Repository navigation

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

Merged

h-guo18 merged 11 commits into

mainfrom

haoguo/dflash2-support

Sep 29, 2026

Commits on Sep 22, 2026
- [Speculative Decoding] Add the DFlash2 draft variant
  h-guo18
  committed
  9b82dd2
Commits on Sep 25, 2026
- feat(speculative): DFlash2 acceptance-annealed objective, strict top-k selector
  
  h-guo18
  and
  claude
  committed
  3967717
- feat(speculative): DFlash2 streaming example, and carry the base RoPE through the fake base
  
  h-guo18
  and
  claude
  committed
  f20a94e
- fix(speculative): make the DFlash2 streaming example runnable, and add a multi-GPU one
  
  h-guo18
  and
  claude
  committed
  1296196
- test(speculative): stop pinning a transformers-5-only layout in the RoPE tests
  
  h-guo18
  and
  claude
  committed
  b35bf76
- refactor(speculative): read the base RoPE through the exporter's reader, and guard it
  
  h-guo18
  and
  claude
  committed
  b82195b
- docs(launcher): one streaming example for DFlash2, under the plain name
  
  h-guo18
  and
  claude
  committed
  9027c42
- test(export): keep the RoPE tests collectable without transformers
  
  h-guo18
  and
  claude
  committed
  e10b003
Commits on Sep 27, 2026
- fix(speculative): give the fake base's rope_parameters a rope_type
  
  h-guo18
  and
  claude
  committed
  3d77d6c
- fix(speculative): give the fake base's rope_parameters a rope_type
  
  h-guo18
  and
  claude
  committed
  9cf116c
- docs(speculative): cut the commentary back to what a reader needs
  
  h-guo18
  and
  claude
  committed
  aaccc4b