transformers
[generate] Completely stop relying on `cache_position` to prepare inputs
#44130
Merged
Merged
[generate] Completely stop relying on `cache_position` to prepare inputs #44130
Cyrilvallez
merged
29 commits into
main
from
generate-input-prep
generalize
7c7e0d8f
better doc
5bb56b6f
don't mixup the logic everywhere
283b721f
doc
c609fd7d
fix all
05309c00
no unsqueeze
ba392f33
more fixes
28d2b41b
add position_ids
4ef23932
simplify all a lot
babbd8f6
doc
6286638a
small oupsi
6aae383c
more small fixes
71b54f15
more fixes
dd3215db
fix assisted decoding
44c24e58
slice
8c783fcc
fix
2baa7a35
remove outdated test
5d761fc0
fix openai
5fc5ba2c
fix mamba2
40a4390e
fix xlm
22ab44eb
fixes
d537d262
fix
a00608af
fix rag
f5636b87
Cyrilvallez
changed the title
Cyrilvallez
changed the title 
can restart from sliced inputs with a full mask
0714af16
fix test
cabbd7f5
zucchini-nlp
approved these changes
on 2026-02-20
skip whisper test for now
20431e03
remove tests for deprecated hub-specific generation generations
53e9b1dc
fix llava test - not related to PR but was failing...
ac7a4fe7
style
ab8ffcfa
Login to write a write a comment.
Login via GitHub