[TRTLLM-15293][perf] Add tiered GVR CuTe DSL top-k decode kernels (stacked on #16457) - #16877

Merged

lfr-0531 merged 19 commits into

NVIDIA:mainfrom

longcheng-nv:perf/gvr-topk-bsx-cutedsl-tiers

Aug 13, 2026

Commits on Aug 5, 2026

[None][perf] Port op43 BSX CuTe DSL top-K tiers (direct/reg/tp) onto GVR decode op

longcheng-nv
and
claude
committed
c06c8cd
[None][perf] bsx tp: hint-ladder admission fast path (R0 parity on high-hit-rate rows)

longcheng-nv
and
claude
committed
666efff
[None][perf] bsx tp admission: tighten acceptance bound to pivot-band target

longcheng-nv
and
claude
committed
c8fc03a
[None][feat] bsx tiers: MTP (next_n>1) and compress_ratio=1 support; guard narrowed

longcheng-nv
and
claude
committed
fb98f75
[None][perf] bsx tp: lean-pivot admission (stage-0c CI override + per-K interpolated tpush) and mask-hoisted streaming loops
longcheng-nv
committed
9535fd9
[None][perf] bsx dispatch: measured fallback band table to the in-tree kernel
longcheng-nv
committed
22b7af1
[None][perf] bsx dispatch: recalibrate 131072 fallback band lower bound to bs=8

longcheng-nv
and
claude
committed
78013ba
[None][test] bsx suite: cut CI wall-clock 46% by deduplicating JIT compile variants

longcheng-nv
and
claude
committed
e0401dd
[None][chore] bsx: scrub internal development codenames from comments

longcheng-nv
and
claude
committed
2a1950c
[None][chore] GVR decode: resolve the review follow-ups from #16457

longcheng-nv
and
claude
committed
4126897
[None][fix] apply ruff-format to the follow-up changes

longcheng-nv
and
claude
committed
b82f6e0
[None][fix] GVR decode: exact tie-aware terminal for boundary plateaus

longcheng-nv
and
claude
committed
4642d92
[None][perf] bsx dispatch: scope the bs=8 fallback band to true 128K shapes

longcheng-nv
and
claude
committed
cd6c106
[None][fix] gvr top-k: plateau terminal on the register-resident admission path
longcheng-nv
committed
e382f98

Commits on Aug 11, 2026

Commits on Aug 12, 2026

[None][chore] apply pre-commit formatting missed in the rename commit

longcheng-nv
and
claude
committed
9dffd71

Navigation Menu

Uh oh!

[TRTLLM-15293][perf] Add tiered GVR CuTe DSL top-k decode kernels (stacked on #16457) - #16877

[TRTLLM-15293][perf] Add tiered GVR CuTe DSL top-k decode kernels (stacked on #16457)#16877
lfr-0531 merged 19 commits into
NVIDIA:mainfrom
longcheng-nv:perf/gvr-topk-bsx-cutedsl-tiers

Commits on Aug 5, 2026

[None][perf] Port op43 BSX CuTe DSL top-K tiers (direct/reg/tp) onto GVR decode op

[None][perf] bsx tp: hint-ladder admission fast path (R0 parity on high-hit-rate rows)

[None][perf] bsx tp admission: tighten acceptance bound to pivot-band target

[None][feat] bsx tiers: MTP (next_n>1) and compress_ratio=1 support; guard narrowed

[None][perf] bsx tp: lean-pivot admission (stage-0c CI override + per-K interpolated tpush) and mask-hoisted streaming loops

[None][perf] bsx dispatch: measured fallback band table to the in-tree kernel

[None][perf] bsx dispatch: recalibrate 131072 fallback band lower bound to bs=8

[None][test] bsx suite: cut CI wall-clock 46% by deduplicating JIT compile variants

[None][chore] bsx: scrub internal development codenames from comments

[None][chore] GVR decode: resolve the review follow-ups from #16457

[None][fix] apply ruff-format to the follow-up changes

[None][fix] GVR decode: exact tie-aware terminal for boundary plateaus

[None][perf] bsx dispatch: scope the bs=8 fallback band to true 128K shapes

[None][fix] gvr top-k: plateau terminal on the register-resident admission path

Commits on Aug 11, 2026

[None][fix] bsx review round: order_row accept-and-ignore, device-derived wave budget, env soft-fail, shared tie-aware checker, 16-bit exact-tail default revert

[None][chore] conftest: explicit optional type for ref_vals_cache (RUF013)

[None][chore] bsx: memoize cluster-cap verdict, doc fallback-bands knob, scrub leftover codenames; pin gvr tests to the in-tree path

[None][chore] drop the BSX working name: the tiers are GVR optimizations, not a different algorithm

Commits on Aug 12, 2026

[None][chore] apply pre-commit formatting missed in the rename commit