - RAYTEST_FAIL_ON_DEAD_NODES=1

      # 'type: gpu' means: use the 'ray-ml' image.
      type: gpu
    anyscale_sdk_2026: true
    cluster_compute: compute_configs/heterogenous_fixed_size_gpu_4x4_aws.yaml

  variations:
    - __suffix__: skip_training.parquet
      run:
        timeout: 1200
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --dataloader_type=ray_data --num_workers=16 --skip_train_step True --skip_validation_at_epoch_end True --image_classification_data_format=parquet

    - __suffix__: skip_training.s3_url
      run:
        timeout: 1200
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --dataloader_type=ray_data --num_workers=16 --skip_train_step True --skip_validation_at_epoch_end True --image_classification_data_format=s3_url

- name: training_ingest_benchmark-task=image_classification-heterogenous_autoscaling
  python: "3.10"
  group: Train tests
  working_dir: train_tests/benchmark

  frequency: nightly
  team: ml

  cluster:
    byod:
      runtime_env:
        # Enable verbose stats for resource manager
        - RAY_DATA_DEBUG_RESOURCE_MANAGER=1
        # Fail the test if a worker OOMs
        - RAYTEST_FAIL_ON_WORKER_OOM=1
        # Fail the test if a node dies
        - RAYTEST_FAIL_ON_DEAD_NODES=1

      # 'type: gpu' means: use the 'ray-ml' image.
      type: gpu
    anyscale_sdk_2026: true
    cluster_compute: compute_configs/heterogenous_autoscaling_gpu_4x4_aws.yaml

  variations:
    - __suffix__: skip_training.parquet
      run:
        timeout: 1200
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --dataloader_type=ray_data --num_workers=16 --skip_train_step True --skip_validation_at_epoch_end True --image_classification_data_format=parquet

    - __suffix__: skip_training.s3_url
      run:
        timeout: 1200
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --dataloader_type=ray_data --num_workers=16 --skip_train_step True --skip_validation_at_epoch_end True --image_classification_data_format=s3_url

- name: training_ingest_benchmark-task=recsys
  python: "3.10"
  group: Train tests
  working_dir: train_tests/benchmark

  # TODO: Keep these manual for now -- there are a still a few issues to resolve.
  frequency: manual
  team: ml

  cluster:
    byod:
      runtime_env:
        # Fail the test if a worker OOMs
        - RAYTEST_FAIL_ON_WORKER_OOM=1
      type: gpu
      post_build_script: byod_install_torchrec.sh
    anyscale_sdk_2026: true
    cluster_compute: compute_configs/compute_gpu_8xA100_aws.yaml

  variations:
    - __suffix__: skip_training.mock_dataloader
      run:
        timeout: 600
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=recsys --dataloader_type=mock --num_workers=8 --train_batch_size=8192 --num_epochs=1 --skip_train_step=True --skip_validation_at_epoch_end=True

    - __suffix__: skip_training.ray_data
      run:
        timeout: 600
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=recsys --dataloader_type=ray_data --num_workers=8 --train_batch_size=8192 --num_epochs=1 --skip_train_step=True --skip_validation_at_epoch_end=True

    - __suffix__: full_training.mock_dataloader
      run:
        timeout: 600
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=recsys --dataloader_type=mock --num_workers=8 --train_batch_size=8192 --validation_batch_size=16384 --num_epochs=1

    - __suffix__: full_training.ray_data
      run:
        timeout: 1200
        script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=recsys --dataloader_type=ray_data --num_workers=8 --train_batch_size=8192 --validation_batch_size=16384 --num_epochs=1

- name: training_ingest_benchmark-soak_test
  python: "3.10"
  group: Train tests
  working_dir: train_tests/benchmark

  frequency: weekly
  team: ml

  cluster:
    byod:
      type: gpu
    anyscale_sdk_2026: true
    cluster_compute: compute_configs/compute_gpu_4x4_cpu_4_aws.yaml

  run:
    timeout: 43200
    long_running: true
    script: RAY_TRAIN_V2_ENABLED=1 python train_benchmark.py --task=image_classification --dataloader_type=ray_data --num_workers=16 --image_classification_data_format=parquet --num_epochs=50

- name: train_multinode_persistence
  python: "3.10"
  group: Train tests
  working_dir: train_tests/multinode_persistence

  frequency: nightly
  team: ml

  cluster:
    anyscale_sdk_2026: true
    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_v1
    - __suffix__: aws_v2
      run:
        script: RAY_TRAIN_V2_ENABLED=1 pytest -v test_persistence.py -s
    - __suffix__: gce_v1
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gce.yaml
    - __suffix__: gce_v2
      env: gce
      frequency: manual
      cluster:
        cluster_compute: compute_gce.yaml
      run:
        script: RAY_TRAIN_V2_ENABLED=1 pytest -v test_persistence.py -s

  alert: default

- name: train_async_checkpointing_validation_benchmark
  python: "3.10"
  group: Train tests
  working_dir: train_tests/async_checkpointing_validation_benchmark
  frequency: nightly
  team: ml
  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu
    cluster_compute: compute_aws.yaml
  run:
    timeout: 10800
    script: RAY_TRAIN_V2_ENABLED=1 python test_async_checkpointing_validation_benchmark.py
    wait_for_nodes:
      num_nodes: 2
  alert: default

- name: train_colocate_trainer
  python: "3.10"
  group: Train tests
  working_dir: train_tests/colocate_trainer

  # Ray Train V2 doesn't support colocation.
  # TODO: Decide whether to remove this test or re-enable it if we add support again.
  frequency: manual
  team: ml

  cluster:
    anyscale_sdk_2026: true
    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: torch_local_mode
  python: "3.10"
  group: Train tests
  working_dir: train_tests/local_mode

  frequency: nightly
  team: ml

  cluster:
    byod:
      type: gpu
    anyscale_sdk_2026: true
    cluster_compute: compute_gpu_2x4_aws.yaml

  run:
    timeout: 1800
    script: python torch_local_mode_launcher.py

# Reference: how to run a release test on the torchft Ray ML image variant.
# The custom BYOD image is the core Ray CUDA image (py3.13) with the torchft ML
# dependency lock (release/ray_release/byod/ml_torchft_py3.13.lock, generated by
# ci/raydepsets/configs/release_ml_torchft_tests.depsets.yaml) installed on top.
# Setting `python_depset` is enough: the BYOD build automatically runs
# `uv pip install --system --no-deps -r python_depset.lock`, so no
# post_build_script is needed to install the deps.
- name: torchft_hello_world
  python: "3.13"
  group: Train tests
  team: ml
  frequency: manual
  working_dir: train_tests/torchft

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: cu123
      python_depset: ml_torchft_py3.13.lock
    cluster_compute: compute_aws.yaml

  run:
    timeout: 1800
    script: python torchft_hello_world.py

- name: torch_lightning
  python: "3.10"
  group: Train tests
  working_dir: train_tests/pytorch_lightning

  frequency: nightly
  team: ml

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu
      post_build_script: byod_pytorch_lightning_test.sh
    cluster_compute: compute_aws.yaml

  run:
    timeout: 1800
    script: RAY_TRAIN_V2_ENABLED=1 python test_lightning.py

- name: huggingface_transformers
  python: "3.10"
  group: Train tests
  working_dir: train_tests/huggingface_transformers

  frequency: nightly
  team: ml

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu
      post_build_script: byod_huggingface_transformers_test.sh
    cluster_compute: compute_aws.yaml

  run:
    timeout: 1800
    script: RAY_TRAIN_V2_ENABLED=1 python test_huggingface_transformers.py

- name: huggingface_accelerate
  python: "3.10"
  group: Train tests
  working_dir: train_tests/huggingface_accelerate

  frequency: nightly
  team: ml

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu
    cluster_compute: compute_aws.yaml

  run:
    timeout: 1800
    script: RAY_TRAIN_V2_ENABLED=1 python test_huggingface_accelerate.py

- name: "{{framework}}_train_batch_inference_benchmark_{{size}}"
  python: "3.10"
  group: Train tests
  working_dir: train_tests/xgboost_lightgbm

  matrix:
    setup:
      framework: [xgboost, lightgbm]
      size: [10G, 100G]

  frequency: nightly-3x
  team: ml
  cluster:
    byod:
      type: gpu
    anyscale_sdk_2026: true
    cluster_compute: compute_aws_{{size}}.yaml

  run:
    timeout: 1800
    script: RAY_TRAIN_V2_ENABLED=1 python train_batch_inference_benchmark.py "{{framework}}" --size={{size}}

  smoke_test:
    frequency: manual

    run:
      timeout: 1800

  alert: default

- name: train_elastic_e2e_with_node_killing
  python: "3.10"
  group: Train tests
  working_dir: train_tests/elastic_training

  frequency: weekly
  team: ml

  cluster:
    byod:
      type: gpu
    anyscale_sdk_2026: true
    cluster_compute: compute_gpu_aws.yaml

  run:
    timeout: 4800
    script: RAY_TRAIN_V2_ENABLED=1 python torch_elastic_e2e.py run --num-epochs 60

    wait_for_nodes:
      num_nodes: 1

  variations:
    - __suffix__: aws

  alert: default

- name: elastic_benchmark
  python: "3.10"
  group: Train tests
  working_dir: train_tests/benchmark

  frequency: weekly
  team: ml

  cluster:
    byod:
      type: gpu
    anyscale_sdk_2026: true
    cluster_compute: compute_configs/compute_gpu_8x4_aws.yaml

  run:
    timeout: 7200
    script: RAY_TRAIN_V2_ENABLED=1 python -m elastic_training.elastic_e2e --task="image_classification" --image_classification_data_format="parquet" --dataloader_type=ray_data --checkpoint_every_n_steps=100 --min_workers=8 --max_workers=32 --num_epochs=4

    wait_for_nodes:
      num_nodes: 1

  variations:
    - __suffix__: aws

  alert: default

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

# -----------------
# RLlib - APPO test
# -----------------
- name: rllib_learning_tests_pong_appo_torch
  python: "3.10"
  group: RLlib tests
  working_dir: rllib_tests

  frequency: nightly
  team: rllib

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu
      post_build_script: byod_rllib.sh
      runtime_env:
        - RLLIB_TEST_NO_JAX_IMPORT=1
    cluster_compute: 1gpu_16cpus.yaml

  run:
    timeout: 1500  # expected 1000 seconds
    script: python example_algorithms/appo/pong_appo.py --num-learners=1 --num-env-runners=12 --as-release-test

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

- name: shuffle_100gb
  python: "3.10"
  group: core-multi-test
  working_dir: nightly_tests

  frequency: nightly

  team: core
  cluster:
    anyscale_sdk_2026: true
    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
  python: "3.10"
  group: core-multi-test
  working_dir: nightly_tests
  env: aws_perf

  frequency: nightly

  team: core
  cluster:
    anyscale_sdk_2026: true
    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:
        anyscale_sdk_2026: true
        cluster_compute: stress_tests/placement_group_tests_compute_gce.yaml

- name: autoscaling_shuffle_1tb_1000_partitions
  python: "3.10"
  group: core-multi-test
  working_dir: nightly_tests

  frequency: nightly

  team: core
  cluster:
    anyscale_sdk_2026: true
    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
  python: "3.10"
  group: core-daily-test
  team: core
  frequency: nightly
  env: aws_perf
  working_dir: microbenchmark

  cluster:
    anyscale_sdk_2026: true
    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:
        anyscale_sdk_2026: true
        cluster_compute: tpl_64_gce.yaml
    - __suffix__: aws.py312
      frequency: weekly
      python: "3.12"

- name: compiled_graphs
  python: "3.10"
  group: core-daily-test
  team: core
  frequency: manual
  working_dir: microbenchmark

  cluster:
    anyscale_sdk_2026: true
    cluster_compute: tpl_64.yaml

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

- name: compiled_graphs_GPU
  python: "3.10"
  group: core-daily-test
  team: core
  frequency: manual
  working_dir: microbenchmark

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu
    cluster_compute: experimental/compute_t4_gpu.yaml

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

- name: compiled_graphs_GPU_multinode
  python: "3.10"
  group: core-daily-test
  team: core
  frequency: manual
  working_dir: microbenchmark

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu
    cluster_compute: experimental/compute_gpu_2x1_aws.yaml

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

- name: "compiled_graphs_GPU_cu130_py{{python}}"
  python: "{{python}}"
  group: core-daily-test
  team: core
  frequency: manual
  working_dir: microbenchmark
  matrix:
    setup:
      python: ["3.10", "3.11", "3.12", "3.13"]

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu-cu130
      python_depset: gpu_cu130_py{{python}}.lock
    cluster_compute: experimental/compute_l4_gpu.yaml

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

- name: "compiled_graphs_GPU_multinode_cu130_py{{python}}"
  python: "{{python}}"
  group: core-daily-test
  team: core
  frequency: manual
  working_dir: microbenchmark
  matrix:
    setup:
      python: ["3.10", "3.11", "3.12", "3.13"]

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu-cu130
      python_depset: gpu_cu130_py{{python}}.lock
    cluster_compute: experimental/compute_l4_gpu_2x1_aws.yaml

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

- name: rdt_single_node_T4_microbenchmark
  python: "3.10"
  group: core-daily-test
  team: core
  frequency: nightly
  working_dir: microbenchmark

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: gpu
    cluster_compute: experimental/compute_t4_gpu.yaml

  run:
    timeout: 1800
    # NIXL currently only works on T4's for <100mb tensors so not enabling.
    script: python experimental/rdt_single_node_microbenchmark.py

- name: rdt_single_node_A100_microbenchmark
  python: "3.12"
  group: core-daily-test
  team: core
  frequency: weekly
  working_dir: microbenchmark

  cluster:
    anyscale_sdk_2026: true
    byod:
      type: llm-cu130 # llm image already has nixl
      post_build_script: byod_llm_pin_cupy_cuda12x.sh # match cupy-cuda12x to ray-llm base lock (BYOD testdeps can float major)
      runtime_env: # env vars for nixl
        - UCX_TLS=all
        - UCX_NET_DEVICES=all
    cluster_compute: experimental/compute_a100_gpu.yaml

  run:
    timeout: 1800
    script: python experimental/rdt_single_node_microbenchmark.py --enable_10gb --enable_nixl

# TODO(dayshah): Add multi-node benchmarks with actual hardware RDMA support
# - name: rdt_GPU_multinode
#   python: "3.10"
#   group: core-daily-test
#   team: core
#   frequency: nightly
#   working_dir: microbenchmark

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

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

- name: benchmark_worker_startup
  python: "3.10"
  group: core-daily-test
  team: core
  frequency: nightly
  working_dir: benchmark-worker-startup

  cluster:
    anyscale_sdk_2026: true
    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:
        anyscale_sdk_2026: true
        cluster_compute: only_head_node_1gpu_64cpu_gce.yaml

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

  frequency: manual # was nightly
  team: core

  cluster:
    anyscale_sdk_2026: true
    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: stress_test_state_api_scale
  python: "3.10"
  group: core-daily-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    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:
      anyscale_sdk_2026: true
      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:
        anyscale_sdk_2026: true
        cluster_compute: stress_tests/stress_tests_compute_large_gce.yaml
      smoke_test:
        frequency: manual
        cluster:
          cluster_compute: stress_tests/smoke_test_compute_gce.yaml


- name: shuffle_20gb_with_state_api
  python: "3.10"
  group: core-daily-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    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
  python: "3.10"
  group: core-daily-test
  working_dir: nightly_tests
  env: aws_perf

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    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:
      anyscale_sdk_2026: true
      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:
        anyscale_sdk_2026: true
        cluster_compute: stress_tests/stress_tests_compute_gce.yaml
      smoke_test:
        frequency: manual
        cluster:
          cluster_compute: stress_tests/smoke_test_compute_gce.yaml

- name: stress_test_dead_actors
  python: "3.10"
  group: core-daily-test
  working_dir: nightly_tests
  env: aws_perf

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    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:
      anyscale_sdk_2026: true
      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:
        anyscale_sdk_2026: true
        cluster_compute: stress_tests/stress_tests_compute_gce.yaml
      smoke_test:
        frequency: manual
        cluster:
          cluster_compute: stress_tests/smoke_test_compute_gce.yaml

# 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
  python: "3.10"
  group: core-daily-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    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:
        anyscale_sdk_2026: true
        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
  python: "3.10"
  group: core-daily-test
  working_dir: nightly_tests

  frequency: nightly
  team: core

  cluster:
    anyscale_sdk_2026: true
    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:
        anyscale_sdk_2026: true
        cluster_compute: stress_tests/smoke_test_compute_gce.yaml
      smoke_test:
        frequency: manual

- name: single_node_oom
  python: "3.10"
  group: core-daily-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  env: aws_perf
  cluster:
    anyscale_sdk_2026: true
    byod:
      runtime_env:
        # Lower the memory usage threshold at which the Ray memory monitor will kill
        # running tasks so that it does not compete with the system OOM killer.
        - RAY_memory_usage_threshold=0.7
    cluster_compute: stress_tests/stress_tests_single_node_oom_compute.yaml

  run:
    # The script parameters are tuned to run for ~30min.
    timeout: 3600
    # The memory allocated per task is tuned along with the `RAY_memory_usage_threshold` setting above
    # to trigger the Ray memory monitor to kick in before the Ray worker cgroup OOMs.
    script: python stress_tests/test_parallel_tasks_memory_pressure.py --num-tasks 64 --mem-pct-per-task .5

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        anyscale_sdk_2026: true
        cluster_compute: stress_tests/stress_tests_single_node_oom_compute_gce.yaml


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

  frequency: manual # was nightly-3x
  team: core

  cluster:
    anyscale_sdk_2026: true
    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
  python: "3.10"
  group: core-daily-test
  working_dir: benchmarks

  frequency: nightly-3x
  team: core
  cluster:
    anyscale_sdk_2026: true
    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:
        anyscale_sdk_2026: true
        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
  python: "3.10"
  group: core-daily-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    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:
        anyscale_sdk_2026: true
        cluster_compute: placement_group_tests/compute_gce.yaml

- name: placement_group_performance_test
  python: "3.10"
  group: core-daily-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    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:
        anyscale_sdk_2026: true
        cluster_compute: placement_group_tests/pg_perf_test_compute_gce.yaml


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

- name: single_node
  python: "3.10"
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly
  team: core
  env: aws_perf
  cluster:
    byod:
      runtime_env:
        - LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so
    cluster_compute: single_node.yaml
    anyscale_sdk_2026: true

  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
  python: "3.10"
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly
  team: core
  env: aws_perf
  cluster:
    anyscale_sdk_2026: true
    byod:
      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: callback_throughput
  python: "3.10"
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly
  team: core
  env: aws_perf
  cluster:
    anyscale_sdk_2026: true
    byod:
      runtime_env:
        - LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so
    cluster_compute: object_store/callback_throughput.yaml

  run:
    timeout: 600
    script: python object_store/test_callback_throughput.py
    wait_for_nodes:
      num_nodes: 11

  variations:
    - __suffix__: aws

- name: small_objects
  python: "3.10"
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly
  team: core
  env: aws_perf
  cluster:
    anyscale_sdk_2026: true
    byod:
      runtime_env:
        - LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so
    cluster_compute: object_store/small_objects.yaml

  run:
    timeout: 3600
    script: python object_store/test_small_objects.py
    wait_for_nodes:
      num_nodes: 5

  variations:
    - __suffix__: aws

- name: large_objects
  python: "3.10"
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly
  team: core
  env: aws_perf
  cluster:
    anyscale_sdk_2026: true
    byod:
      runtime_env:
        - LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so
    cluster_compute: object_store/large_objects.yaml

  run:
    timeout: 3600
    script: python object_store/test_large_objects.py
    wait_for_nodes:
      num_nodes: 10

  variations:
    - __suffix__: aws

- name: many_actors
  python: "3.10"
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly-3x
  team: core
  env: aws_perf
  cluster:
    anyscale_sdk_2026: true
    byod:
      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:
        anyscale_sdk_2026: true
        cluster_compute: distributed_gce.yaml

- name: many_actors_smoke_test
  python: "3.10"
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    byod:
      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
  python: "3.10"
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly
  team: core
  env: aws_perf
  cluster:
    anyscale_sdk_2026: true
    byod:
      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:
        anyscale_sdk_2026: true
        cluster_compute: distributed_gce.yaml

- name: many_pgs
  python: "3.10"
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly-3x
  team: core
  env: aws_perf
  cluster:
    anyscale_sdk_2026: true
    byod:
      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:
        anyscale_sdk_2026: true
        cluster_compute: distributed_gce.yaml


- name: many_pgs_smoke_test
  python: "3.10"
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    byod:
      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
  python: "3.10"
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly-3x
  team: core
  env: aws_perf
  cluster:
    anyscale_sdk_2026: true
    byod:
      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
  python: "3.10"
  group: core-scalability-test
  working_dir: benchmarks

  frequency: nightly
  team: core
  cluster:
    byod:
      runtime_env:
        - LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so
    cluster_compute: scheduling.yaml
    anyscale_sdk_2026: true

  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

# - 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


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

- name: chaos_many_tasks_baseline
  python: "3.10"
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    cluster_compute: chaos_test/compute_template.yaml

  run:
    timeout: 3600
    wait_for_nodes:
      num_nodes: 10
    script: python chaos_test/test_chaos.py --workload=tasks --disable-resource-killer

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

- name: chaos_many_tasks_kill_raylet
  python: "3.10"
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    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.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
  python: "3.10"
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    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.py --workload=tasks

  variations:
    - __suffix__: aws

- name: chaos_many_tasks_iptable_failure_injection
  python: "3.10"
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    byod:
      runtime_env:
        - RAY_health_check_period_ms=10000
        - RAY_health_check_timeout_ms=100000
        - RAY_health_check_failure_threshold=10
        - RAY_gcs_rpc_server_connect_timeout_s=60
    cluster_compute: chaos_test/compute_template.yaml

  run:
    timeout: 3600
    wait_for_nodes:
      num_nodes: 10
    script: >
      python simulate_cross_az_network_failure.py --network-failure-interval 60 --network-failure-duration 5 --command python chaos_test/test_chaos.py --workload=tasks --disable-resource-killer

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

- name: chaos_many_actors_baseline
  python: "3.10"
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    cluster_compute: chaos_test/compute_template.yaml

  run:
    timeout: 4200
    wait_for_nodes:
      num_nodes: 10
    script: python chaos_test/test_chaos.py --workload=actors --disable-resource-killer

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

- name: chaos_many_actors_kill_raylet
  python: "3.10"
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    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.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
  python: "3.10"
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    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.py --workload=actors

  variations:
    - __suffix__: aws

- name: chaos_many_actors_iptable_failure_injection
  python: "3.10"
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    byod:
      runtime_env:
        - RAY_health_check_period_ms=10000
        - RAY_health_check_timeout_ms=100000
        - RAY_health_check_failure_threshold=10
        - RAY_gcs_rpc_server_connect_timeout_s=60
    cluster_compute: chaos_test/compute_template.yaml

  run:
    timeout: 4200
    wait_for_nodes:
      num_nodes: 10
    script: >
      python simulate_cross_az_network_failure.py --network-failure-interval 60 --network-failure-duration 5 --command python chaos_test/test_chaos.py --workload=actors --disable-resource-killer

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

- name: chaos_streaming_generator_baseline
  python: "3.10"
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    cluster_compute: chaos_test/compute_template.yaml

  run:
    timeout: 3600
    wait_for_nodes:
      num_nodes: 10
    script: python chaos_test/test_chaos.py --workload=streaming --disable-resource-killer

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

- name: chaos_streaming_generator_kill_raylet
  python: "3.10"
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    cluster_compute: chaos_test/compute_template.yaml

  run:
    timeout: 3600
    wait_for_nodes:
      num_nodes: 10
    prepare: python setup_chaos.py --no-start --max-to-kill 10
    script: python chaos_test/test_chaos.py --workload=streaming

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

- name: chaos_streaming_generator_terminate_instance
  python: "3.10"
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    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 --max-to-kill 10
    script: python chaos_test/test_chaos.py --workload=streaming

  variations:
    - __suffix__: aws

- name: chaos_streaming_generator_iptable_failure_injection
  python: "3.10"
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    byod:
      runtime_env:
        - RAY_health_check_period_ms=10000
        - RAY_health_check_timeout_ms=100000
        - RAY_health_check_failure_threshold=10
        - RAY_gcs_rpc_server_connect_timeout_s=60
    cluster_compute: chaos_test/compute_template.yaml

  run:
    timeout: 3600
    wait_for_nodes:
      num_nodes: 10
    script: >
      python simulate_cross_az_network_failure.py --network-failure-interval 60 --network-failure-duration 5 --command python chaos_test/test_chaos.py --workload=streaming --disable-resource-killer

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

- name: chaos_object_ref_borrowing_baseline
  python: "3.10"
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    cluster_compute: chaos_test/compute_template.yaml

  run:
    timeout: 3600
    wait_for_nodes:
      num_nodes: 10
    script: python chaos_test/test_chaos.py --workload=borrowing --disable-resource-killer

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

- name: chaos_object_ref_borrowing_kill_raylet
  python: "3.10"
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    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.py --workload=borrowing

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

- name: chaos_object_ref_borrowing_terminate_instance
  python: "3.10"
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    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.py --workload=borrowing

  variations:
    - __suffix__: aws

- name: chaos_object_ref_borrowing_iptable_failure_injection
  python: "3.10"
  group: core-nightly-test
  working_dir: nightly_tests

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    byod:
      runtime_env:
        - RAY_health_check_period_ms=10000
        - RAY_health_check_timeout_ms=100000
        - RAY_health_check_failure_threshold=10
        - RAY_gcs_rpc_server_connect_timeout_s=60
    cluster_compute: chaos_test/compute_template.yaml

  run:
    timeout: 3600
    wait_for_nodes:
      num_nodes: 10
    script: >
      python simulate_cross_az_network_failure.py --network-failure-interval 60 --network-failure-duration 5 --command python chaos_test/test_chaos.py --workload=borrowing --disable-resource-killer

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

#####################
# Observability tests
#####################
- name: agent_stress_test
  python: "3.10"
  group: core-observability-test
  working_dir: dashboard

  frequency: nightly
  team: core
  cluster:
    anyscale_sdk_2026: true
    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

  # Failing since Oct 2024.
  # https://github.com/ray-project/ray/issues/36190
  frequency: manual
  team: serve
  cluster:
    anyscale_sdk_2026: true
    cluster_compute: compute_tpl.yaml

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

- name: azure_cluster_launcher
  python: "3.10"
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/
  # TODO: the test seems to be leaking resources, and will
  # start failing over time; needs more investigation.
  frequency: manual
  team: clusters
  cluster:
    byod:
      post_build_script: byod_azure_cluster_launcher.sh
    cluster_compute: azure/tests/azure_compute.yaml
  run:
    timeout: 2400
    script: bash release/azure_docker_login.sh && python -I launch_and_verify_cluster.py azure/tests/azure-cluster.yaml --num-expected-nodes 3 --retries 10

  variations:
    - __suffix__: v1
      run:
        script: RAY_UP_enable_autoscaler_v2=0 python -I launch_and_verify_cluster.py azure/tests/azure-cluster.yaml --num-expected-nodes 3 --retries 10
    - __suffix__: v2
      run:
        script: RAY_UP_enable_autoscaler_v2=1 python -I launch_and_verify_cluster.py azure/tests/azure-cluster.yaml --num-expected-nodes 3 --retries 10

- name: aws_cluster_launcher
  python: "3.10"
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  frequency: nightly
  team: clusters

  cluster:
    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

  variations:
    - __suffix__: v1
      run:
        script: RAY_UP_enable_autoscaler_v2=0 python launch_and_verify_cluster.py aws/tests/aws_cluster.yaml --num-expected-nodes 2 --retries 10
    - __suffix__: v2
      run:
        script: RAY_UP_enable_autoscaler_v2=1 python launch_and_verify_cluster.py aws/tests/aws_cluster.yaml --num-expected-nodes 2 --retries 10

- name: aws_cluster_launcher_nightly_image
  python: "3.10"
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  frequency: manual
  team: clusters
  cluster:
    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

  variations:
    - __suffix__: v1
      run:
        script: RAY_UP_enable_autoscaler_v2=0 python launch_and_verify_cluster.py aws/tests/aws_cluster.yaml --num-expected-nodes 2 --retries 10 --docker-override nightly
    - __suffix__: v2
      run:
        script: RAY_UP_enable_autoscaler_v2=1 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
  python: "3.10"
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  frequency: manual
  team: clusters
  cluster:
    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

  variations:
    - __suffix__: v1
      run:
        script: RAY_UP_enable_autoscaler_v2=0 python launch_and_verify_cluster.py aws/tests/aws_cluster.yaml --num-expected-nodes 2 --retries 10 --docker-override latest
    - __suffix__: v2
      run:
        script: RAY_UP_enable_autoscaler_v2=1 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
  python: "3.10"
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  frequency: manual
  team: clusters
  cluster:
    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

  variations:
    - __suffix__: v1
      run:
        script: RAY_UP_enable_autoscaler_v2=0 python launch_and_verify_cluster.py aws/tests/aws_cluster.yaml --num-expected-nodes 2 --retries 10 --docker-override commit
    - __suffix__: v2
      run:
        script: RAY_UP_enable_autoscaler_v2=1 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
  python: "3.10"
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

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

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

  variations:
    - __suffix__: v1
      run:
        script: RAY_UP_enable_autoscaler_v2=0 python launch_and_verify_cluster.py aws/example-minimal.yaml
    - __suffix__: v2
      run:
        script: RAY_UP_enable_autoscaler_v2=1 python launch_and_verify_cluster.py aws/example-minimal.yaml

- name: aws_cluster_launcher_full
  python: "3.10"
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  frequency: nightly
  team: clusters
  cluster:
    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

  variations:
    - __suffix__: v1
      run:
        script: RAY_UP_enable_autoscaler_v2=0 python launch_and_verify_cluster.py aws/example-full.yaml --num-expected-nodes 2 --retries 20 --docker-override latest
    - __suffix__: v2
      run:
        script: RAY_UP_enable_autoscaler_v2=1 python launch_and_verify_cluster.py aws/example-full.yaml --num-expected-nodes 2 --retries 20 --docker-override latest

- name: gcp_cluster_launcher_minimal
  python: "3.10"
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

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

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

  variations:
    - __suffix__: v1
      run:
        script: RAY_UP_enable_autoscaler_v2=0 RAY_TRIM_AUTOSCALER_SSH_KEYS=1 python launch_and_verify_cluster.py gcp/example-minimal-pinned.yaml
    - __suffix__: v2
      run:
        script: RAY_UP_enable_autoscaler_v2=1 RAY_TRIM_AUTOSCALER_SSH_KEYS=1 python launch_and_verify_cluster.py gcp/example-minimal-pinned.yaml

- name: gcp_cluster_launcher_full
  python: "3.10"
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  env: gce
  frequency: nightly
  team: clusters
  cluster:
    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

  variations:
    - __suffix__: v1
      run:
        script: RAY_UP_enable_autoscaler_v2=0 RAY_TRIM_AUTOSCALER_SSH_KEYS=1 python launch_and_verify_cluster.py gcp/example-full.yaml --num-expected-nodes 2 --retries 30 --docker-override latest
    - __suffix__: v2
      run:
        script: RAY_UP_enable_autoscaler_v2=1 RAY_TRIM_AUTOSCALER_SSH_KEYS=1 python launch_and_verify_cluster.py gcp/example-full.yaml --num-expected-nodes 2 --retries 30 --docker-override latest

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

  env: gce
  frequency: manual
  team: clusters
  cluster:
    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

  variations:
    - __suffix__: v1
      run:
        script: RAY_UP_enable_autoscaler_v2=0 RAY_TRIM_AUTOSCALER_SSH_KEYS=1 python launch_and_verify_cluster.py gcp/example-full.yaml --num-expected-nodes 2 --retries 20  --docker-override latest
    - __suffix__: v2
      run:
        script: RAY_UP_enable_autoscaler_v2=1 RAY_TRIM_AUTOSCALER_SSH_KEYS=1 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
  python: "3.10"
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  env: gce
  frequency: manual
  team: clusters
  cluster:
    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

  variations:
    - __suffix__: v1
      run:
        script: RAY_UP_enable_autoscaler_v2=0 RAY_TRIM_AUTOSCALER_SSH_KEYS=1 python launch_and_verify_cluster.py gcp/example-full.yaml --num-expected-nodes 2 --retries 20  --docker-override nightly
    - __suffix__: v2
      run:
        script: RAY_UP_enable_autoscaler_v2=1 RAY_TRIM_AUTOSCALER_SSH_KEYS=1 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
  python: "3.10"
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  env: gce
  frequency: manual
  team: clusters
  cluster:
    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

  variations:
    - __suffix__: v1
      run:
        script: RAY_UP_enable_autoscaler_v2=0 RAY_TRIM_AUTOSCALER_SSH_KEYS=1 python launch_and_verify_cluster.py gcp/example-full.yaml --num-expected-nodes 2 --retries 20  --docker-override commit
    - __suffix__: v2
      run:
        script: RAY_UP_enable_autoscaler_v2=1 RAY_TRIM_AUTOSCALER_SSH_KEYS=1 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
  python: "3.10"
  group: cluster-launcher-test
  working_dir: ../python/ray/autoscaler/

  env: gce
  frequency: weekly
  team: clusters
  cluster:
    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

  variations:
    - __suffix__: v1
      run:
        script: RAY_UP_enable_autoscaler_v2=0 RAY_TRIM_AUTOSCALER_SSH_KEYS=1 python launch_and_verify_cluster.py gcp/example-gpu-docker.yaml
    - __suffix__: v2
      run:
        script: RAY_UP_enable_autoscaler_v2=1 RAY_TRIM_AUTOSCALER_SSH_KEYS=1 python launch_and_verify_cluster.py gcp/example-gpu-docker.yaml

- name: autoscaler_aws
  python: "3.10"
  group: autoscaler-test
  working_dir: autoscaling_tests

  frequency: nightly
  team: core

  cluster:
    cluster_compute: aws.yaml

  run:
    timeout: 1800
    script: python run.py


##############
# LLM Serve
##############

# Runs performance benchmark tests against vLLM service
- name: llm_serve_llama_3dot1_8B_quantized_tp_1
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
    anyscale_sdk_2026: true
    cluster_compute: llm_auto_select_worker.yaml
    # NOTE: Important for getting the correct secrets
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 3600
    long_running: false
    script: python run_llm_serve_test_and_bms.py --serve-config-file configs/serve_llama_3dot1_8b_quantized_tp1.yaml --run-vllm-profiler --run-serve-llm-profiler

- name: llm_serve_llama_3dot1_8B_quantized_tp_1_direct_streaming
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
      runtime_env:
        - RAY_SERVE_LLM_ENABLE_DIRECT_STREAMING=1
        - RAY_SERVE_ENABLE_HA_PROXY=1
        - RAY_SERVE_DEFAULT_HTTP_HOST=0.0.0.0
    anyscale_sdk_2026: true
    cluster_compute: llm_auto_select_worker.yaml
    # NOTE: Important for getting the correct secrets
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 3600
    long_running: false
    script: python run_llm_serve_test_and_bms.py --serve-config-file configs/serve_llama_3dot1_8b_quantized_tp1.yaml --run-vllm-profiler --run-serve-llm-profiler

# Direct streaming with the prefix-cache-aware request router. Guards #64326,
# where a body-aware router got no routing key on the direct-streaming path.
- name: llm_serve_llama_3dot1_8B_quantized_tp_1_direct_streaming_prefix_cache
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
      runtime_env:
        - RAY_SERVE_LLM_ENABLE_DIRECT_STREAMING=1
        - RAY_SERVE_ENABLE_HA_PROXY=1
        - RAY_SERVE_DEFAULT_HTTP_HOST=0.0.0.0
    anyscale_sdk_2026: true
    cluster_compute: llm_auto_select_worker.yaml
    # NOTE: Important for getting the correct secrets
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 3600
    long_running: false
    script: python run_llm_serve_test_and_bms.py --serve-config-file configs/serve_llama_3dot1_8b_quantized_tp1_prefix_cache.yaml

# Runs performance benchmark tests against vLLM service
- name: llm_serve_llama_3dot1_8B_tp_2
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
    anyscale_sdk_2026: true
    cluster_compute: llm_auto_select_worker.yaml
    # NOTE: Important for getting the correct secrets
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 3600
    long_running: false
    script: python run_llm_serve_test_and_bms.py --serve-config-file configs/serve_llama_3dot1_8b_tp2.yaml --run-vllm-profiler --run-serve-llm-profiler

- name: llm_serve_llama_3dot1_8B_tp_2_direct_streaming
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
      runtime_env:
        - RAY_SERVE_LLM_ENABLE_DIRECT_STREAMING=1
        - RAY_SERVE_ENABLE_HA_PROXY=1
        - RAY_SERVE_DEFAULT_HTTP_HOST=0.0.0.0
    anyscale_sdk_2026: true
    cluster_compute: llm_auto_select_worker.yaml
    # NOTE: Important for getting the correct secrets
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 3600
    long_running: false
    script: python run_llm_serve_test_and_bms.py --serve-config-file configs/serve_llama_3dot1_8b_tp2.yaml --run-vllm-profiler --run-serve-llm-profiler

- name: llm_serve_multi_node
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
      runtime_env:
        - RAY_SERVE_USE_COMPACT_SCHEDULING_STRATEGY=1 # For test_llm_serve_prefill_decode_with_data_parallelism
    anyscale_sdk_2026: true
    cluster_compute: llm_2x_4xl4.yaml

  run:
    timeout: 3600
    script: >
      pytest -sv test_llm_serve_multi_node_integration.py

- name: llm_serve_multi_node_direct_streaming
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
      runtime_env:
        - RAY_SERVE_USE_COMPACT_SCHEDULING_STRATEGY=1 # For test_llm_serve_prefill_decode_with_data_parallelism
        - RAY_SERVE_LLM_ENABLE_DIRECT_STREAMING=1
        - RAY_SERVE_ENABLE_HA_PROXY=1
        - RAY_SERVE_DEFAULT_HTTP_HOST=0.0.0.0
    anyscale_sdk_2026: true
    cluster_compute: llm_2x_4xl4.yaml

  run:
    timeout: 3600
    script: >
      pytest -sv test_llm_serve_multi_node_integration.py

- name: llm_serve_dp_engine_fault
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
    anyscale_sdk_2026: true
    cluster_compute: llm_2x_4xl4.yaml

  run:
    timeout: 3600
    script: >
      pytest -sv test_llm_serve_dp_engine_fault.py

- name: llm_serve_llama_3dot2_1B_no_accelerator
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
    anyscale_sdk_2026: true
    cluster_compute: llm_g5-4xlarge.yaml
    # NOTE: Important for getting the correct secrets
    cloud: serve_release_tests_cloud
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 3600
    long_running: false
    script: python run_llm_serve_test_and_bms.py --serve-config-file configs/serve_llama_3dot2_1b_no_accelerator.yaml --run-serve-llm-profiler


- name: llm_serve_llama_3dot1_8B_lora
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
    anyscale_sdk_2026: true
    cluster_compute: llm_auto_select_worker.yaml
    # NOTE: Important for getting the correct secrets
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 3600
    long_running: false
    script: python run_llm_serve_test_and_bms.py --serve-config-file configs/serve_llama_3dot1_8b_lora.yaml --timeout 900

- name: llm_serve_llama_3dot2_1B_s3
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
    anyscale_sdk_2026: true
    cluster_compute: llm_auto_select_worker.yaml
    # NOTE: Important for getting the correct secrets
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 3600
    long_running: false
    script: python run_llm_serve_test_and_bms.py --serve-config-file configs/serve_llama_3dot2_1b_s3.yaml --skip-hf-token true

- name: llm_serve_correctness
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
    anyscale_sdk_2026: true
    cluster_compute: llm_four_L4_gpu_head_node.yaml
    # NOTE: Important for getting the correct secrets
    cloud: serve_release_tests_cloud
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq
  run:
    timeout: 3600
    long_running: false
    script: pytest -vs test_llm_serve_correctness.py

- name: llm_serve_vllm_integration_tests
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
    anyscale_sdk_2026: true
    cluster_compute: llm_four_L4_gpu_head_node.yaml
    # NOTE: Important for getting the correct secrets
    cloud: serve_release_tests_cloud
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq
  run:
    timeout: 3600
    long_running: false
    script: pytest -vs test_llm_serve_integration.py test_llm_serve_fault_tolerance.py

- name: llm_serve_vllm_integration_tests_direct_streaming
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
      runtime_env:
        - RAY_SERVE_LLM_ENABLE_DIRECT_STREAMING=1
        - RAY_SERVE_ENABLE_HA_PROXY=1
    anyscale_sdk_2026: true
    cluster_compute: llm_four_L4_gpu_head_node.yaml
    # NOTE: Important for getting the correct secrets
    cloud: serve_release_tests_cloud
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq
  run:
    timeout: 3600
    long_running: false
    script: pytest -vs test_llm_serve_integration.py test_llm_serve_fault_tolerance.py

- name: llm_serve_sleep_wakeup
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
    anyscale_sdk_2026: true
    cluster_compute: llm_four_L4_gpu_head_node.yaml
    # NOTE: Important for getting the correct secrets
    cloud: serve_release_tests_cloud
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 3600
    long_running: false
    script: pytest -vs test_llm_serve_sleep_wakeup.py

- name: llm_serve_pause_resume
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
    anyscale_sdk_2026: true
    cluster_compute: llm_four_L4_gpu_head_node.yaml
    # NOTE: Important for getting the correct secrets
    cloud: serve_release_tests_cloud
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 3600
    long_running: false
    script: pytest -vs test_llm_serve_pause_resume.py

- name: llm_serve_rlhf
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
    anyscale_sdk_2026: true
    cluster_compute: llm_four_L4_gpu_head_node.yaml
    # NOTE: Important for getting the correct secrets
    cloud: serve_release_tests_cloud
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 3600
    long_running: false
    script: python test_llm_serve_rlhf.py

- name: llm_serve_llama_3dot1_8B_quantized_tp1_1p1d
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
    anyscale_sdk_2026: true
    cluster_compute: llm_auto_select_worker.yaml
    # NOTE: Important for getting the correct secrets
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 3600
    long_running: false
    script: python run_llm_serve_test_and_bms.py --serve-config-file configs/serve_llama_3dot1_8b_quantized_tp1_1p1d.yaml --skip-hf-token true

- name: llm_serve_llama_3dot1_8B_quantized_tp1_1p1d_direct_streaming
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
      runtime_env:
        - RAY_SERVE_LLM_ENABLE_DIRECT_STREAMING=1
        - RAY_SERVE_ENABLE_HA_PROXY=1
        - RAY_SERVE_DEFAULT_HTTP_HOST=0.0.0.0
    anyscale_sdk_2026: true
    cluster_compute: llm_auto_select_worker.yaml
    # NOTE: Important for getting the correct secrets
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 3600
    long_running: false
    script: python run_llm_serve_test_and_bms.py --serve-config-file configs/serve_llama_3dot1_8b_quantized_tp1_1p1d.yaml --skip-hf-token true

- name: llm_serve_llama_3dot1_8B_quantized_tp1_2p6d
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
    anyscale_sdk_2026: true
    cluster_compute: llm_auto_select_worker.yaml
    # NOTE: Important for getting the correct secrets
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 3600
    long_running: false
    script: python run_llm_serve_test_and_bms.py --serve-config-file configs/serve_llama_3dot1_8b_quantized_tp1_2p6d.yaml --skip-hf-token true

- name: llm_serve_llama_3dot1_8B_quantized_tp1_2p6d_direct_streaming
  frequency: nightly
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
      runtime_env:
        - RAY_SERVE_LLM_ENABLE_DIRECT_STREAMING=1
        - RAY_SERVE_ENABLE_HA_PROXY=1
        - RAY_SERVE_DEFAULT_HTTP_HOST=0.0.0.0
    anyscale_sdk_2026: true
    cluster_compute: llm_auto_select_worker.yaml
    # NOTE: Important for getting the correct secrets
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 3600
    long_running: false
    script: python run_llm_serve_test_and_bms.py --serve-config-file configs/serve_llama_3dot1_8b_quantized_tp1_2p6d.yaml --skip-hf-token true

- name: llm_serve_llama_3dot1_8B_quantized_tp1_2p6d_lmcache
  frequency: manual # todo(ray-llm): fix this test with new/old lmcache version and new vllm version and re-enable it.
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
      post_build_script: byod_llm_lmcache_test.sh
    anyscale_sdk_2026: true
    cluster_compute: llm_auto_select_worker.yaml
    # NOTE: Important for getting the correct secrets
    cloud_id: cld_wy5a6nhazplvu32526ams61d98
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 3600
    long_running: false
    script: python run_llm_serve_test_and_bms.py --serve-config-file configs/serve_llama_3dot1_8b_quantized_tp1_2p6d_lmcache.yaml --skip-hf-token true

- name: llm_serve_sglang_e2e
  frequency: manual
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/serve

  cluster:
    byod:
      type: llm-cu130
      runtime_env:
        - RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES=1
      post_build_script: byod_llm_sglang_test.sh
    anyscale_sdk_2026: true
    cluster_compute: llm_four_L4_gpu_head_node.yaml

  run:
    timeout: 3600
    script: pytest -vs test_llm_serve_sglang.py

- name: llm_serve_kv_router
  frequency: manual
  python: "3.12"
  group: llm-serve
  team: llm
  working_dir: llm_tests/kv_router_test

  cluster:
    byod:
      type: llm-cu130
      runtime_env:
        - RAY_SERVE_LLM_ENABLE_DIRECT_STREAMING=1
        - RAY_SERVE_ENABLE_HA_PROXY=1
        - RAY_SERVE_DEFAULT_HTTP_HOST=0.0.0.0
      post_build_script: byod_llm_kv_router_test.sh
    anyscale_sdk_2026: true
    # Kept in this working_dir rather than referenced across dirs
    # so it ships in the test's runfiles.
    cluster_compute: llm_4xl4.yaml

  run:
    timeout: 3600
    script: pytest -vs .

##############
# LLM Batch
##############

- name: llm_batch_vllm
  frequency: nightly
  python: "3.12"
  group: llm-batch
  team: llm
  working_dir: llm_tests/batch

  cluster:
    byod:
      type: llm-cu130
    anyscale_sdk_2026: true
    cluster_compute: llm_4xl4.yaml
    # NOTE: Important for getting the correct secrets
    cloud: serve_release_tests_cloud
    project_id: prj_lhlrf1u5yv8qz9qg3xzw8fkiiq

  run:
    timeout: 3600
    script: >
      pytest -sv test_batch_vllm.py

- name: llm_batch_sglang_llama
  frequency: nightly
  python: "3.12"
  group: llm-batch
  team: llm
  working_dir: llm_tests/batch

  cluster:
    byod:
      type: llm-cu130
      post_build_script: byod_llm_sglang_test.sh
    anyscale_sdk_2026: true
    cluster_compute: llm_4xl4.yaml

  run:
    timeout: 3600
    script: >
      pytest -sv test_batch_sglang.py

- name: llm_batch_vllm_multi_node
  frequency: nightly
  python: "3.12"
  group: llm-batch
  team: llm
  working_dir: llm_tests/batch

  cluster:
    byod:
      type: llm-cu130
    anyscale_sdk_2026: true
    cluster_compute: llm_2x_4xl4.yaml

  run:
    timeout: 3600
    script: >
      pytest -sv test_batch_multi_node_vllm.py

- name: llm_batch_single_node_baseline_benchmark
  python: "3.12"
  group: llm-batch
  working_dir: llm_tests/batch
  frequency: nightly
  team: llm

  cluster:
    byod:
      runtime_env:
        - VLLM_DISABLE_COMPILE_CACHE=1
      type: llm-cu130
      post_build_script: byod_llm_single_node_baseline_benchmark.sh
      python_depset: llm_batch/llm_batch_single_node_benchmark_py312_cu130.lock
    anyscale_sdk_2026: true
    cluster_compute: llm_single_node_benchmark_l4.yaml

  run:
    timeout: 3600
    script: pytest -v test_batch_single_node_vllm.py::test_single_node_baseline_benchmark


#######################
# Ray examples tests
#######################


- name: video_analysis_pipeline  # do not use dashes (regex sensitive)
  frequency: weekly
  python: "3.10"
  group: ray-examples
  team: serve
  working_dir: //doc/source/serve/tutorials/video-analysis

  cluster:
    byod:
      type: cpu
      post_build_script: byod_serve_video_analysis.sh
    cluster_compute: ci/aws.yaml

  run:
    timeout: 3600
    script: bash ci/tests.sh  # relative to working_dir

  variations:
    - __suffix__: aws  # uses default specs above
    # probably wont work on gce as tests depend on s3 access
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: ci/gce.yaml  # relative to working_dir

- name: transformers_reinforcement_learning  # do not use dashes (regex sensitive)
  python: "3.10"
  frequency: weekly
  group: ray-examples
  team: ml
  working_dir: //doc/source/train/examples/transformers/transformer_reinforcement_learning

  cluster:
    byod:
      type: gpu
      post_build_script: byod_huggingface_trl.sh  # release/ray_release/byod/
    cluster_compute: ci/aws.yaml  # relative to working_dir

  run:
    timeout: 3600
    script: bash ci/tests.sh  # relative to working_dir

  variations:
    - __suffix__: aws
    - __suffix__: gce
      env: gce
      frequency: manual
      cluster:
        cluster_compute: ci/gce.yaml  # relative to working_dir