irhum / esmjax Goto Github PK

Somehow I am getting worse performance with 15B parameter model than a 650M parameter model. This might explain it.

Have you ran tests comparing the embedding produced by the smaller models in jax with the same model in PyTorch?

Right now there are two places the named axes are translated into the mesh axes (X and Y):

• params: Inside the library at https://github.com/irhum/esmjax/blob/main/esmjax/modules/partitioning.py#L45
• activations: By the instantiated mesh in the final script with the nn_partitioning.axis_rules(partitioning.DEFAULT_TPU_RULES) context manager, see examples/inference_15B.ipynb for example.

This works fine when default translation rules are used, but may lead to bugs when custom rules are used (if the rules in both places are not separately specified as non-default)

Proposed fix: Rely on Flax to perform the translation for both params and activations inside the context manager.

The training time for the model is very slow - 400 hours per epoch for 3B parameter as compared to 1.5 hours per epoch for my Keras data-parallel implementation of 650M model

The original splits over UniRef50 can be found in the original repo. Using a subset of them, need to compute either:

• the randomized masking perplexity
• the pseudo-perplexity with exhaustive single-token masking

Will likely do the former for computational efficiency.