FlaxGPTJ #14396
FlaxGPTJ #14396
Conversation
patil-suraj
force-pushed
the
flax-gptj
branch
from
66b2707 to
740110d
Compare
| # overwrite from common since `attention_mask` in combination | ||
| # with `causal_mask` behaves slighly differently | ||
| @is_pt_flax_cross_test | ||
| def test_equivalence_pt_to_flax(self): |
There was a problem hiding this comment.
Thanks for adding this test - could you also add it in test_modeling_gptj.py so that the test fetcher always correctly checks PT<->Flax (e.g. at the moment if PT GPTJ is changed the test fetcher will not run this test, thus we need an identical test in the PyTorch test suite)
There was a problem hiding this comment.
These are already added in test_modeling_common.py, so they run for all pt models now
transformers/tests/test_modeling_common.py
Line 1630 in faacd74
There was a problem hiding this comment.
Ok! So for the Flax files the PT<=>Flax tests need to be overwritten whereas for the PT files the test doesn't have to be overwritten? If this is the case - the PR is good to be merged for me :-)
|
Looks great - just one small test for PT<>Flax compatibility should be added in |
What does this PR do?
This PR adds the GPTJ model in flax.