llama.cpp
model: add NVIDIA Nemotron-3-Puzzle-75B-A9B (NemotronHPuzzle) support
#25444
Merged
Merged
model: add NVIDIA Nemotron-3-Puzzle-75B-A9B (NemotronHPuzzle) support #25444
CISC
merged
11 commits into
ggml-org:master
from
YanissAmz:puzzle-port
YanissAmz
changed the title 
danbev
commented
on 2026-09-01
danbev
approved these changes
on 2026-09-02
CISC
commented
on 2026-09-02
hparams: add per-layer n_ff_exp/n_expert_used arrays with scalar-or-a…
470ee50c
nemotron-h: use per-layer n_ff_exp(il) and n_expert_used(il) at MoE c…
e8f939ea
arch/*.cpp + gguf-py: mechanical rename n_ff_exp->n_ff_exp_impl, n_ex…
555e84ad
convert: support NemotronHPuzzleForCausalLM (per-block MoE config)
9e4ce83f
nemotron-h: distinguish Nemotron 3 Puzzle (75B.A9B) from Super (120B.…
f07b9b3c
convert: accept the official Puzzle BF16 checkpoint's tensor naming
dc67d0bf
laguna: use n_ff_exp_impl for the uniform-MoE FFN size
3e1786ab
arch: extend the n_ff_exp/n_expert_used rename to archs added upstream
38d39068
convert: keep Puzzle opted out of the NemotronH MTP export path
293541f5
llama: replace n_ff_exp/n_expert_used scalars with per-layer accessors
9d09b3a6
llama: restore per-key required flags on the expert hparam reads
b257ab90
CISC
approved these changes
on 2026-09-02
CISC
commented
on 2026-09-03
Login to write a write a comment.
Login via GitHub