[Speculative Decoding] DFlash2 draft variant (sublayer convolution + candidate selector) - #2216

Open

h-guo18 wants to merge 10 commits into

mainfrom

haoguo/dflash2-support

Commits on Aug 19, 2026

refactor(speculative): add no-op sublayer wrappers to the DFlash draft layer

h-guo18
and
claude
committed
1419d47
feat(speculative): DFlash2 draft variant (sublayer conv + candidate selector)

h-guo18
and
claude
committed
6b4ca0d
test(speculative): CPU unit tests for the DFlash2 draft variant

h-guo18
and
claude
committed
9975a5b
docs(changelog): note the DFlash2 draft variant

h-guo18
and
claude
committed
197dd85
feat(recipes): DFlash2 training recipe

h-guo18
and
claude
committed
13e8e2f
fix(speculative): read rope_theta from rope_parameters first

h-guo18
and
claude
committed
ba377e7
feat(export): match the published DFlash2 config contract

h-guo18
and
claude
committed
fe8d5b8
chore: drop MODEL_SPECIFIC_REFACTOR.md from this branch

h-guo18
and
claude
committed
c446f6e

Commits on Aug 23, 2026