Array has been deleted

[David Ding]

Hi, we are trying out the orbax (0.4.1) AsyncCheckpointer (used through CheckpointManager). We are getting "Array has been deleted" errors. It seems as if the async checkpointer is trying to copy a jax.Array from device to memory, but that array is no longer available. The Orbax documentations says that "From start to finish, async checkpointing for a train state of arrays works by first performing a blocking copy of the arrays from device to host", but I wonder if there any gotchas in how we should use orbax checkpointing.

Here is the stack trace:

Exception in thread Thread-314 (_finalize):
Traceback (most recent call last):
File "/usr/lib/python3.11/threading.py", line 1045, in _bootstrap_inner
     self.run()
File "/usr/lib/python3.11/threading.py", line 982, in run
     self._target(*self._args, **self._kwargs)
File ".venv/lib/python3.11/site-packages/orbax/checkpoint/checkpoint_manager.py", line 956, in _finalize
    self.wait_until_finished(join_finalize_thread=Fale)
File ".venv/lib/python3.11/site-packages/orbax/checkpoint/checkpoint_manager.py", line 888, in wait_until_finished
     checkpointer.wait_until_finished()  # pytype: disable=attribute-error
     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File ".venv/lib/python3.11/site-packages/orbax/checkpoint/async_checkpointer.py", line 262, in wait_until_finished
     self._async_manager.wait_until_finished()
File ".venv/lib/python3.11/site-packages/orbax/checkpoint/async_checkpointer.py", line 154, in wait_until_finished
     self.check_for_errors()
File ".venv/lib/python3.11/site-packages/orbax/checkpoint/async_checkpointer.py", line 145, in check_for_errors
     raise exception  # pylint: disable=raising-bad-type
     ^^^^^^^^^^^^^^^
File ".venv/lib/python3.11/site-packages/orbax/checkpoint/async_checkpointer.py", line 97, in _thread_func
     future.result()
File "/usr/lib/python3.11/concurrent/futures/_base.py", line 456, in result
     return self.__get_result()
            ^^^^^^^^^^^^^^^^^^^
File "/usr/lib/python3.11/concurrent/futures/_base.py", line 401, in __get_result
     raise self._exception
File "/usr/lib/python3.11/concurrent/futures/thread.py", line 58, in run
    result = self.fn(*self.args, **self.kwargs)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File ".venv/lib/python3.11/site-packages/orbax/checkpoint/aggregate_handlers.py", line 75, in _serialize_fn
     msgpack = msgpack_utils.msgpack_serialize(serializable_dict)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File ".venv/lib/python3.11/site-packages/orbax/checkpoint/msgpack_utils.py", line 216, in msgpack_serialize
     return msgpack.packb(pytree, default=_msgpack_ext_pack, strict_types=True)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File ".venv/lib/python3.11/site-packages/msgpack/__init__.py", line 36, in packb
     return Packer(**kwargs).pack(o)
File "msgpack/_packer.pyx", line 285, in msgpack._cmsgpack.Packer._pack
File ".venv/lib/python3.11/site-packages/orbax/checkpoint/msgpack_utils.py", line 78, in _msgpack_ext_pack
     return msgpack.ExtType(_MsgpackExtType.NDARRAY, _ndarray_to_bytes(x))
                                                     ^^^^^^^^^^^^^^^^^^^^
File ".venv/lib/python3.11/site-packages/orbax/checkpoint/msgpack_utils.py", line 40, in _ndarray_to_bytes
     arr = np.array(arr)
           ^^^^^^^^^^^^^
File ".venv/lib/python3.11/site-packages/jax/_src/array.py", line 377, in __array__
     return np.asarray(self._value, dtype=dtype)
                       ^^^^^^^^^^^
File ".venv/lib/python3.11/site-packages/jax/_src/profiler.py", line 340, in wrapper
     return func(*args, **kwargs)
            ^^^^^^^^^^^^^^^^^^^^^
File ".venv/lib/python3.11/site-packages/jax/_src/array.py", line 562, in _value
     self._check_if_deleted()
File ".venv/lib/python3.11/site-packages/jax/_src/array.py", line 530, in _check_if_deleted
     raise RuntimeError(
RuntimeError: Array has been deleted with shape=float32[256].

[Niket Kumar]

Hi David, Can you please try 0.4.7, which might help.

[David Ding]

Hi Niket,

I tried upgrading to 0.4.7, and I'm still hitting the error. The only difference is that the error now crashes the training loop, whereas previously training would continue. If I do wait=True, the error disappears, but obviously that's not performant.

Do I need to pass in np.array(jax.experimental.multihost_utils.global_array_to_host_local_array(params)) explicitly? And do I need to be careful about inserting jax.block_until_ready calls?

[David Ding]

I did a bit more debugging. I think the error occurs when aggregate=True in the save_args. Reading the code, it seems like in the aggregate=False branch, the type_handlers.ArrayHandler.async_save correctly awaits the copy futures, but when aggregate=True, the threads that are writing the checkpoints access to DeviceArrays directly (https://github.com/google/orbax/blob/main/checkpoint/orbax/checkpoint/aggregate_handlers.py#L72)?

Note that aggregate=True is automatically set by flax.orbax_utils.save_args_from_target(ckpt) for non-sharded params (e.g. all params in a smaller model that is not model sharded or zero sharded). Do you recommend setting aggregate=True? I guess even with a non-sharded model, parameters can be big, so perhaps checking if the array is sharded or not isn't the most accurate heuristic?

[Niket Kumar]

Thanks for debugging the issue and associating it with SaveArgs.aggregate option!

While we recreate the issue in our dev setup, please switch to aggregate=False if that works for your use case. SaveArgs.aggregate is mainly meant for performance optimization. I hope that unblocks you.

To recreate the issue, I will need your help:

Do I need to pass in np.array(jax.experimental.multihost_utils.global_array_to_host_local_array(params)) explicitly?
Which api do you want to pass this to? Input to save(...)?

And do I need to be careful about inserting jax.block_until_ready calls?
Which code snippet were you referring to?

Note that aggregate=True is automatically set by flax.orbax_utils.save_args_from_target(ckpt) for non-sharded params (e.g. all params in a smaller model that is not model sharded or zero sharded).
Please note that aggregate=True is set if 1) array is not a jax.Array or 2) it is jax.Array and is_fully_replicated. Please note that it is about replication.

Can you please share a code snippet or pointer which details the Orbax usage, so that I can recreate the issue and resolve it?