From a8108bda3ae1666d07861719c8943053f09f4856 Mon Sep 17 00:00:00 2001
From: Avnish Narayan <38871737+avnishn@users.noreply.github.com>
Date: Wed, 24 May 2023 15:35:22 -0700
Subject: [PATCH] [RLlib] Don't add a cpu to bundle for learner when using gpu
 (#35529) (#35676)

solves https://github.com/ray-project/ray/issues/35409

Prevent fragmentation of resources by not placing gpus
with cpus in bundles for the learner workers, making it
so that an actor that requires only cpu does not
potentially take a bundle that has both a cpu and gpu.

The long term fix will be to allow the specification
of placement group bundle index via tune and ray train.

Signed-off-by: avnishn <avnishnarayan@gmail.com>
---
 release/release_tests.yaml                    |  31 +++++
 ..._test_basic_multi_node_training_learner.py | 115 ++++++++++++++++++
 2 files changed, 146 insertions(+)

diff --git a/release/release_tests.yaml b/release/release_tests.yaml
index 0cf56764e236db..e8f6f46c7c6b7d 100644
--- a/release/release_tests.yaml
+++ b/release/release_tests.yaml
@@ -3542,6 +3542,37 @@
       cluster:
         cluster_compute: multi_node_checkpointing_compute_config_gce.yaml
 
+- name: rllib_multi_node_e2e_training_smoke_test
+  group: RLlib tests
+  working_dir: rllib_tests
+
+  frequency: nightly
+  team: rllib
+
+  cluster:
+    cluster_env: app_config.yaml
+    cluster_compute: multi_node_checkpointing_compute_config.yaml
+
+  run:
+    timeout: 3600
+    script: pytest smoke_tests/smoke_test_basic_multi_node_training_learner.py
+
+    wait_for_nodes:
+      num_nodes: 3
+
+  alert: default
+
+  variations:
+    - __suffix__: aws
+    - __suffix__: gce
+      env: gce
+      frequency: manual
+      cluster:
+        cluster_env: app_config.yaml
+        cluster_compute: multi_node_checkpointing_compute_config_gce.yaml
+
+
+
 - name: rllib_learning_tests_a2c_tf
   group: RLlib tests
   working_dir: rllib_tests
diff --git a/release/rllib_tests/smoke_tests/smoke_test_basic_multi_node_training_learner.py b/release/rllib_tests/smoke_tests/smoke_test_basic_multi_node_training_learner.py
index e69de29bb2d1d6..f332d7b6b808b9 100644
--- a/release/rllib_tests/smoke_tests/smoke_test_basic_multi_node_training_learner.py
+++ b/release/rllib_tests/smoke_tests/smoke_test_basic_multi_node_training_learner.py
@@ -0,0 +1,115 @@
+import ray
+from ray import air, tune
+from ray.rllib.algorithms.ppo import PPOConfig
+
+
+def run_with_tuner_n_rollout_worker_2_gpu(config):
+    """Run training with n rollout workers and 2 learner workers with gpu."""
+    config = config.rollouts(num_rollout_workers=5)
+    tuner = tune.Tuner(
+        "PPO",
+        param_space=config,
+        run_config=air.RunConfig(
+            stop={"timesteps_total": 128},
+            failure_config=air.FailureConfig(fail_fast=True),
+        ),
+    )
+    tuner.fit()
+
+
+def run_with_tuner_0_rollout_worker_2_gpu(config):
+    """Run training with 0 rollout workers with 2 learner workers with gpu."""
+    config = config.rollouts(num_rollout_workers=0)
+    tuner = tune.Tuner(
+        "PPO",
+        param_space=config,
+        run_config=air.RunConfig(
+            stop={"timesteps_total": 128},
+            failure_config=air.FailureConfig(fail_fast=True),
+        ),
+    )
+    tuner.fit()
+
+
+def run_tuner_n_rollout_workers_0_gpu(config):
+    """Run training with n rollout workers, multiple learner workers, and no gpu."""
+    config = config.rollouts(num_rollout_workers=5)
+    config = config.resources(
+        num_cpus_per_learner_worker=1,
+        num_learner_workers=2,
+    )
+
+    tuner = tune.Tuner(
+        "PPO",
+        param_space=config,
+        run_config=air.RunConfig(
+            stop={"timesteps_total": 128},
+            failure_config=air.FailureConfig(fail_fast=True),
+        ),
+    )
+    tuner.fit()
+
+
+def run_tuner_n_rollout_workers_1_gpu_local(config):
+    """Run training with n rollout workers, local learner, and 1 gpu."""
+    config = config.rollouts(num_rollout_workers=5)
+    config = config.resources(
+        num_gpus_per_learner_worker=1,
+        num_learner_workers=0,
+    )
+
+    tuner = tune.Tuner(
+        "PPO",
+        param_space=config,
+        run_config=air.RunConfig(
+            stop={"timesteps_total": 128},
+            failure_config=air.FailureConfig(fail_fast=True),
+        ),
+    )
+    tuner.fit()
+
+
+def test_multi_node_training_smoke():
+    """A smoke test to see if we can run multi node training without pg problems.
+
+    This test is run on a 3 node cluster. The head node is a m5.xlarge (4 cpu),
+    the worker nodes are 2 g4dn.xlarge (1 gpu, 4 cpu) machines.
+
+    """
+
+    ray.init()
+
+    config = (
+        PPOConfig()
+        .training(
+            _enable_learner_api=True,
+            model={
+                "fcnet_hiddens": [256, 256, 256],
+                "fcnet_activation": "relu",
+                "vf_share_layers": True,
+            },
+            train_batch_size=128,
+        )
+        .rl_module(_enable_rl_module_api=True)
+        .environment("CartPole-v1")
+        .resources(
+            num_gpus_per_learner_worker=1,
+            num_learner_workers=2,
+        )
+        .rollouts(num_rollout_workers=2)
+        .reporting(min_time_s_per_iteration=0, min_sample_timesteps_per_iteration=10)
+    )
+    for fw in ["tf2", "torch"]:
+        config = config.framework(fw, eager_tracing=True)
+
+        run_with_tuner_0_rollout_worker_2_gpu(config)
+        run_with_tuner_n_rollout_worker_2_gpu(config)
+        run_tuner_n_rollout_workers_0_gpu(config)
+        run_tuner_n_rollout_workers_1_gpu_local(config)
+
+
+if __name__ == "__main__":
+    import sys
+    import pytest
+
+    sys.exit(pytest.main(["-v", __file__]))