FastText RAM usage

[menshikh-iv]

Problem description

The FastText model takes too much RAM. We saw this issue many times in our CI systems, this typically looks like

____________________ TestFastTextModel.test_cbow_hs_online _____________________
self = <gensim.test.test_fasttext.TestFastTextModel testMethod=test_cbow_hs_online>
    @unittest.skipIf(IS_WIN32, "avoid memory error with Appveyor x32")
    def test_cbow_hs_online(self):
        model = FT_gensim(
>           sg=0, cbow_mean=1, alpha=0.05, window=2, hs=1, negative=0, min_count=3, iter=1, seed=42, workers=1
        )
self       = <gensim.test.test_fasttext.TestFastTextModel testMethod=test_cbow_hs_online>
/venv/lib/python3.7/site-packages/gensim/test/test_fasttext.py:733: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
/venv/lib/python3.7/site-packages/gensim/models/fasttext.py:595: in __init__
    self.trainables.prepare_weights(hs, negative, self.wv, update=False, vocabulary=self.vocabulary)
...
/venv/lib/python3.7/site-packages/gensim/models/fasttext.py:1130: in prepare_weights
    self.init_ngrams_weights(wv, update=update, vocabulary=vocabulary)
 ...
mtrand.pyx:1307: in mtrand.RandomState.uniform
    ???
        ...
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
>   ???
E   MemoryError

Steps/code/corpus to reproduce

from gensim.models import FastText

m = FastText()  # and measure RAM after that

It will "eat" around 1.6GB after __init__. I guess than issue in "bucket" matrix (too large).
Itself, this isn't a problem, but this is an issue in our tests (because we almost never pin this parameter).

Versions

Please provide the output of:

>>> import platform; print(platform.platform())
Linux-5.3.0-46-generic-x86_64-with-Ubuntu-19.10-eoan
>>> import sys; print("Python", sys.version)
('Python', '2.7.17 (default, Nov  7 2019, 10:07:09) \n[GCC 9.2.1 20191008]')
>>> import numpy; print("NumPy", numpy.__version__)
('NumPy', '1.16.1')
>>> import scipy; print("SciPy", scipy.__version__)
('SciPy', '1.2.3')
>>> import gensim; print("gensim", gensim.__version__)
('gensim', '3.8.1')
>>> from gensim.models import word2vec;print("FAST_VERSION", word2vec.FAST_VERSION)
('FAST_VERSION', 1)

I'm sure than other py versions and more early gensim versions affected in the same way