EOFError error during remote_worker_envs flags

[XavierGeerinck]

What happened + What you expected to happen

I am trying to get training to work while setting remote_worker_envs to true, but I am getting an EOFError

Traceback (most recent call last):
  File "~/.venv/lib/python3.11/site-packages/ray/rllib/env/env_runner_group.py", line 169, in __init__
    self._setup(
  File "~/.venv/lib/python3.11/site-packages/ray/rllib/env/env_runner_group.py", line 239, in _setup
    self.add_workers(
  File "~/.venv/lib/python3.11/site-packages/ray/rllib/env/env_runner_group.py", line 799, in add_workers
    raise result.get()
  File "~/.venv/lib/python3.11/site-packages/ray/rllib/utils/actor_manager.py", line 500, in _fetch_result
    result = ray.get(ready)
             ^^^^^^^^^^^^^^
  File "~/.venv/lib/python3.11/site-packages/ray/_private/auto_init_hook.py", line 21, in auto_init_wrapper
    return fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^
  File "~/.venv/lib/python3.11/site-packages/ray/_private/client_mode_hook.py", line 103, in wrapper
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "~/.venv/lib/python3.11/site-packages/ray/_private/worker.py", line 2639, in get
    values, debugger_breakpoint = worker.get_objects(object_refs, timeout=timeout)
                                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "~/.venv/lib/python3.11/site-packages/ray/_private/worker.py", line 866, in get_objects
    raise value
ray.exceptions.ActorDiedError: The actor died because of an error raised in its creation task, ray::SingleAgentEnvRunner.__init__() (pid=83332, ip=127.0.0.1, actor_id=1ca45720433ca900e1057f5801000000, repr=<ray.rllib.env.single_agent_env_runner.SingleAgentEnvRunner object at 0x149cc1210>)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "~/.venv/lib/python3.11/site-packages/ray/rllib/env/single_agent_env_runner.py", line 79, in __init__
    self.make_env()
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "~/.venv/lib/python3.11/site-packages/ray/rllib/env/single_agent_env_runner.py", line 764, in make_env
    gym.vector.make(
  File "~/.venv/lib/python3.11/site-packages/gymnasium/vector/__init__.py", line 82, in make
    return AsyncVectorEnv(env_fns) if asynchronous else SyncVectorEnv(env_fns)
           ^^^^^^^^^^^^^^^^^^^^^^^
  File "~/.venv/lib/python3.11/site-packages/gymnasium/vector/async_vector_env.py", line 169, in __init__
    self._check_spaces()
  File "~/.venv/lib/python3.11/site-packages/gymnasium/vector/async_vector_env.py", line 504, in _check_spaces
    results, successes = zip(*[pipe.recv() for pipe in self.parent_pipes])
                              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "~/.venv/lib/python3.11/site-packages/gymnasium/vector/async_vector_env.py", line 504, in <listcomp>
    results, successes = zip(*[pipe.recv() for pipe in self.parent_pipes])
                               ^^^^^^^^^^^
  File "/Users/xaviergeerinck/.pyenv/versions/3.11.8/lib/python3.11/multiprocessing/connection.py", line 250, in recv
    buf = self._recv_bytes()
          ^^^^^^^^^^^^^^^^^^
  File "/Users/xaviergeerinck/.pyenv/versions/3.11.8/lib/python3.11/multiprocessing/connection.py", line 430, in _recv_bytes
    buf = self._recv(4)
          ^^^^^^^^^^^^^
  File "/Users/xaviergeerinck/.pyenv/versions/3.11.8/lib/python3.11/multiprocessing/connection.py", line 399, in _recv
    raise EOFError
EOFError

Versions / Dependencies

platform : macOS 14.2 23C64 (arm64) memory : 48.0 GB cpu : 16 cores mac : 6a:3c:67:54:c1:4d ip : 192.168.4.76 model_info : Mac15,9 (MUW73LL/A) kernel_version : 23.2.0 git_commit_sha : Unknown python_version : 3.11.8 (~/.venv/bin/python) pip_version : 24.0 (~/.venv/lib/python3.11/site-packages/pip) torch_version : 2.3.0 docker_version : 24.0.7, kubernetes_version : 1.28.2 ray_version : 2.31.0 nvidia_smi : Unknown, nvidia-smi was not found nvidia_cuda : Unknown, nvcc was not found is_tty : True

Reproduction script

.env_runners(
    num_env_runners=1,
    num_envs_per_env_runner=8,
    num_cpus_per_env_runner=1,
    num_gpus_per_env_runner=0,
    sample_timeout_s=60,
    remote_worker_envs=True,
    rollout_fragment_length="auto",
)

Issue Severity

High: It blocks me from completing my task.

[simonsays1980]

@XavierGeerinck Thanks for raising this issue. Can you provide a reproducable example? I guess this might be in the new API stack which does not support asynchronous vector environments (yet - we wait for a gymnasium update).

[XavierGeerinck]

Awesome! We indeed are thinking the same and are awaiting the 1.0.0a2 release to start testing . Is there any ETA currently that you are aware of?

[simonsays1980]

This should come soon, but we know of no ETA in regard to it. Would it help, for the time being to just sample with more Env Runners but a single env in each of them?