[`BERT`] Add support for sdpa #28802
amyeroberts
merged
22 commits into
huggingface:main
from
hackyon:sdpa-bert
hackyon
commented
on 2024-01-31
fxmarty
commented
on 2024-02-02
fxmarty
commented
on 2024-02-02
hackyon
commented
on 2024-02-08
hackyon
commented
on 2024-02-08
hackyon
commented
on 2024-02-08
hackyon
commented
on 2024-02-08
fxmarty
approved these changes
on 2024-02-19
hackyon
commented
on 2024-02-22
Adding SDPA support for BERT
ed6efd0c
Using the proper input name for testing model input in inference()
0f34d589
Adding documentation for SDPA in BERT model page
c85acbd2
Use the stable link for the documentation
5ce07b3f
Adding a gate to only call .contiguous() for torch < 2.2.0
288cc1d5
Additions and fixes to the documentation
2afd61f5
Minor updates to documentation
fa8b5ad3
Adding extra requirements needed for the contiguous() bug
05d5c4e1
Adding "Adapted from" in plcae of the "Copied from"
95ec569c
Add benchmark speedup tables to the documentation
a07fd890
Minor fixes to the documentation
64334c10
Use ClapText as a replacemenet for Bert in the Copied-From
6a7376d4
Some more fixes for the fix-copies references
5ddb6e15
Overriding the test_eager_matches_sdpa_generate in bert tests to not …
35577eba
Undo changes to separate test
0e62fe00
Refactored SDPA self attention code for KV projections
5c64480d
Change use_sdpa to attn_implementation
9a9bb9b4
Merge remote-tracking branch 'upstream/main' into sdpa-bert
1a0af20b
ArthurZucker
approved these changes
on 2024-04-05
Merge remote-tracking branch 'upstream/main' into sdpa-bert
09653991
Merge remote-tracking branch 'upstream/main' into sdpa-bert
b4813a02
Merge remote-tracking branch 'upstream/main' into sdpa-bert
e312cd12
Fix test_sdpa_can_dispatch_on_flash by preparing input (required for …
66a24c1e
Reviewers
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub