GitLab

Also, do Transformer inference in fp16, as well as training, when --dtype=fp16. In TF 2, layers now cannot run in multiple different dtypes, so we must use the same dtype for training and inference.

261171038  by gjn<gjn@google.com>:

    Remove weight_decay_rate 0 early exit check

    Removing this code path should be fine since this was actually not doing
    what it meant to do. Since weight_decay_rate is actually a tensor, the
    equality check was only looking at the id of the object and comparing to
    0. This should never be true. Evaluating a tensor is also not what we
    want to do at this point of the code. Thus it should be fine to simply
    remove this code.

--
261169862  by haoyuzhang<haoyuzhang@google.com>:

    Internal change

261153520  by haoyuzhang<haoyuzhang@google.com>:

    Internal change

261140302  by hongkuny<hongkuny@google.com>:

    Clean up

--

PiperOrigin-RevId: 261171038

* Merged commit includes the following changes:
249776315  by tianlin<tianlin@google.com>:

    Internal change

249763206  by tianlin<tianlin@google.com>:

    For TF 2.0 (related to Beam Search), expand cond dims in tf.where(cond, x, y) to make all parameters broadcastable.

--
249392724  by hongkuny<hongkuny@google.com>:

    Internal change

PiperOrigin-RevId: 249776315

* Merged commit includes the following changes:
249823043  by tianlin<tianlin@google.com>:

    Bring back v2 test for predict and eval.

--

PiperOrigin-RevId: 249823043