Generate: Add new decoding strategy "DoLa" in `.generate()` #29619
gante
commented
on 2024-03-13
gante
commented
on 2024-03-20
gante
approved these changes
on 2024-03-21
add support for DoLa decoding
5c358932
add docs; remove deprecated function
bb66df53
add test code for DoLa decoding
9769d6d3
update docs and paper link
4dea208c
solved the issues that made tests failed on CircleCI
1bcdf79b
ruff reformatted
c31d6646
update DoLa decoding; test cases for llama/mistral/mixtral/gemma; docs
8ea188b5
fix formatting; fix failed test cases
76708c88
fix other test cases
419be60a
ruff reformatted; mamba cache issue unsolved
69a11e6d
remove keyword argument 'model_inputs' to match upstream changes
87fc406c
improve documentation
efe3e34c
fixed suggestions from @gante
8654787b
ruff reformated
7fcf9903
moved config warning of dola generation from `utils.py` to `configura…
21a646a6
fixed suggestions from @amyeroberts
2b81d731
fixed format issue; removed print; added explanation
cc067cca
remove trailing whitespace
679afeac
ruff reformat to pass test
7e9367b3
fixed suggestions from @amyeroberts on Mar 28
4bd54c8f
fix failed CI tests
92de5cd7
ruff reformatted; fixed missing argument generation_config
cff56616
make `dola_layers` not optional
ba61bf47
fix divergence w main
87ea8d8c
fix dola test on mamba
57c89afd
rwkv test (wont fix
1c449006
slow tests running in fp16
9d9f8945
gante
force pushed
to
9d9f8945
1 year ago
make fixup
a8993ef5
remove redundant fn
aaf560f9
final rebase divergences
dc2192c3
this one was missing
520202db
a few more nits
ce64a5f1
skip stateful models
8b6653cb
gante
merged
d094d8d9
into main 1 year ago
Assignees
No one assigned
Login to write a write a comment.
Login via GitHub