Skip to content

TFRT

Jump to bottom
Ashok Bhat edited this page May 25, 2020 · 23 revisions

Description

  • New runtime that will replace the existing TensorFlow runtime.
  • Responsible for efficient execution of kernels – low-level device-specific primitives – on targeted hardware.

TFRT’s role in graph and eager execution within the TensorFlow training stack

Why?

  • Efficient use of multithreaded host CPUs
  • Supports fully asynchronous programming models
  • Low-level efficiency
  • Help Hardware makers integrate edge and datacenter devices into TensorFlow in a modular way.

TFTR and MLIR

  • TFRT utilizes MLIR’s compiler infrastructure to generate an optimized, target-specific representation of your computational graph that the runtime executes.
  • TFRT uses MLIR’s extensible type system to support arbitrary C++ types in the runtime, which removes tensor-specific limitations.

TFRT CPU kernels

  • Where is it?
  • What does it use?
  • Does it use Eigen or OneDNN?

FAQ

  • What is TFRT?
  • When is TFRT going to be the default runtime?
  • What are the main problems being solved by TFRT?
  • What are the main use-cases for TFRT?
  • What does TFRT rely on for CPU kernels?
  • What is the relation of TFRT and MLIR?
  • What is the relation between TFRT and TensorRT?
  • What is the relation between TFRT and OneDNN?
  • What is the relation between TFRT and TensorFlow Lite?

See also

Clone this wiki locally