[TRTLLM-14814][feat] Kimi K3 serving parsers, chat template, and speculative decoding (suffix automaton + DFlash scaffold) - #17327

Merged

brnguyen2 merged 15 commits into

NVIDIA:mainfrom

brnguyen2:k3/14812-14814-serving-specdec

Aug 8, 2026

Commits on Aug 8, 2026

[TRTLLM-14812][feat] Add Kimi K3 reasoning parser, tool parser, and chat template support

moraxu
authored and
brnguyen2
committed
fbcdc4e
[TRTLLM-14814][feat] Kimi K3 speculative decoding: suffix automaton and DFlash scaffold
brnguyen2
committed
b8e5050
[TRTLLM-14814][fix] Restrict DSpark non-causal validation to the block-decode path
brnguyen2
committed
85deb14
[TRTLLM-14814][fix] Skip specdec tests whose companion files ship in follow-up PRs
brnguyen2
committed
defbc79
[TRTLLM-14812][fix] Update api_stability CLI references for kimi_k3 parsers
brnguyen2
committed
d54ee0c
[TRTLLM-14814][fix] Address review comments on specdec harness and tests
brnguyen2
committed
83cba1b
[TRTLLM-14814][fix] Guard graph-safe MLA latent-cache append to hybrid KV managers
brnguyen2
committed
a2430fa
[TRTLLM-14814][chore] Reference TRTLLM-15193 in the MLA append guard TODO
brnguyen2
committed
559601d
[TRTLLM-14814][chore] Note TRTLLM-15199 at the warmup-row clamp in the graph-safe append
brnguyen2
committed
3bba8cb
Address trivial review comments
brnguyen2
committed
6bda014
[TRTLLM-14814][chore] Address review comments: clarify DSpark None placeholders and K3 MoE backend pinning
brnguyen2
committed
6d57079
[TRTLLM-14814][chore] Warn that DFlash accept-stats recording is experimental
brnguyen2
committed
038537f
[TRTLLM-14814][doc] Update K3 docs for serving/spec-decode landing: drop chat-completions dependency note, mark SA speculation supported for aggregated serving
brnguyen2
committed
f9376ca
Merge branch 'main' into k3/14812-14814-serving-specdec
brnguyen2
authored
51897d4
Merge branch 'main' into k3/14812-14814-serving-specdec
brnguyen2
authored
fac670d

Navigation Menu

Uh oh!

[TRTLLM-14814][feat] Kimi K3 serving parsers, chat template, and speculative decoding (suffix automaton + DFlash scaffold) - #17327

[TRTLLM-14814][feat] Kimi K3 serving parsers, chat template, and speculative decoding (suffix automaton + DFlash scaffold)#17327
brnguyen2 merged 15 commits into
NVIDIA:mainfrom
brnguyen2:k3/14812-14814-serving-specdec

Commits on Aug 8, 2026

[TRTLLM-14812][feat] Add Kimi K3 reasoning parser, tool parser, and chat template support

[TRTLLM-14814][feat] Kimi K3 speculative decoding: suffix automaton and DFlash scaffold

[TRTLLM-14814][fix] Restrict DSpark non-causal validation to the block-decode path

[TRTLLM-14814][fix] Skip specdec tests whose companion files ship in follow-up PRs

[TRTLLM-14812][fix] Update api_stability CLI references for kimi_k3 parsers

[TRTLLM-14814][fix] Address review comments on specdec harness and tests

[TRTLLM-14814][fix] Guard graph-safe MLA latent-cache append to hybrid KV managers

[TRTLLM-14814][chore] Reference TRTLLM-15193 in the MLA append guard TODO

[TRTLLM-14814][chore] Note TRTLLM-15199 at the warmup-row clamp in the graph-safe append

Address trivial review comments

[TRTLLM-14814][chore] Address review comments: clarify DSpark None placeholders and K3 MoE backend pinning

[TRTLLM-14814][chore] Warn that DFlash accept-stats recording is experimental

[TRTLLM-14814][doc] Update K3 docs for serving/spec-decode landing: drop chat-completions dependency note, mark SA speculation supported for aggregated serving

Merge branch 'main' into k3/14812-14814-serving-specdec

Merge branch 'main' into k3/14812-14814-serving-specdec