The current lazy loader relies on a feature of the Python zipfile module
that was added in Python 3.7.0:
https://bugs.python.org/issue22908
This commit adds compatibility for Python 3.6.
If `probs` is zero then `log_probs` will be negative infinity, and the
calculation of `neg_entropy` would then give NaN because zero times
infinity is a mathematically indeterminate value.
We need to use nansum so that those NaN values are treated as zeros to
ignore them in the entropy calculation.
This warning happens when `np.log` gets an input containing zeros.
In that case, NumPy will throw a warning and output negative infinity.
Negative infinity is the correct behaviour here, so we can safely ignore
the warning.
* Error messages are now shown when memory, author's note, etc. exceeds
budget by itself
* Formatting options no longer break if there are empty chunks in the
story (although there shouldn't be any in the first place)
* Number of generated tokens is now kept track of from Python
A Python `for` loop was replaced with a `jax.lax.scan` call so that JAX
only compiles the `transformer.generate_initial` function one time
instead of `numseqs` times. This is because JAX unrolls Python built-in
loops like `for`. The compilation times should now be about the same as
they were before the upgrade to JAX 0.2.21.
The original one supported versions of JAX up to 0.2.12, and possibly also some
earlier versions. This new code supports exclusively JAX 0.2.21 and does not
work with any earlier or later versions of JAX. However, this new code benefits
from not needing to recompile when changing "Amount To Generate" and also from
supporting stopping generation early, which makes an implementation of Dynamic
World Info Scan finally possible.
Gnome Ann
henk717