What is the roadmap of TFRT?

[haozech]

Hi, I'm interested in this project(TFRT) and I notice that this repo has been no big updates for a while. So I wonder what's the current developing pace of TFRT? Is some of the releases closed-source? And will the underlying runtime framework of tensorflow be replaced with TFRT according to the roadmap?

[mhong]

Thank you for your question and interest. While we don't have a publicly shareable roadmap at this point, we've been making continued progress in building out TFRT and integrating it with TF training and serving. Here are some recent updates: - TFRT-based TF serving <https://github.com/tensorflow/serving> has been increasingly covering more CPU models at Google. The low level infra code for runtime is here <https://github.com/tensorflow/runtime>, and compiler passes here <https://github.com/tensorflow/tensorflow/tree/master/tensorflow/compiler/mlir/tfrt>. Additional code for TPU support is not open sourced. - For GPU support, TFRT is being integrated with XLA GPU. Here's a recent CL <tensorflow/tensorflow@3cea30a> that open sourced the integration with TF training, and this is WIP. - TFRT-based CPU and TPU runtimes have also been integrated with Jax (example code <https://github.com/google/jax/blob/28d9e9236285b4dd6be6145ba866067ea0e192be/jax/lib/xla_bridge.py#L70-L74>), and the corresponding GPU runtime (through XLA GPU) is WIP.

…

On Wed, Jul 28, 2021 at 4:42 AM haozech ***@***.***> wrote: Hi, I'm interested in this project(TFRT) and I notice that this repo has been no big updates for a while. So I wonder what's the current developing pace of TFRT? Is some of the releases closed-source? And will the underlying runtime framework of tensorflow be replaced with TFRT according to the roadmap? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#71>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAHMYR5ALBEHHTHXEKNKTOLTZ7UKVANCNFSM5BEF3PXQ> .

[haozech]

Thank you for your question and interest. While we don't have a publicly shareable roadmap at this point, we've been making continued progress in building out TFRT and integrating it with TF training and serving. Here are some recent updates: - TFRT-based TF serving https://github.com/tensorflow/serving has been increasingly covering more CPU models at Google. The low level infra code for runtime is here https://github.com/tensorflow/runtime, and compiler passes here https://github.com/tensorflow/tensorflow/tree/master/tensorflow/compiler/mlir/tfrt. Additional code for TPU support is not open sourced. - For GPU support, TFRT is being integrated with XLA GPU. Here's a recent CL <tensorflow/tensorflow@3cea30a> that open sourced the integration with TF training, and this is WIP. - TFRT-based CPU and TPU runtimes have also been integrated with Jax (example code https://github.com/google/jax/blob/28d9e9236285b4dd6be6145ba866067ea0e192be/jax/lib/xla_bridge.py#L70-L74), and the corresponding GPU runtime (through XLA GPU) is WIP.
…
On Wed, Jul 28, 2021 at 4:42 AM haozech @.***> wrote: Hi, I'm interested in this project(TFRT) and I notice that this repo has been no big updates for a while. So I wonder what's the current developing pace of TFRT? Is some of the releases closed-source? And will the underlying runtime framework of tensorflow be replaced with TFRT according to the roadmap? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#71>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAHMYR5ALBEHHTHXEKNKTOLTZ7UKVANCNFSM5BEF3PXQ .

Thank you for your reply!