https://github.com/huggingface/transformers/blob/546dc24e0883e5e9f5eb06ec8060e3e6ccc5f6d7/src/transformers/models/gpt2/modeling_gpt2.py#L698

Assertions can't be relied upon for control flow because they can be disabled, as per the following:

$ python --help
usage: python [option] ... [-c cmd | -m mod | file | -] [arg] ...
...
-O     : remove assert and __debug__-dependent statem

It would be great if hyp2f1 could be added under jax.scipy.special! Are there any plans to do this?

My application uses a special case of the hyp2f1 that seems like it can be written using special functions already implemented in jax ([gammaln](https://jax.readthedocs.io/en/latest/_autosummar

Is there a way to train a bidirectional RNN (like LSTM or GRU) on trax nowadays?

In SVI latent random variables sampled from the guide completely mask those sampled in the model during inference. However, nothing prevents us from specifying different shapes for such a sample site in model and guide respectively. This makes it easy to introduce confusing bugs when code in the model expects a certain shape that is different than what guide provides. This can easily happen if the

Description

TIL that pip has a super nice Issue template page

that uses a [.github/ISSUE_TEMPLATE/config.yml](https://github.com/pypa/pip/blob/e6414d6db6db37951988f6f2b11ec530ed0b191d/.github/ISSUE_TEMPLATE/conf

See wikipedia table for example patterns: https://en.wikipedia.org/wiki/List_of_integrals_of_Gaussian_functions