[Release/Horovod] Add user test for Horovod (#19661)

* infra

* wip

* add test

* typo

* typo

* update

* rename

* fix

* full path

* formatting

* reorder

* update

* update

* Update release/horovod_tests/workloads/horovod_user_test.py

Co-authored-by: matthewdeng <matthew.j.deng@gmail.com>

* bump num_workers

* update installs

* try

* add pip_packages

* min_workers

* fix

* bump pg timeout

* Fix symlink

* fix

* fix

* cmake

* fix

* pin filelock

* final

* update

* fix

* Update release/horovod_tests/workloads/horovod_user_test.py

* fix

* fix

* separate compute template

* test latest and master

Co-authored-by: matthewdeng <matthew.j.deng@gmail.com>

python/ray/tune/examples/horovod_cifar_pbt_example.py

@@ -1 +1 @@
-../../../../release/horovod_tests/workloads/horovod_test.py
+../../../../release/horovod_tests/workloads/horovod_tune_test.py

python/ray/util/horovod/horovod_example.py

@@ -115,11 +115,20 @@ def train_fn(data_dir=None,
                     100. * batch_idx / len(train_loader), loss.item()))
 
 
-def main(num_workers, use_gpu, **kwargs):
-    settings = RayExecutor.create_settings(timeout_s=30)
+def main(num_workers,
+         use_gpu,
+         timeout_s=30,
+         placement_group_timeout_s=100,
+         kwargs=None):
+    kwargs = kwargs or {}
+    if use_gpu:
+        kwargs["use_cuda"] = True
+    settings = RayExecutor.create_settings(
+        timeout_s=timeout_s,
+        placement_group_timeout_s=placement_group_timeout_s)
     executor = RayExecutor(settings, use_gpu=use_gpu, num_workers=num_workers)
     executor.start()
-    executor.run(train_fn, **kwargs)
+    executor.run(train_fn, kwargs=kwargs)
 
 
 if __name__ == "__main__":

release/.buildkite/build_pipeline.py

@@ -20,7 +20,12 @@ import yaml
 
 
 class ReleaseTest:
-    def __init__(self, name: str, smoke_test: bool = False, retry: int = 0):
+    def __init__(
+            self,
+            name: str,
+            smoke_test: bool = False,
+            retry: int = 0,
+    ):
         self.name = name
         self.smoke_test = smoke_test
         self.retry = retry
@@ -243,6 +248,19 @@ MANUAL_TESTS = {
     ],
 }
 
+HOROVOD_INSTALL_ENV_VARS = [
+    "HOROVOD_WITH_GLOO", "HOROVOD_WITHOUT_MPI", "HOROVOD_WITHOUT_TENSORFLOW",
+    "HOROVOD_WITHOUT_MXNET", "HOROVOD_WITH_PYTORCH"
+]
+
+HOROVOD_SETUP_COMMANDS = [
+    "sudo apt update", "sudo apt -y install build-essential",
+    "pip install cmake"
+] + [
+    f"export {horovod_env_var}=1"
+    for horovod_env_var in HOROVOD_INSTALL_ENV_VARS
+]
+
 # This test suite holds "user" tests to test important user workflows
 # in a particular environment.
 # All workloads in this test suite should:
@@ -251,6 +269,17 @@ MANUAL_TESTS = {
 #   3. Use GPUs if applicable
 #   4. Have the `use_connect` flag set.
 USER_TESTS = {
+    "~/ray/release/horovod_tests/horovod_tests.yaml": [
+        ConnectTest(
+            "horovod_user_test_latest",
+            setup_commands=HOROVOD_SETUP_COMMANDS,
+            requirements_file="release/horovod_tests/driver_requirements.txt"),
+        ConnectTest(
+            "horovod_user_test_master",
+            setup_commands=HOROVOD_SETUP_COMMANDS,
+            requirements_file="release/horovod_tests"
+            "/driver_requirements_master.txt")
+    ],
     "~/ray/release/train_tests/train_tests.yaml": [
         ConnectTest(
             "train_tensorflow_mnist_test",
@@ -260,7 +289,7 @@ USER_TESTS = {
             "train_torch_linear_test",
             requirements_file="release/train_tests"
             "/driver_requirements.txt")
-    ]
+    ],
 }
 
 SUITES = {
@@ -484,22 +513,21 @@ def create_test_step(
             }]
         }
 
-    step_conf["commands"] = [
-        "pip install -q -r release/requirements.txt",
-        "pip install -U boto3 botocore",
-        f"git clone -b {ray_test_branch} {ray_test_repo} ~/ray", cmd,
-        "sudo cp -rf /tmp/artifacts/* /tmp/ray_release_test_artifacts "
-        "|| true"
-    ]
-
     if isinstance(test_name, ConnectTest):
         # Add driver side setup commands to the step.
         pip_requirements_command = [f"pip install -U -r "
                                     f"{test_name.requirements_file}"] if \
             test_name.requirements_file else []
         step_conf["commands"] = test_name.setup_commands \
-            + pip_requirements_command \
-            + step_conf["commands"]
+            + pip_requirements_command
+
+    step_conf["commands"] += [
+        "pip install -q -r release/requirements.txt",
+        "pip install -U boto3 botocore",
+        f"git clone -b {ray_test_branch} {ray_test_repo} ~/ray", cmd,
+        "sudo cp -rf /tmp/artifacts/* /tmp/ray_release_test_artifacts "
+        "|| true"
+    ]
 
     step_conf["label"] = (
         f"{test_name} "

release/horovod_tests/app_config.yaml

@@ -14,7 +14,7 @@ post_build_cmds:
   - sudo rm -rf /home/ray/anaconda3/lib/python3.7/site-packages/numpy
   - pip3 install numpy || true
   - pip3 install -U {{ env["RAY_WHEELS"] | default("ray") }}
-  - pip3 install 'ray[rllib]'
+  - pip3 install 'ray[tune]'
   - pip3 install torch torchvision
-  - HOROVOD_WITH_GLOO=1 HOROVOD_WITHOUT_MPI=1 HOROVOD_WITHOUT_TENSORFLOW=1 HOROVOD_WITHOUT_MXNET=1 HOROVOD_WITH_PYTORCH=1 pip3 install -U git+https://github.com/horovod/horovod.git
+  - HOROVOD_WITH_GLOO=1 HOROVOD_WITHOUT_MPI=1 HOROVOD_WITHOUT_TENSORFLOW=1 HOROVOD_WITHOUT_MXNET=1 HOROVOD_WITH_PYTORCH=1 pip3 install -U horovod
   - {{ env["RAY_WHEELS_SANITY_CHECK"] | default("echo No Ray wheels sanity check") }}

release/horovod_tests/app_config_master.yaml

@@ -0,0 +1,20 @@
+base_image: "anyscale/ray-ml:pinned-nightly-py37-gpu"
+env_vars: {}
+debian_packages:
+  - curl
+
+python:
+  pip_packages:
+    - pytest
+    - awscli
+  conda_packages: []
+
+post_build_cmds:
+  - pip uninstall -y numpy ray || true
+  - sudo rm -rf /home/ray/anaconda3/lib/python3.7/site-packages/numpy
+  - pip3 install numpy || true
+  - pip3 install -U {{ env["RAY_WHEELS"] | default("ray") }}
+  - pip3 install 'ray[tune]'
+  - pip3 install torch torchvision
+  - HOROVOD_WITH_GLOO=1 HOROVOD_WITHOUT_MPI=1 HOROVOD_WITHOUT_TENSORFLOW=1 HOROVOD_WITHOUT_MXNET=1 HOROVOD_WITH_PYTORCH=1 pip3 install -U git+https://github.com/horovod/horovod.git
+  - {{ env["RAY_WHEELS_SANITY_CHECK"] | default("echo No Ray wheels sanity check") }}

release/horovod_tests/base_driver_reqs.txt

@@ -0,0 +1,8 @@
+# Make sure the driver versions are the same as cluster versions.
+# The cluster uses ray-ml Docker image.
+# ray-ml Docker image installs dependencies from ray/python/requirements/ml/ directory.
+# We constrain on these requirements file so that the same versions are installed.
+-c ../../python/requirements/ml/requirements_dl.txt
+
+torch
+torchvision

release/horovod_tests/compute_tpl.yaml

@@ -10,8 +10,8 @@ head_node_type:
 worker_node_types:
     - name: worker_node
       instance_type: g3.8xlarge
-      min_workers: 3
       max_workers: 3
+      min_workers: 3
       use_spot: false
 
 aws:

release/horovod_tests/compute_tpl_autoscaling.yaml

@@ -0,0 +1,24 @@
+cloud_id: {{env["ANYSCALE_CLOUD_ID"]}}
+region: us-west-2
+
+max_workers: 3
+
+head_node_type:
+    name: head_node
+    instance_type: g3.8xlarge
+
+worker_node_types:
+    - name: worker_node
+      instance_type: g3.8xlarge
+      max_workers: 3
+      min_workers: 0
+      use_spot: false
+
+aws:
+  TagSpecifications:
+    - ResourceType: "instance"
+      Tags:
+        - Key: anyscale-user
+          Value: '{{env["ANYSCALE_USER"]}}'
+        - Key: anyscale-expiration
+          Value: '{{env["EXPIRATION_1D"]}}'

release/horovod_tests/driver_requirements.txt

@@ -0,0 +1,3 @@
+-r ./base_driver_reqs.txt
+
+horovod

release/horovod_tests/driver_requirements_master.txt

@@ -0,0 +1,4 @@
+-r ./base_driver_reqs.txt
+
+# Horovod master.
+git+https://github.com/horovod/horovod.git

release/horovod_tests/horovod_tests.yaml

@@ -1,6 +1,6 @@
 - name: horovod_test
   cluster:
-    app_config: app_config.yaml
+    app_config: app_config_master.yaml
     compute_template: compute_tpl.yaml
 
   run:
@@ -12,3 +12,25 @@
   smoke_test:
     run:
       timeout: 1800
+
+- name: horovod_user_test_latest
+  cluster:
+    app_config: app_config.yaml
+    compute_template: compute_tpl_autoscaling.yaml
+
+  run:
+    use_connect: True
+    autosuspend_mins: 10
+    timeout: 1200
+    script: python workloads/horovod_user_test.py
+
+- name: horovod_user_test_master
+  cluster:
+    app_config: app_config_master.yaml
+    compute_template: compute_tpl_autoscaling.yaml
+
+  run:
+    use_connect: True
+    autosuspend_mins: 10
+    timeout: 1200
+    script: python workloads/horovod_user_test.py

release/horovod_tests/workloads/horovod_user_test.py

@@ -0,0 +1,33 @@
+import json
+import os
+import time
+
+import ray
+from ray.util.horovod.horovod_example import main
+
+if __name__ == "__main__":
+    start = time.time()
+
+    addr = os.environ.get("RAY_ADDRESS")
+    job_name = os.environ.get("RAY_JOB_NAME", "horovod_user_test")
+    if addr is not None and addr.startswith("anyscale://"):
+        ray.init(address=addr, job_name=job_name)
+    else:
+        ray.init(address="auto")
+
+    main(
+        num_workers=6,
+        use_gpu=True,
+        placement_group_timeout_s=900,
+        kwargs={"num_epochs": 20})
+
+    taken = time.time() - start
+    result = {
+        "time_taken": taken,
+    }
+    test_output_json = os.environ.get("TEST_OUTPUT_JSON",
+                                      "/tmp/horovod_user_test.json")
+    with open(test_output_json, "wt") as f:
+        json.dump(result, f)
+
+    print("Test Successful!")