TF: GPT-J compatible with XLA generation #17986
Merged
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
This PR modifies TF GPT-J so as to be compatible with XLA generation. It borrows the new code from FLAX -- in essence, instead of computing the embedded positions (
sincos
) at each call, given the size of the sequence (which could be obtained from the size of the past), now pre-computes the embedded positions and gathers them given theposition_ids
.@tooslow
, due to the size of the model. I've reworked the tests BEFORE touching GPT-J code, to test all needed features correctly. All but the XLA test were passing before GPT-J was changed, and all tests pass after the changes. We still have two XLA tests being run in CI frequently (test_xla_generate_fast
andtest_xla_generate_slow
), as well as a couple of generic generate tests -- they just don't use the trained model weights.