Generate: validate model_kwargs on TF (and catch typos in generate arguments)
#18651
Conversation
|
The documentation is not available anymore as the PR was closed or merged. |
There was a problem hiding this comment.
moved from test_modeling_tf_common, no changes
There was a problem hiding this comment.
moved from test_modeling_tf_common, added the @slow (takes >30s)
There was a problem hiding this comment.
Cool that makes sense (similar to PyTorch).
Also just FYI in PyTorch we're testing currently much more than in TF mainly because we've allowed to return hidden_states and attentios. We could do the same for TF at some point
There was a problem hiding this comment.
Same as for PyTorch (here), with self.forward replaced with self.call
There was a problem hiding this comment.
The model is not relevant for the test, but not using a model from hf-internal-testing was an oversight in the previous PR :D
There was a problem hiding this comment.
Yes def a good idea!
gante
changed the title
model_kwargs on TF (and catch typos in generate arguments)
There was a problem hiding this comment.
Haha, fine with me! Can also increase all numbers otherwise
patrickvonplaten
left a comment
patrickvonplaten
left a comment
There was a problem hiding this comment.
Thanks for the clean-up
gante
force-pushed
the
tf_generate_kwarg_valid
branch
from
293120d to
3eb5072
Compare
3 participants
What does this PR do?
TF version of #18261
Adds
model_kwargsvalidation to TFgenerate, which also catches typos in the arguments. See the PR above for more details and an example of the error message users will see.Since TF had no dedicated file for
generatetests, I took the liberty to create it and move some existing tests there (>70% of the diff is due to moving things around :) ). The test for this new check was also added there.