search

IBM / tensorflow-large-model-support

Large Model Support in Tensorflow
Apache License 2.0
202 stars 38 forks source link

Tensorflow 2 eager mode support #45

Closed anji993 closed 4 years ago

anji993 commented 4 years ago

Hello, can LMS works in the eager model in tensorflow 2?

And, is there a plan to release patches for higher tensorflow versions than 2.1?

jayfurmanek commented 4 years ago

Hi, Yes. In TensorFlow 2 we moved the LMS function from the Python layer (where it relied on the pre-built graph existing) to the core memory manager. The Tensorflow 2.1 package from IBM's WML CE (https://public.dhe.ibm.com/ibmdl/export/pub/software/server/ibm-ai/conda/) includes LMS built in and this will work with eager mode (which is enabled by default in TF2).

The answer to you second question is unfortunately no for now.

jayfurmanek commented 4 years ago

Oh, I failed to mention that there is a 2.2 build enabled with LMS in the "Early Access" WML CE channel:

https://public.dhe.ibm.com/ibmdl/export/pub/software/server/ibm-ai/conda-early-access/

That was the last update we did. I'll investigate about pushing the patch set out for that build.

smatzek commented 4 years ago

In addition, we just merged a PR in this repository which adds the TensorFlow 2.2.0 patch to the patches directory, updates the examples, readme, and allocator stats documentation.

No work has been done for LMS on versions later than 2.2.0 and as @jayfurmanek said there are unfortunately no current plans to do so.