[`Mxfp4`] Add a way to save with a quantization method #40176
ArthurZucker
merged
36 commits into
main
from
save-post-quantize
add a test
bc656531
tempdir
ff5b81a4
fix import issue[
86a7d473
wow I am tired
d1f1533d
properly init
05a379ca
i am not super familiar with quantizer api :|
4c69662e
set to TRUE fro now
a325d2d7
full support
d9e8845e
push current changes
8b019870
will clean this later but the imports are a shitshow here
5d9a0040
winglian
commented
on 2025-08-16
winglian
commented
on 2025-08-16
this correctly saves the block and scales but forward seems broken
75616fa3
quanitze was not correct
069d1ad8
fix storage
ed0049c7
why were bias even included
825f3d02
finally!
f9cc70e6
style
d42f27ab
fix style
5570c5fa
remove print
f0c14523
lazy import
cf167895
up
bb84ae10
not sure what happens this works now?
8ef69e23
holy molly it was not so far
131b9028
okay this seems to work!
988cdd92
workings!!!
fd040090
allow save_pretrained to create PR
85e982cd
Apply suggestions from code review
59f75816
fixup
e0839c9c
add deqyabtze fakse as wek
b05218a3
working new
c250b05f
fix
a5aadbe4
ArthurZucker
commented
on 2025-08-21
rm swizzle and unswizzle during saving
9b575d8e
rm print
a8fa97e3
ArthurZucker
commented
on 2025-08-21
Update src/transformers/modeling_utils.py
a698e179
SunMarc
approved these changes
on 2025-08-21
Merge remote-tracking branch 'upstream/main' into save-post-quantize
ff1a1a0b
fix
54be1a15
style
dfc9ef3d
Login to write a write a comment.
Login via GitHub