🔴[`Attention`] Attention refactor for Whisper-based models #38235
start refactoring whisper
0d65fecd
revert for now
45cd987c
first step
91fb2ff2
carry over attn fixes
57dd18db
check if this works
04abc5d1
whisper has an off by one somewhere - cutting mask in any interface
6f8af141
make it based on interface
7c9bc047
remove some tests that were skipped but now work
bd0618f1
some fixes for whisper tests
c46a56f0
interface changes
8c9613b0
change the order of fix
1a05fe6c
some attention adjustments for eager + TP
d293c378
fix scaling
44e050bc
mask changes
2f3036a4
why does whisper contain those extra seq lens?
b6ab4290
fix from config for fa2 as input_ids is invalid
50b85506
fix another test
0288ef9f
another fix
21515337
disable flex attn due to compile issues
00da5dbe
copies and refactor for qwen audio since it somewhat relies on whisper
ba3092a9
fix scaling and smaller things
97444913
vasqu
marked this pull request as ready for review 210 days ago
retrigger
0d5f01be
vasqu
commented
on 2025-05-21
vasqu
commented
on 2025-05-21
vasqu
commented
on 2025-05-21
vasqu
commented
on 2025-05-21
vasqu
commented
on 2025-05-21
vasqu
commented
on 2025-05-21
vasqu
commented
on 2025-05-21
vasqu
commented
on 2025-05-21
vasqu
commented
on 2025-05-21
vasqu
commented
on 2025-05-21
vasqu
commented
on 2025-05-21
vasqu
changed the title [`WIP`][`Attention`] Attention refactor for Whisper-based models [[`Attention`] Attention refactor for Whisper-based models 210 days ago
vasqu
changed the title [[`Attention`] Attention refactor for Whisper-based models [`Attention`] Attention refactor for Whisper-based models 210 days ago
Merge branch 'main' into vas-whisper-attn-refactor
0c72d880
vasqu
changed the title [`Attention`] Attention refactor for Whisper-based models 🔴[`Attention`] Attention refactor for Whisper-based models 210 days ago
gante
commented
on 2025-05-21
vasqu
commented
on 2025-05-21
new new interface version + more fixups
54ca390e
adjust qwen
7392d328
Merge branch 'main' into vas-whisper-attn-refactor
a12f2f43
vasqu
commented
on 2025-05-22
add comment
2602b0f9
forgot this one
10666c60
change copies as whisper cuts on the mask
28951f0b
add guard
eeb5cf6f
add flex attention
98836e6e
switch to new mask function + add skips for torchscript
d4b0bd8b
gante
commented
on 2025-05-23
gante
approved these changes
on 2025-05-26
remove old api with cache position
2d02cb19
Merge branch 'main' into vas-whisper-attn-refactor
395de624
last changes?
b6cf40c4
trigger ci
338921c3
vasqu
merged
badc71b9
into main 203 days ago
vasqu
deleted the vas-whisper-attn-refactor branch 203 days ago
Assignees
No one assigned
Login to write a write a comment.
Login via GitHub