Mamba2 conversion script for original models #32580
ArthurZucker
merged
17 commits into
huggingface:main
from
vasqu:base-mamba2-conversion
vasqu
commented
on 2024-08-10
vasqu
commented
on 2024-08-10
vasqu
commented
on 2024-08-10
molbap
commented
on 2024-08-12
vasqu
commented
on 2024-08-13
vasqu
commented
on 2024-08-13
first attempt at allowing both conversions from codestral and from th…
376621b6
allow fp16, seems default for mamba2
11bde9a5
dtype fix
fc36bc10
simplify codestral check, dont overwrite pad/eos/bos when codestral
01bed7d1
change file -> directory
22b48adb
use path join to be safe
0fd08a00
style
a2f0008c
apply code review
50dc02d8
fix copies
e98147b4
add tokenizer to docs
32ba3dfb
empty commit to check for weird err
a77d15be
make conversion user dependent on model type, defaults for original p…
ae43243d
small comment nit
52ca5494
remove norm_before_gate in conversion
abd77545
simplify model dict by using shared keys directly + remove unnecessar…
6a37735d
fix tokenization: remove separate mamba2 tokenizer, add padding optio…
42d8afc5
simplify even further as we pass padding side via **kwargs already
f57616a8
ArthurZucker
approved these changes
on 2024-08-29
Reviewers
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub