mne-python: BUG: mne.parallel crashes when too much data

I’m encountering some weird error when I try to decod too much data e.g.

import numpy as np
from mne import create_info, EpochsArray
from mne.decoding import TimeDecoding

# simulate data
n_trial, n_chan, n_time = 800, 300, 1445
info = create_info(n_chan, 128., 'mag')
epochs = EpochsArray(np.random.rand(n_trial, n_chan, n_time),
                     info=info,
                     events=np.zeros((n_trial, 3), int))
y = np.random.randint(0, 2, n_trial)
epochs._data[y==1, :, :] += 1

# fit
td = TimeDecoding(n_jobs=-1, predict_mode = 'mean-prediction')
td.fit(epochs, y=y)

works fine but

td.predict(epochs)

returns

---------------------------------------------------------------------------
SystemError                               Traceback (most recent call last)
<ipython-input-22-274c297025ed> in <module>()
     13 td = TimeDecoding(n_jobs=-1)
     14 td.fit(epochs, y=y)
---> 15 y_pred = td.predict(epochs)

/home/ubuntu/mne-python/mne/decoding/time_gen.py in predict(self, epochs)
   1339         """  # noqa
   1340         self._prep_times()
-> 1341         super(TimeDecoding, self).predict(epochs)
   1342         self._clean_times()
   1343         return self.y_pred_

/home/ubuntu/mne-python/mne/decoding/time_gen.py in predict(self, epochs)
    291             n_orig_epochs=n_orig_epochs, test_epochs=test_epochs,
    292             **dict(zip(['X', 'train_times'], _chunk_data(X, chunk))))
--> 293             for chunk in chunks)
    294 
    295         # Concatenate chunks across test time dimension.

/home/ubuntu/anaconda/lib/python2.7/site-packages/joblib/parallel.pyc in __call__(self, iterable)
    808                 # consumption.
    809                 self._iterating = False
--> 810             self.retrieve()
    811             # Make sure that we get a last message telling us we are done
    812             elapsed_time = time.time() - self._start_time

/home/ubuntu/anaconda/lib/python2.7/site-packages/joblib/parallel.pyc in retrieve(self)
    725                 job = self._jobs.pop(0)
    726             try:
--> 727                 self._output.extend(job.get())
    728             except tuple(self.exceptions) as exception:
    729                 # Stop dispatching any new job in the async callback thread

/home/ubuntu/anaconda/lib/python2.7/multiprocessing/pool.pyc in get(self, timeout)
    565             return self._value
    566         else:
--> 567             raise self._value
    568 
    569     def _set(self, i, obj):

SystemError: error return without exception set

or

IOError: bad message length

There is no error if I set n_jobs=1, if I reduce the dimensionality of the data (either n_chan, n_time or n_trial), and if I predict a subset: e.g.

td.predict(epochs[:596]) 

works fine, but from 597, it crashes

td.predict(epochs[:597]) 

I am running on an aws m4.4xlarge running with ubuntu anaconda and mne dev, joblib is 0.9.4.

I have no idea what to do…