Unexpected PJRT_Plugin_Attributes_Args size: expected 32, got 24.

[forcadellvincent]

Hello @alvarosg,

I've come across an error trying to run the graphcast_demo.ipynb notebook on a TPU v5litepod-1.

The following error is triggered when the rollout is performed, likely when the model is loaded or when the forward operator is called:

pjrt_c_api_helpers.cc:258: Unexpected PJRT_Plugin_Attributes_Args size: expected 32, got 24. The plugin is likely built with a later version than the framework. This plugin is built with PJRT API version 0.75

This error ultimately leads to an Aborted (core dumped), stopping the rollout.

I feel that it has to do with my TPU not being compatible with the version of jax or libtpu libraries I have installed in my python venv.

Did you ever encounter this problem?

Thank you very much

[forcadellvincent]

Hi,

Found the bug.

It apparently comes from two different sources:

• the requirements mismatch in between jax and the underlying libtpu and jaxlib:
  • the solution is to change from jax to jax[tpu], to make sure the tpu library libtpu is installed with correct version aligned with jax. It should be mentionned in the README.md for TPU users.
• the introduction of breaking changes in jax-v0.7.0 (https://github.com/jax-ml/jax/releases/tag/jax-v0.7.0).

  • jax.stages.OutInfo has been replaced with jax.ShapeDtypeStruct.

  • This change throws an error when the autoregressive module is imported. The most recent version of jax to correctly run the graphcast_demo.ipynb seems to be jax[tpu]==0.6.2. The jax version in the requirements should be lowered to this version, waiting for a solution to make it work on jax >= v0.7.0.

[andrewlkd]

Hey,

Apologies for the delayed response. I'm glad you found a solution.

Could you confirm whether applying this fix https://github.com/google-deepmind/graphcast/commit/7077d40a36db6541e3ed72ccaed1c0d202fa6014 in graphcast_demo.ipynb also solved the issue? In which case we will also make the change there.

Best,

Andrew