llama.cpp
model: add DSpark speculator support for DeepSeek-V4 targets
#25683
Closed
Closed
model: add DSpark speculator support for DeepSeek-V4 targets #25683
YanissAmz
wants to merge
18 commits into
ggml-org:master
from
YanissAmz:dspark-dsv4
spec: add DSpark speculative decoding
8006ae30
spec: read draft block size in the dflash impl
3e048b82
docs: add DSpark section to speculative.md
5bfd3219
spec: keep dspark block size read in the dspark impl
f2bd1906
dspark : add TODOs for incomplete parts
576908bd
spec: fold DSpark into the DFlash arch
c52ba2b2
spec: add confidence-based draft pruning for DSpark
42865369
fold the dspark impl into dflash, selected by spec type
80d1d08f
address review comments
dc900904
dspark: clean up and improve naming
66f37202
remove trailing whitespace
27cc3bae
convert: support DSpark draft models with a DeepSeek-V4 backbone
0ac0c5f5
spec: support DSpark draft models with a DeepSeek-V4 backbone
f704c0e3
spec: dflash draft sampler was discarding its sampled token
8e0a7d02
spec: dflash draft did not purge the previous block's noise K/V
913fde39
spec: dflash injected target states on top of the stale draft region
b4f05f9d
server: draft acceptance stats were wrong whenever the target could n…
e7383bba
dflash: use the DeepSeek-V4 rotary pairing for a DeepSeek-V4 backbone
c2e51866
Reviewers
No reviews
Assignees
No one assigned
Labels
documentation
model
server
conversion
Milestone
No milestone
Login to write a write a comment.
Login via GitHub