[None][feat] Kimi K3: DSpark speculative decoding under Helix CP - #18166

Draft

lancelly wants to merge 15 commits into

NVIDIA:mainfrom

lancelly:user/laliao/kimi-k3-helix-dspark

Commits on Aug 31, 2026

[None][feat] Kimi K3 helix x DSpark: verify groups on the V2 ledger
lancelly
committed
2eaa4eb
[None][fix] helix x DSpark review fixes: global base, overlap-off recompute, loud fallback rejection
lancelly
committed
1f549dd
[None][fix] helix x DSpark: sanitize zero-bound verify-group rows per token
lancelly
committed
ab27100
[None][fix] address review: enable DSpark in the K3 spec-mode assert, sanitize zero-bound rows per token, drop leaked dead code
lancelly
committed
421a85f
[None][fix] helix x DSpark: CUDA-graph capture hygiene
lancelly
committed
39a44a8
[None][fix] helix x DSpark: make the embedded-drafter rejection fire, tighten metadata guards
lancelly
committed
addfde5
[None][chore] helix x DSpark: tighten the added comments
lancelly
committed
7afd98c
[None][chore] helix x DSpark: factor the shared extend-loop packing
lancelly
committed
7af0512
[None][fix] Drop the speculative_model-presence guard
lancelly
committed
b17afdd
[None][fix] Helix: account KDA state and draft KV with the repurposed mapping
lancelly
committed
65bc5e8
[None][chore] Use the canonical spec-mode check in the helix allowlist
lancelly
committed
980c1db
[None][fix] Re-thread kv_bounds through the post-#18131 kernel entry points
lancelly
committed
2c4ed6a
[None][fix] Guard helix x speculation against V1 managers and the acceptance gate
lancelly
committed
94128e5
[None][chore] Apply pre-commit formatting
lancelly
committed
af11d51
[None][fix] Make MoonViT replication helix-aware
lancelly
committed
abf746c