[TRTLLM-15293][perf] Add self-sampling (GVR V2) top-K decode kernels - #17821

Merged

juney-nvidia merged 42 commits into

NVIDIA:mainfrom

longcheng-nv:feat/gvr-selfsampling-topk

Aug 26, 2026

Commits on Aug 17, 2026

[None][perf] Add self-sampling GVR top-K decode kernels (CuTeDSL, standalone)
longcheng-nv
committed
1fb56b8

Commits on Aug 18, 2026

[TRTLLM-15293][perf] self-sampling top-K: fix pre_idx hint contract — raw indices, uniform across DSv3.2/Flash/Pro
longcheng-nv
committed
29354d1
[TRTLLM-15293][perf] self-sampling top-K: n <= topK short path (production pad convention)
longcheng-nv
committed
b1c5e67
[TRTLLM-15293][perf] self-sampling top-K: opt-in values output (default off)
longcheng-nv
committed
3953e90
[TRTLLM-15293][test] self-sampling top-K: degenerate-hint hardening tests
longcheng-nv
committed
5877b60
[TRTLLM-15293][perf] self-sampling top-K: run_varlen production-contract entry (reference engine)
longcheng-nv
committed
6103232
[TRTLLM-15293][perf] self-sampling top-K: two-time-scale dispatch split (route_static/route_dynamic)
longcheng-nv
committed
0349d15
[TRTLLM-15293][perf] self-sampling top-K: per-row in-kernel varlen engine (gvr_main port)
longcheng-nv
committed
cb23e1d
[TRTLLM-15293][test] self-sampling top-K: CUDA-graph capture/replay test for the varlen engine
longcheng-nv
committed
4ea46e5
[TRTLLM-15293][perf] self-sampling top-K: opt-in dsa.py decode seam (TRTLLM_GVR_SELF_SAMPLING=1)
longcheng-nv
committed
7986d32
[TRTLLM-15293][fix] self-sampling top-K: short-row sentinel must survive the TGT*2 scan target
longcheng-nv
committed
b54eb2b
[TRTLLM-15293][fix] self-sampling top-K: adversarial-review hardening of the varlen engine and seam
longcheng-nv
committed
2efd66e

Commits on Aug 19, 2026

Commits on Aug 20, 2026

Commits on Aug 21, 2026

Commits on Aug 22, 2026

[TRTLLM-15293][perf] self-sampling top-K: cheapen the per-row varlen mirror chain

longcheng-nv
and
claude
committed
80484e8
Merge branch 'main' into feat/gvr-selfsampling-topk

longcheng-nv
and
claude
committed
dc7822f
Merge branch 'main' into feat/gvr-selfsampling-topk (round 2)

longcheng-nv
and
claude
committed
3b459c9
[TRTLLM-15293][chore] self-sampling top-K: pre-commit formatting for the TopK module edit

longcheng-nv
and
claude
committed
7fe0d51
[TRTLLM-15293][chore] self-sampling top-K: lint fixes for the kernel module

longcheng-nv
and
claude
committed
e126c4c
[TRTLLM-15293][chore] self-sampling top-K: ruff-format line wrapping in the register family

longcheng-nv
and
claude
committed
3584c42

Commits on Aug 23, 2026

Commits on Aug 24, 2026

[TRTLLM-15293][chore] self-sampling top-K: slim code comments

longcheng-nv
and
claude
committed
41d61c9
[TRTLLM-15293][chore] self-sampling top-K: comment slim tail

longcheng-nv
and
claude
committed
ff08b1d

Commits on Aug 25, 2026

[TRTLLM-15293][chore] self-sampling top-K: dedupe the per-row prologue contract comments
longcheng-nv
committed
62cfe5d