Repository navigation

perf(speculative): speed up DFlash/DSpark drafter training 3.49x - #2279

Merged

h-guo18 merged 19 commits into

mainfrom

haoguo/dflash-training-perf

Oct 23, 2026

Merged

perf(speculative): speed up DFlash/DSpark drafter training 3.49x#2279
h-guo18 merged 19 commits into
mainfrom
haoguo/dflash-training-perf

Commits on Oct 10, 2026
Commits on Oct 12, 2026
- refactor(speculative): reject FlexAttention conflicts at convert time; one tile rule for all layers
  
  h-guo18
  and
  claude
  committed
  cb365bd
- fix(speculative): make dflash_use_torch_compile opt-in
  
  h-guo18
  and
  claude
  committed
  24ba465
- revert(speculative): move the streaming /done fix to #2756
  
  h-guo18
  and
  claude
  committed
  38cd3e7
Commits on Oct 20, 2026
- fix(speculative): keep the draft mask_mod in bounds on padded flex rows
  
  h-guo18
  and
  claude
  committed
  ee5516a
Commits on Oct 22, 2026
- test(speculative): check DFlash2 selector coverage in bf16 after the fp32 overfit
  
  h-guo18
  and
  claude
  committed
  0f13196