[RLlib] Don't add a cpu to bundle for learner when using gpu (ray-pro…

…ject#35529)

solves ray-project#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 <[email protected]>

release/release_tests.yaml

@@ -3465,6 +3465,37 @@
  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

release/rllib_tests/multi_node_checkpointing_compute_config.yaml

@@ -5,7 +5,7 @@ max_workers: 3
 
 head_node_type:
  name: head_node
- instance_type: m5.2xlarge
+ instance_type: m5.xlarge
 
 worker_node_types:
  - name: worker_node

release/rllib_tests/multi_node_checkpointing_compute_config_gce.yaml

@@ -7,7 +7,7 @@ max_workers: 3
 
 head_node_type:
  name: head_node
- instance_type: n2-standard-8 # m5.2xlarge
+ instance_type: n2-standard-4 # m5.xlarge
 
 worker_node_types:
  - name: worker_node

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__]))

rllib/algorithms/algorithm.py

@@ -2272,13 +2272,22 @@ def default_resource_request(
  # resources for remote learner workers
  learner_bundles = []
  if cf._enable_learner_api and cf.num_learner_workers > 0:
- learner_bundles = [
- {
- "CPU": cf.num_cpus_per_learner_worker,
- "GPU": cf.num_gpus_per_learner_worker,
- }
- for _ in range(cf.num_learner_workers)
- ]
+ # can't specify cpus for learner workers at the same
+ # time as gpus
+ if cf.num_gpus_per_learner_worker:
+ learner_bundles = [
+ {
+ "GPU": cf.num_gpus_per_learner_worker,
+ }
+ for _ in range(cf.num_learner_workers)
+ ]
+ elif cf.num_cpus_per_learner_worker:
+ learner_bundles = [
+ {
+ "CPU": cf.num_cpus_per_learner_worker,
+ }
+ for _ in range(cf.num_learner_workers)
+ ]
 
  bundles = [driver] + rollout_bundles + evaluation_bundles + learner_bundles
 

rllib/algorithms/algorithm_config.py

@@ -887,6 +887,20 @@ def validate(self) -> None:
  "via `config.training(_enable_learner_api=True)` (or set both to "
  "False)."
  )
+ # TODO @Avnishn: This is a short-term work around due to
+ # https://github.com/ray-project/ray/issues/35409
+ # Remove this once we are able to specify placement group bundle index in RLlib
+ if (
+ self.num_cpus_per_learner_worker > 1
+ and self.num_gpus_per_learner_worker > 0
+ ):
+ raise ValueError(
+ "Cannot set both `num_cpus_per_learner_worker` and "
+ " `num_gpus_per_learner_worker` > 0! Users must set one"
+ " or the other due to issues with placement group"
+ " fragmentation. See "
+ "https://github.com/ray-project/ray/issues/35409 for more details."
+ )
 
  if bool(os.environ.get("RLLIB_ENABLE_RL_MODULE", False)):
  # Enable RLModule API and connectors if env variable is set
@@ -1149,7 +1163,8 @@ def resources(
  num_gpus_per_learner_worker: Number of GPUs allocated per worker. If
  `num_learner_workers = 0`, any value greater than 0 will run the
  training on a single GPU on the head node, while a value of 0 will run
- the training on head node CPU cores.
+ the training on head node CPU cores. If num_gpus_per_learner_worker is
+ set, then num_cpus_per_learner_worker cannot be set.
  local_gpu_idx: if num_gpus_per_worker > 0, and num_workers<2, then this gpu
  index will be used for training. This is an index into the available
  cuda devices. For example if os.environ["CUDA_VISIBLE_DEVICES"] = "1"
@@ -3227,7 +3242,11 @@ def get_learner_group_config(self, module_spec: ModuleSpec) -> LearnerGroupConfi
  )
  .resources(
  num_learner_workers=self.num_learner_workers,
- num_cpus_per_learner_worker=self.num_cpus_per_learner_worker,
+ num_cpus_per_learner_worker=(
+ self.num_cpus_per_learner_worker
+ if not self.num_gpus_per_learner_worker
+ else 0
+ ),
  num_gpus_per_learner_worker=self.num_gpus_per_learner_worker,
  local_gpu_idx=self.local_gpu_idx,
  )

rllib/core/learner/learner.py

@@ -588,8 +588,6 @@ def compile_results(
  loss_per_module: A dict mapping module IDs (including ALL_MODULES) to the
  individual loss tensors as returned by calls to
  `compute_loss_for_module(module_id=...)`.
- postprocessed_gradients: The postprocessed gradients dict, (flat) mapping
- gradient tensor refs to the already postprocessed gradient tensors.
  metrics_per_module: The collected metrics defaultdict mapping ModuleIDs to
  metrics dicts. These metrics are collected during loss- and
  gradient computation, gradient postprocessing, and gradient application.

rllib/examples/learner/multi_agent_cartpole_ppo.py

@@ -95,10 +95,10 @@ def policy_mapping_fn(agent_id, episode, worker, **kwargs):
 
  config = (
  PPOConfig()
- .rollouts(rollout_fragment_length=500)
+ .rollouts(rollout_fragment_length="auto", num_rollout_workers=3)
  .environment(MultiAgentCartPole, env_config={"num_agents": args.num_agents})
  .framework(args.framework)
- .training(num_sgd_iter=10)
+ .training(num_sgd_iter=10, sgd_minibatch_size=2**9, train_batch_size=2**12)
  .multi_agent(policies=policies, policy_mapping_fn=policy_mapping_fn)
  .rl_module(_enable_rl_module_api=True)
  .training(_enable_learner_api=True)
@@ -115,7 +115,7 @@ def policy_mapping_fn(agent_id, episode, worker, **kwargs):
  results = tune.Tuner(
  "PPO",
  param_space=config.to_dict(),
- run_config=air.RunConfig(stop=stop, verbose=1),
+ run_config=air.RunConfig(stop=stop, verbose=3),
  ).fit()
 
  if args.as_test: