[new model] step 3.7 #46658
itazap
merged
49 commits into
main
from
step_3_7_flash
mini model original
46674d81
mini model and modular
f7812c0e
update modular
9fc9f9bf
update moe logic to match standard
c9ded186
updates debug scripts
49f3fa67
original code, debug scripts, mini model, etc.
d0574dac
fix sparseblockmoe and decoderlayer to match standards, by converting…
19ba9f49
slim down moe modular, fix config conversion
b8151fe5
rm Step3p7CausalLMOutputWithPast, unused functions, return BaseModelO…
74a7fa5f
standard renaming
83c3f321
tests and siglip vision encoder
2bda59ba
build mini checkpoint in right format + conversion script
3912d271
new demo and compare outputs workflow
6f114093
major cut of modular, no image_embeds, etc.
659e4b63
fix tests
ece9810e
remove debug files
766d08d8
address feedback - add processor and config to modular
56dbd962
apply more feedback, reduce image processing
4ab537cb
apply round 2 of feedback!
4ed32c3c
more feedback and weight converter
4010606a
feedback
ccbe0073
gemma3attention and rm SharedExperts class
a9bcb989
Untrack local debug scripts
27e5927d
Untrack local debug scripts p2
a957f165
Untrack local debug scripts p3
c091fcb4
Revert core model loading changes
113eaa5a
standard rope
80b145ad
clean up
032a14c8
add tok
1cd75ae2
tests integration
6ac8f272
docs and tests unskip
c9f01aaf
tests
877f1507
laguana parents
d3461fc0
fix step3p5_vision conversion after rebase onto main
49964f44
add checkpoint links to step3p7 config docstrings
73247d97
allowlist step3p7's unused legacy config attributes, add contribution…
b9495d08
more feedback / cleanup!
09c24c15
modular conversion missed
722c9d4a
modualr
3047100d
mini checkpoint
96e333e5
feedback
f71103c2
modeling from converter
4566026b
revert core model loading changes from rebase
7d1867d0
config attrs
c18dec03
date md
fac4b0ea
Drop redundant num_heads arg from Step3p7Attention
a1c1386d
add date
a7b7698f
Merge branch 'main' into step_3_7_flash
0a39205f
rebase and date
c3c84840
Reviewers
Assignees
No one assigned
Labels
New model
Milestone
No milestone
Login to write a write a comment.
Login via GitHub