feat(LightSpec): Support LightSpec Dynamic Speculative Decoding with EAGLE3, DSpark, and DFlash - #1465

Open

shihaobai wants to merge 31 commits into

mainfrom

Commits on Aug 4, 2026

feat: add unified LightSpec speculative decoding
flyinglandlord
committed
b1623ca
add qwen3.5 dflash support
Owleye4
committed
0b00906

Commits on Aug 5, 2026

remove dflash dynamic draft & add qwen35 dspark
shihaobai
committed
464bc48

Commits on Aug 10, 2026

WIP: simplify LightSpec implementation (not ready for review)
shihaobai
committed
8a75a1f
fix: read Qwen3.5 target layers from text config
shihaobai
committed
6aea8ce

Commits on Aug 11, 2026

refactor: simplify LightSpec draft integration
shihaobai
committed
3f2ee16
refactor: simplify speculative decoding pipeline
shihaobai
committed
9aa6dfc
fix: align speculative draft inputs and DSpark RoPE
shihaobai
committed
f5293ba

Commits on Aug 13, 2026

refactor: simplify dynamic speculative scheduling
shihaobai
committed
f713b82
Merge remote-tracking branch 'origin/main' into lightspec
shihaobai
committed
640036f

Commits on Aug 14, 2026

Commits on Aug 17, 2026

refactor: isolate hidden collector inference state
hiworldwzj
committed
fe70b2f
refactor: reuse forward in autotune warmup
hiworldwzj
committed
249667f
fix
hiworldwzj
committed
55cf376
refactor: initialize decode group metadata on demand
hiworldwzj
committed
56421db
refactor: split FA3 decode state initialization
hiworldwzj
committed
b25e38f
refactor: move FA3 causality into attention state
hiworldwzj
committed
8f32f61
refactor: align linear attention MTP state naming
hiworldwzj
committed
538189a
refactor: clarify MTP SSM buffer shape
hiworldwzj
committed
552b02a
refactor: share GPU attention workspaces
hiworldwzj
committed
a423012
fix
hiworldwzj
committed
4e18cee
refactor: move FA3 utility test to file end
hiworldwzj
committed
027ce10
refactor: align MTP utility naming
hiworldwzj
committed
bf339c1

Commits on Aug 18, 2026