🚨All attention refactor🚨 #35235
Cyrilvallez
merged
99 commits into
main
from
all-attention-refactor
ArthurZucker
commented
on 2024-12-13
refactor LlamaAttention
79cb53c1
minimal changes
4bb485b4
fix llama
f3709070
update
d3ef5397
modular gemmas
45eac582
modular nits
e52af494
modular updates
5ed37aee
nits
38cafc1a
simplify
a862eacd
gpt2
5639b81a
more modualr and fixes
452d8edc
granite
81a0b664
modular modular modular
bc72c3f5
nits
48caa890
update
df68dd0d
qwen2 + starcoder2
0325dc46
mostly gemma2
ecd814bd
Update image_processing_auto.py
f5fc638d
fix
5e56d9c0
Update modular_starcoder2.py
598b7bb5
fix
0f565fbf
remove all copied from attentions
c9ac84d4
remove gcv
d189fe74
make fix-copies
9c83d969
oups
138368ec
oups2.0
7225a4f3
fix some modulars + all copied from
a3b9195f
should be good now
8d93708e
Merge branch 'all-attention-refactor' of github.com:huggingface/trans…
3cc2b4df
Merge branch 'all-attention-refactor' of github.com:huggingface/trans…
074e469d
revert unwanted changes
54d9b954
Update modeling_decision_transformer.py
944e26e9
finish cleanup
911833f8
Update modeling_olmo.py
ea269109
consistency
bc421af3
re-add gradient checkpointing attribute
8664ddcd
fix
607e928e
style
46125952
make config necessary
20c376cb
bis
0ac9db2b
bis
349b7ab8
Update modeling_my_new_model2.py
defa88ff
is_causal attr
fbf4b552
fix
9104d0a4
remove past kv return from decoder layer
0b093400
fix
46a0df79
default rope config
aedd88a7
correctly fix rope config
57e9b49f
fix bias
fe90ec05
fix gpt2 attention output
a3f50d0f
fix test
6a92c706
fix inits
a28ad195
fix default sdpa
9bd6c948
fix default sdpa implementation
fae05e16
harmonize classes
838d211d
fix mistral
e0d10f65
fix sliding window models
b275fdc8
mixtral
71eb6a2d
be more explicit
4e257534
style
1e8712b2
fix
854537b2
several fixes
99bddf01
Update modeling_dbrx.py
2f666b35
fix test
bafa020d
olmo + phi
00a98e71
rotary
8c254112
syle
4bb2f257
phi
44ff5e3d
phi again
95f7b963
again
7d550361
kwargs
24ac9ab8
Update test_modeling_common.py
bd8ede8a
skip fx tracing tests
0d3d3e39
Update modeling_utils.py
49135d04
gemma 2
f80a2c33
again
3e461bd1
Update modeling_recurrent_gemma.py
7a882d55
gemma2
78700734
granite
5b4ebaad
style
7bdf61c6
starcoder
7d5b0b53
Update sdpa_attention.py
70ef2fd9
switch args
b8429c5e
Update modeling_mllama.py
533657c9
fix
fe20d63a
cache type tests
248a6072
ArthurZucker
commented
on 2024-12-17
gpt2
46460142
Update test_modeling_common.py
ad16b1bd
fix
1df6e29b
consistency
6c01005c
fix shape with encoder
f651cd0d
should be the last one
98b7f974
tests non model
88e2fe56
most comments
5a3bdc44
small oupsi
f3923b6e
be more explicit in modulars
a6a2ff9e
more explicit modulars
aeea33bd
CIs! it works locally
ec3bef3d
LysandreJik
approved these changes
on 2024-12-18
add kwargs to _flash_attention_forward
fc74e397
yuanyao-nv
commented
on 2025-06-30
Reviewers
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub