who who who

I found the default setting is faster than ```accum_steps=1``` on a single device.

You could test the code on my PR which I have fixed several bugs.

Performance issue from other ops, you could test it via follow example: ```python # -*- coding: utf-8 -*- """edt model speed.ipynb Automatically generated by Colaboratory. Original file is located at...

I made a mistake, I should trace it by nvprof and the kernel is really slow.

@seanpmorgan @bhack I could add rocm support for addons and I have passed custom layer ops tests in my local environment. Could tensorflow team provides a CI/CD environment?

MirroredStrategy supports single GPU too.

In order to use tflite model, you have to convert strings to token ids, such as 'He'-> 1332. @yoeo

For improving model performance, I recommend ```tf.keras.layers.TextVectorization``` + FastText model which is similar to the current model. For more details, taking a look at https://www.tensorflow.org/text/guide/word_embeddings

I'm interested in it.

Just use open source repo or integrate them into a repo?