[TFBart] Split TF-Bart#9497
Conversation
patrickvonplaten
force-pushed
the
split_tf_bart
branch
from
a82cf67 to
1b2bc4f
Compare
There was a problem hiding this comment.
Awesome work!! Just left few smalll comments. I think we should first find a proper fix #9478 and then merging this one. Switching on/off some tests everytime we touch a model is really not a long term solution, I think a proper template as to be stated first and then afterwards we do the models.
IMO this PR should be merged and the s2s fix should be applied afterward as said offline. This PR is blocking a new release currently |
Ok, nevermind, I didn't know you wanted to have it in the next release. |
LysandreJik
left a comment
LysandreJik
left a comment
There was a problem hiding this comment.
LGTM! Left a few nits, but great work all around! Thank you for taking care of this.
| class TFBartModelIntegrationTest(unittest.TestCase): | ||
| def test_inference_no_head(self): | ||
| model = TFBartModel.from_pretrained("facebook/bart-large", from_pt=True) | ||
| model = TFBartForConditionalGeneration.from_pretrained("facebook/bart-large").model |
There was a problem hiding this comment.
doesn't work, the weights don't load correctly...this didn't work previously either. Could be dealt with in a future PR
sgugger
left a comment
sgugger
left a comment
There was a problem hiding this comment.
Thanks a lot for doing this! There are just a few repeated mistakes, that probably come from the templates.
4 participants
What does this PR do?
TF mirror of: #9343
After PR is merged TODO:
facebook/blenderbot_small-90Mtokenizer - cannot download files from hub. Weird issue