transformers
Refactor the way we handle outputs for new llamas and new models
#39120
Merged
Merged
Refactor the way we handle outputs for new llamas and new models #39120
ydshieh
merged
124 commits into
main
from
clean-llamas
just update 2 files
7433c443
update other models as well just making fix-copies
37b4ef02
also add the changes needed to modeling utils
7f113b43
put this on the pretrained model instead
abf9d39d
nits and fixes
eb6747bc
update generic, fix to use config value
0f1d7e0a
update other modelings
e437edd7
use transformers kwargs instead
96aabd77
update
63df15bb
update
98f402cd
update other models
a7e0ce23
update
c9bb39ef
updates
cb5da530
update
0dc08262
update
fca73ad7
update
98739ba4
fix
124cd829
finally
4a14287a
very small nits
ea87eb70
this fixes more tests
8c66f4d0
fix other models as well!
3caf7d76
LysandreJik
approved these changes
on 2025-06-30
update modularqwen2
113219be
update models based on qwen2
e7705c98
update
a74974d9
update
3fb6b710
remove the **flash stuff in favor of noraml kwargs
7266aafa
update
c7d195fe
propagate gemma?
e63ef640
remove output attentions
1303470a
propagate
063e510d
ArthurZucker
commented
on 2025-06-30
Merge branch 'main' of github.com:huggingface/transformers into clean…
8c96926f
support cross attention edge case
01d4da85
same
780141ca
test this
3c0c56b8
fixes
7a0512a1
more fix
a13a98c6
update
15a8ff4f
update
22423738
update
2748b993
fix conflicts
da50ccc5
update
209d5022
fix emu3
10fb88ae
fix emu3
00afce98
move the fix a bit
3ac6c52f
quel enfer
0b119ffb
some fixes, loss_kwargs should never had been
f7a1f0da
finish fixing gemma3n
6a132a07
fix small lm3
9fa5f266
fix another one
aaae861f
fix csm now
5e5ae84a
fux csm and mistral
075bd0c2
fix mistral now
d04c2b1a
small fixes
5065b9a2
fix janusss
6a5f410d
only for some models
4834aeca
fixup
d8ee27e4
phix phi3
e2973440
more fixes?
0c9f6de0
dose this fix it?
501aead2
update
253307a3
holy shit it was just graph breaks
a267d8d4
protect torch
17cf5424
updates
c4d43c53
fix samhq?
4fc83fa3
fix moonshine
499ae87e
more moonshine fixes, 3 failures left!
b3c8641f
nits
b81df9bd
generic needs to support more
cfe62b6b
more fixes to moonshine!
6eb5e53e
fix cross attention outputs!
a9690f43
fix csm!
d462a8ea
nits
0f3c3683
fix stupid kosmos2
3cba8ac3
current updates
5af5bccd
fixes
9968c85e
use output recorder?
fbfaf040
nicer!
1f559c67
a little bit of magic
cd63172c
update
cf2e98c9
fix protect
c278e1cb
fix
e3c82cb7
small fixes
c5592be0
protect import
f6190cbf
fix a bunch of more models
d0be3319
fix fixups
22f0eaea
fix some of the last ones
422122d6
nit
feba9a03
partly fix phi
9a3708ae
update
7a0f14a7
fix import path
c4f314b3
Merge branch 'main' of github.com:huggingface/transformers into clean…
c6c5efbe
make something that is fullgraph compatible just to be sure
5f3722cf
typing was wrong on llama so the rest was wrong as well
7781368d
fucking ugly but at least it is still exportable
c9493081
syle
eaa7392b
supposed to fix moonshine, it still breaks
4b6a535c
fix some default
9976ed8b
fix the last bits of sam
6d723988
update samhq
ddea6837
Merge branch 'main' of github.com:huggingface/transformers into clean…
f021967a
more fixes to am hq
2e296b55
nit
8aaa10e2
fix all output+hidden states and output_attentions!
b8d6666d
fix?
cb16ef8c
fix diffllama
faf2a427
updates to fix initialization on the sam pips
6c83dcc7
ups there was a bug
bd567297
fix the last sam hq test
4213b183
fix gotocr
df766048
fix gotocr2!
a50382b3
fixes
73d74500
skip stupid tests
59ba6fab
there was one left :)
e9a3e47f
fixup
141a01ff
fix fix copies issues with this test file
cb7a8815
fix copies for sam_hq
90e36aae
rm some comments
459062f0
skip 2 more failing tests
2614116e
fix
f5695c00
fix everything
da4875af
Apply suggestions from code review
44da8484
add more doc!
c209f7ed
fix public init
4ae2049d
fix modular qwen3
7548ec23
Reviewers
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub