Improve TF weight loading, especially PT crossloading #21792
Rocketknight1
merged
24 commits into
main
from
tf_weight_loading_fixes
First commit for the improved PT-TF weight loading
f9cee34f
Remove workarounds from TFEncoderDecoder tests
9d18cea5
Allow a custom weight renaming function in from_pretrained and use th…
cfde5786
make fixup
cef740d5
First attempt at visionencoderdecoder
84bbacb6
Disable tensorfloat32 in tests to get consistent outputs
bc020f8d
Quick fix to tf_vision_encoder_decoder tests
8bdcd36d
make fixup
1c5d7cf9
Update Blenderbot tests
44db428c
Remove unused arg in modeling_tf_opt
edea418e
load_tf_sharded_weights had strict=True! This meant transfer learning…
c7e3d797
Support prefixes when loading sharded TF checkpoints
6b352e1e
make fixup
f66de506
Rocketknight1
commented
on 2023-02-28
Add test to load sharded models with a weight prefix
d7c17700
Fix sharded weight loading test
309958c8
Add a test for transfer from a sharded checkpoint
0f699914
make fixup
2fd46b41
Add test to check that crossloading from PT with a prefix works
00da6cea
Refactor from_pretrained in the encoderdecoder classes
27631203
Refactor from_pretrained in the encoderdecoder classes
9c26cfdf
gante
approved these changes
on 2023-02-28
missmatched -> mismatched
faf91fc4
Explicitly check for None
58cab95e
No comments showing my very impressive and attractive knowledge of Py…
dda10e47
Disable TF32 across all TF tests
893d669e
Login to write a write a comment.
Login via GitHub