[RLlib] Enhancements for multi-node/multi-GPU training and better EnvRunner error msg.

rllib/env/multi_agent_env_runner.py

@@ -828,16 +828,22 @@ def make_env(self):
                 remote=self.config.remote_worker_envs,
             )
 
+        # No env provided -> Error.
+        if not self.config.env:
+            raise ValueError(
+                "`config.env` is not provided! You should provide a valid environment "
+                "to your config through `config.environment([env descriptor e.g. "
+                "'CartPole-v1'])`."
+            )
         # Register env for the local context.
         # Note, `gym.register` has to be called on each worker.
-        if isinstance(self.config.env, str) and _global_registry.contains(
+        elif isinstance(self.config.env, str) and _global_registry.contains(
             ENV_CREATOR, self.config.env
         ):
             entry_point = partial(
                 _global_registry.get(ENV_CREATOR, self.config.env),
                 env_ctx,
             )
-
         else:
             entry_point = partial(
                 _gym_env_creator,

rllib/env/single_agent_env_runner.py

@@ -794,16 +794,22 @@ def make_env(self) -> None:
                 remote=self.config.remote_worker_envs,
             )
 
+        # No env provided -> Error.
+        if not self.config.env:
+            raise ValueError(
+                "`config.env` is not provided! You should provide a valid environment "
+                "to your config through `config.environment([env descriptor e.g. "
+                "'CartPole-v1'])`."
+            )
         # Register env for the local context.
         # Note, `gym.register` has to be called on each worker.
-        if isinstance(self.config.env, str) and _global_registry.contains(
+        elif isinstance(self.config.env, str) and _global_registry.contains(
             ENV_CREATOR, self.config.env
         ):
             entry_point = partial(
                 _global_registry.get(ENV_CREATOR, self.config.env),
                 env_ctx,
             )
-
         else:
             entry_point = partial(
                 _gym_env_creator,

rllib/utils/test_utils.py

@@ -1392,15 +1392,18 @@ def run_rllib_example_script_experiment(
         # Define compute resources used automatically (only using the --num-gpus arg).
         # New stack.
         if config.enable_rl_module_and_learner:
+            # Do we have GPUs available in the cluster?
+            num_gpus = ray.cluster_resources().get("GPU", 0)
+            if args.num_gpus > 0 and num_gpus < args.num_gpus:
+                raise ValueError(
+                    f"You are running your script with --num-gpus={args.num_gpus}, "
+                    "but your cluster only has {resources['GPU']} GPUs! Will "
+                )
             # Define compute resources used.
             config.resources(num_gpus=0)
             config.learners(
                 num_learners=args.num_gpus,
-                num_gpus_per_learner=(
-                    1
-                    if torch and torch.cuda.is_available() and args.num_gpus > 0
-                    else 0
-                ),
+                num_gpus_per_learner=1 if num_gpus >= args.num_gpus > 0 else 0,
             )
             config.resources(num_gpus=0)
         # Old stack.