Add support for SmallThinker model series #14898
CISC
merged
18 commits into
ggml-org:master
from
wdl339:smallthinker
support smallthinker
efe27eb5
support 20b softmax, 4b no sliding window
a6d6eafe
Merge branch 'master' into smallthinker
a5274b79
new build_moe_ffn_from_probs, and can run 4b
8e2cb21f
fix 4b rope bug
e28d2c56
Merge branch 'master' into smallthinker
ebd78ccc
fix python type check
8c6af02a
remove is_moe judge
92b518b4
ggerganov
commented
on 2025-07-27
remove set_dense_start_swa_pattern function and modify set_swa_patter…
4186babc
trim trailing whitespace
f1d4698f
remove get_vocab_base of SmallThinkerModel in convert_hf_to_gguf.py
f10cd467
better whitespace
4af8b591
use GGML_ASSERT for expert count validation
e2c900ce
Improve null pointer check for probs
594af993
use template parameter for SWA attention logic
29e1fe0a
better whitespace
5d09d11b
move the creation of inp_out_ids before the layer loop
bb3dd583
remove redundant judge for probs
e338c30c
CISC
approved these changes
on 2025-07-28
Login to write a write a comment.
Login via GitHub