JB (Don)

Still working through another 10 failing cases (where it's not so obvious), so only marking as draft for now. tests/models/flava/test_modeling_flava.py .....F tests/models/encodec/test_modeling_encodec.py F tests/models/fsmt/test_modeling_fsmt.py F tests/models/lxmert/test_modeling_lxmert.py F tests/models/marian/test_modeling_marian.py F. tests/models/musicgen/test_modeling_musicgen.py...

Cool, I fixed the remaining missing _tie_weights(), and also added some more skip tests for some special audio/vision models (many failing due to use of `nn.utils.weight_norm`). I ran the following...

The test failure in tests_tf looks unrelated. Any chance you can kick off a re-run of the CI checks? 🙏 Also, I've verified again that `pytest -k test_save_load_low_cpu_mem_usage tests/` passes.

Thanks for the review! I added the explanation of tie_weights() from my research, but it'd be great to get some feedback from someone who's more knowledgeable on this.

cc @SunMarc @muellerzr Don't meant to be pushy, but the tests for the models in this change are currently broken in main/HEAD, so I'd be grateful if you could give...

> cc @SunMarc @muellerzr > > Don't meant to be pushy, but the tests for the models in this change are currently broken in main/HEAD, so I'd be grateful if...

Sorry I added the wrong link in the PR description, this issue is a follow up of #28948. There's context in that link (tl;dr adding the tie_weights() enable those models...

The latest commit should be based off of main/HEAD and has all the necessary changes to tie_weights(). It includes additional tests for safe tensors and also checkpoint bins, and also...

@SunMarc @younesbelkada Any other insights or concerns over the use of tie_weights() here? Thanks!

@SunMarc @younesbelkada I just sync'd latest HEAD. Would you mind taking a quick look? Thanks!