difference between TF_XLA_FLAGS=--tf_xla_auto_jit=2 and tf.config.optimizer.set_jit(True)

[t kevin]

Hi Folks,

What's the difference between these two methods to enable XLA?

Are they identical? Please let me know which piece of code should I look into if not.

https://www.tensorflow.org/xla#explicit_compilation_with_tffunctionjit_compiletrue

This page said nothing about tf.config

https://www.tensorflow.org/api_docs/python/tf/config/optimizer/set_jit

and this page has a weird deprecation warning.

Thanks

Kevin

[George Karpenkov]

Hi Kevin,

They are identical and they both enable autoclustering. TF_XLA_FLAGS=--tf_xla_auto_jit=2 is preferred, as optimizer.set_jit has some weird caching behavior, and non-obvious non-local effects.

In general, jit_compile=True is preferred if possible.

--
You received this message because you are subscribed to the Google Groups "XLA development" group.
To unsubscribe from this group and stop receiving emails from it, send an email to xla-dev+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/xla-dev/cd9d2760-962a-473c-a7da-ed752adeb2f9n%40googlegroups.com.