[`Ernie 4.5`] Add ernie text models #39228
vasqu
merged
83 commits into
huggingface:main
from
vasqu:ernie4_5
init
055fa740
copied from remote
cf318494
add proper structure and llama like structure
69d942cd
fixup
0fa32aa2
revert to state that works
2fc44ee2
get closer to llama
39176fd8
slow and steady
f9d5789c
some removal
de53a33f
masks work
999e7e95
it is indeed the rope implementation, how dafuq does it mesh with the…
1816efb8
nice
d0c5877d
getting closer
03618225
closer to transformers style
06820859
let's simplify this, batching works now
a551f220
simplified
4a9c20bb
working version with modular
393c2c77
it is indeed the rotation per weights, make it complete llama style
936aa04b
cleanup conversion, next to look at -> tokenizer
b33edd1c
remove llama artefacts
4bcb7f04
fix modeling tests (common ones)
41295066
style
e837bf53
integration test + first look into tokenization (will need more work,…
be5a7b00
style
f98c8d95
working moe version, based on remote
de0389f7
lets keep it simple and go step by step - transformers annotations fo…
1a417194
more cleanup
484da46c
refactor namings and remove addition forXXX classes
5627e5a5
our moe won't cut it it seems, correction bias seems to be missing in…
a146b134
tokenization change (remote)
d1b11447
our moe version works when adding normalization :D
54749e5d
cleanup moe
c81082f4
nits
0efb8a77
cleanup modeling -> let's get to modular next
8132cf43
style
0f47d74c
modular v1
be22c9aa
minor things + attempt at conversion (which doesn't work)
522b8d9c
no conversion follow glm, fixup modular and other nits
016bc8ef
modular cleanup
ac62ca73
fixes
1f185e21
tests, tests, tests + some moe dtype forcing
d51b2575
simplify modular, fix fatal fa2 bug, remaining tests
72f1627b
fix import issue?
ba4b05e2
some initial docs, fix bnb faulty behavior --> needs to fix some test…
cce6eaa3
fix sdpa test, load on init dtype only
64fc88a1
Merge branch 'main' into ernie4_5
e9eab1fc
fixup post merge
cf37555c
style
76dc3f90
vasqu
changed the title 
fix doc links
654bcd36
Merge branch 'main' into ernie4_5
c595c3e3
Merge branch 'main' into ernie4_5
694837a8
tokenization cleanup beginnings
40a5ade3
simplify tokenizer by a lot as its basically llama
b0bb530b
tokenizer is full llama with different defaults + extra special tokens
090f78e4
sync og special tokens of ernie
29101689
fix decoding with numbers (also in remote done what a timing), begin …
90a69333
align with remote and preserve special tokens, adjust tests to ernie …
68e3c94a
nits
98a37598
docs
1eb7b9ec
Merge branch 'main' into ernie4_5
f6d768e6
my daily post merge it is
157c6227
check
4bcd77a6
tokenization update with explanations and conversion script
82539060
review on modular (til), revert some tokenizer things i did prior, re…
c91193a5
Merge branch 'main' into ernie4_5
c22e0c27
post merge fixes
b8506a41
fixup tokenization, llama fast is the way to go
49b46399
more fixups
ad76c7cf
check
c08509da
import fixes
3c0014a4
correction bias following the paddle code
2948aed6
fix
81bafc6c
Merge branch 'main' into ernie4_5
f42e3129
fix TP plan, fix correction bias sharding during forward
508b683b
Merge branch 'main' into ernie4_5
624820ca
style
b8d8f5c8
whoops
c72e4d93
fix tied weights
b9f8db7a
Merge branch 'main' into ernie4_5
680353ec
docs and last nit
60ebe117
license
a4635883
Merge branch 'main' into ernie4_5
416cda6e
ArthurZucker
approved these changes
on 2025-07-21
flasky tests
f31f41d1
move repo id, update when merged on the hub
961423da
Login to write a write a comment.
Login via GitHub