[rllib][windows] CUDA error when work is distributed

When I run my program in the distributed setting, I am getting a CUBLAS_STATUS_ALLOC_FAILED error. Specifically, I am instantiating several PPOTrainer objects and wanting them to train in parallel. My setup is similar in spirit to population-based training in RL (in fact, It would probably be good for me follow how you all have implemented PBT, but I was having trouble finding that.)

Some googling around is telling me that this error is either coming from either:

  1. trying to index into a tensor with a bad value or
  2. memory allocation error on the GPU (1060)

However, when I initialize Ray, I am doing so with ray.init(num_gpus=0) I have 64GB of RAM on my machine, so that shouldn’t be the problem as the networks have ~10k params each.

When I put ray into local_mode to look for bugs the error disappeared which makes me think the problem is case 2.

So, I wanted to follow-up on this. I ran my program again, but was watching my GPU usage the whole time.

Despite running with ray.init(num_gpus=0), my models are being sent to the GPU. As soon as the GPU Memory hits 98%, the program crashes with a CUBLAS_STATUS_ALLOC_FAILED error.

If I run the program in local_mode the GPU utilization never goes above 20% and the program completes. But that’s still bad behavior because it’s using the GPU at all when I explicitly told ray that the number of GPUs was zero.

Hey @aadharna , hmm, I thought that in local_mode, the model would not be moved to the GPU, but I could be wrong or this could be a Win problem.

Could you post your ray version as well as your particular PPO config? We changed the logic quite a lot in the past as to when the model is actually moved to the GPU (when a GPU is available vs when num_gpus > 0, etc…).

For sure! I can also provide additional code as well if desired. One thing to note is that I have several ppoTrainers (/other trainers depending on a flag – impalaTrainer).

ray version: 1.2.0
So, I am also using Chris Bamford’s Griddly, just FYI.
Config: This almost entirely defaults –

{'_fake_gpus': False,
 '_use_trajectory_view_api': True,
 'batch_mode': 'truncate_episodes',
 'callbacks': <class 'ray.rllib.agents.callbacks.DefaultCallbacks'>,
 'clip_actions': True,
 'clip_param': 0.3,
 'clip_rewards': None,
 'collect_metrics_timeout': 180,
 'compress_observations': False,
 'create_env_on_driver': False,
 'custom_eval_function': None,
 'custom_resources_per_worker': {},
 'eager_tracing': False,
 'entropy_coeff': 0.0,
 'entropy_coeff_schedule': None,
 'env': 'limited_zelda_custom',
 'env_config': {'environment_name': 'limited_zelda_custom',
                'global_observer_type': <ObserverType.VECTOR: 4>,
                'level_string': DYNAMICALLY_ADDED_IN,
                'max_steps': 500,
                'player_observer_type': <ObserverType.VECTOR: 4>,
                'random_level_on_reset': False,
                'record_video_config': {'directory': '.\\videos',
                                        'frequency': 10000000},
                'yaml_file': 'levels\\limited_zelda.yaml'},
 'evaluation_config': {},
 'evaluation_interval': None,
 'evaluation_num_episodes': 10,
 'evaluation_num_workers': 0,
 'exploration_config': {'type': 'StochasticSampling'},
 'explore': True,
 'extra_python_environs_for_driver': {},
 'extra_python_environs_for_worker': {},
 'fake_sampler': False,
 'framework': 'torch',
 'gamma': 0.99,
 'grad_clip': None,
 'horizon': None,
 'ignore_worker_failures': False,
 'in_evaluation': False,
 'input': 'sampler',
 'input_evaluation': ['is', 'wis'],
 'kl_coeff': 0.2,
 'kl_target': 0.01,
 'lambda': 1.0,
 'local_tf_session_args': {'inter_op_parallelism_threads': 8,
                           'intra_op_parallelism_threads': 8},
 'log_level': 'WARN',
 'log_sys_usage': True,
 'logger_config': None,
 'lr': 5e-05,
 'lr_schedule': None,
 'memory': 0,
 'memory_per_worker': 0,
 'metrics_smoothing_episodes': 100,
 'min_iter_time_s': 0,
 'model': {'custom_model': 'AIIDE_PINSKY_MODEL', 'custom_model_config': {}},
 'monitor': False,
 'multiagent': {'count_steps_by': 'env_steps',
                'observation_fn': None,
                'policies': {},
                'policies_to_train': None,
                'policy_mapping_fn': None,
                'replay_mode': 'independent'},
 'no_done_at_end': False,
 'normalize_actions': False,
 'num_cpus_for_driver': 1,
 'num_cpus_per_worker': 1,
 'num_envs_per_worker': 1,
 'num_gpus': 0,
 'num_gpus_per_worker': 0,
 'num_sgd_iter': 30,
 'num_workers': 1,
 'object_store_memory': 0,
 'object_store_memory_per_worker': 0,
 'observation_filter': 'NoFilter',
 'optimizer': {},
 'output': None,
 'output_compress_columns': ['obs', 'new_obs'],
 'output_max_file_size': 67108864,
 'postprocess_inputs': False,
 'preprocessor_pref': 'deepmind',
 'remote_env_batch_wait_ms': 0,
 'remote_worker_envs': False,
 'replay_sequence_length': 1,
 'rollout_fragment_length': 200,
 'sample_async': False,
 'sample_collector': <class 'ray.rllib.evaluation.collectors.simple_list_collector.SimpleListCollector'>,
 'seed': None,
 'sgd_minibatch_size': 128,
 'shuffle_buffer_size': 0,
 'shuffle_sequences': True,
 'simple_optimizer': False,
 'soft_horizon': False,
 'synchronize_filters': True,
 'tf_session_args': {'allow_soft_placement': True,
                     'device_count': {'CPU': 1},
                     'gpu_options': {'allow_growth': True},
                     'inter_op_parallelism_threads': 2,
                     'intra_op_parallelism_threads': 2,
                     'log_device_placement': False},
 'timesteps_per_iteration': 0,
 'train_batch_size': 4000,
 'use_critic': True,
 'use_gae': True,
 'vf_clip_param': 10.0,
 'vf_loss_coeff': 1.0,
 'vf_share_layers': -1}

For completeness, here’s a copy of the remote call to optimize:

    def optimize(trainer_constructor, trainer_config, registered_gym_name, level_string_monad, network_weights,
        """Run one step of optimization!!

        :param trainer_constructor: constructor for algo to optimize wtih e.g. ppo.PPOTrainer for rllib to run optimization.
        :param trainer_config: config dict for e.g. PPO.
        :param registered_gym_name: name of env registered with ray via `env_register`
        :param level_string_monad:  callback to allow for dynamically created strings
        :param network_weights: torch state_dict
        :return: dict of {optimized weights, result_dict}

        # todo same as rollout.py
        # todo will probably have to change this to first instantiate a generator model
        # and then query it for the levels.
        #  That will allow something like PAIRED to function?
        trainer_config['env_config']['level_string'], _ = level_string_monad()
        trainer = trainer_constructor(config=trainer_config, env=registered_gym_name)
        result = trainer.train()

        return {0: {'weights': trainer.get_policy('default_policy').model.state_dict(), # key of zero b/c this will be adapted to multi-agent in the future.
                    "result_dict": result,
                    'pair_id': kwargs.get('pair_id', 0)

The call looks like this:

list_of_agent_env_pairs = [...]
for p in list_of_agent_env_pairs:
    opt_refs = optimize.remote(ppoTrainer, config, gym_name, p.generator_fn_wrapper(), p.solver.get_state_dict())

For good measure, here’s a copy of the NN code:

class AIIDEActor(TorchModelV2, nn.Module):
    def __init__(self, obs_space: gym.spaces.Space,
                 action_space: gym.spaces.Space, num_outputs: int,
                 model_config: ModelConfigDict, name: str):
        TorchModelV2.__init__(self, obs_space, action_space, num_outputs,
                              model_config, name)

        self._num_objects = obs_space.shape[2]
        self._num_actions = num_outputs

        self.embedding = nn.Sequential(
            layer_init(nn.Conv2d(in_channels=self._num_objects, out_channels=8, kernel_size=1)),
            layer_init(nn.Conv2d(in_channels=8, out_channels=32, kernel_size=2)),
            layer_init(nn.Linear(512, 128)),  

        self.policy_head = nn.Sequential(
            layer_init(nn.Linear(128, num_outputs))

        self.value_head = nn.Sequential(
            layer_init(nn.Linear(128, 1))

    def forward(self, input_dict, state, seq_lens):
        # print(input_dict['obs'].shape)
        x = input_dict['obs'].permute(0, 3, 1, 2)
        self._last_batch_size = x.shape[0]

        embed = self.embedding(x)
        logits = self.policy_head(embed)
        value = self.value_head(embed)
        self._value = value.reshape(-1)
        return logits, state

    def value_function(self):
        return self._value
1 Like