perf(speculative): speed up DFlash/DSpark drafter training 3.12x - #2279

Draft

h-guo18 wants to merge 6 commits into

haoguo/gemma4-e4b-supportfrom

haoguo/dflash-training-perf

Commits on Aug 28, 2026

perf(speculative): compute draft attention with a FlexAttention block-sparse mask
h-guo18
committed
b27639b
perf(speculative): project teacher rows on demand and collapse the DSpark metric syncs
h-guo18
committed
dd5b7fc
perf(speculative): chunk the DSpark TVD with Tensor.split instead of slicing
h-guo18
committed
1e288e5
perf(speculative): pin the draft's block count so the compiled attention has one shape
h-guo18
committed
31b2514

Commits on Aug 29, 2026