Make tensorflow profiler into OpenXLA

[yliu120]

Hi,

I assume that many users in the community has the same issue that we need to pip install tensorflow even though we don't use tensorflow. The reason for that is that the tensorboard depends on tensorflow.

I guess only some very small part of the profiler frontend has some dependencies on the util libraries inside tensorflow. Can we make tensorboard depending purely on XLA so that all JAX users don't have to install tensorflow to use the profiler tool?

The reasons why we don't like the tensorflow package:

1. Very very big, ~1.5GiB in size and we don't use it.
2. depends on very old protobuf library which gives us many dependency conflicts.

I see two directions:

1. in a short term, I guess we can potentially update the protobuf library of XLA to some latest version.
2. in the long run, we should completely lift tensorboard from TF code base.

[cliveverghese]

Hi,

With version 2.19.0 of the profiler, You don't have to install tensorflow.

Just the following should be sufficient.

pip install tensorboard tensorboard-plugin-profile

[yliu120]

I have an additional question:

https://github.com/tensorflow/tensorboard/blob/master/WORKSPACE#L147-L155

If we don't constraint on being compatible with TF, we don't need to pin protobuf version to 3.x?

[Matt-Hurd]

While we don't require tensorflow, there are still codepaths that do use it. Example. If all of these are removed, then we can do a pass over our dependencies.