[None][perf] GVR top-K decode: enable R0 histogram-ladder admission by default#16457

Open

longcheng-nv wants to merge 20 commits into

NVIDIA:mainfrom

longcheng-nv:perf/gvr-topk-r0-histogram-ladder

Commits on Jul 15, 2026

[None][perf] gvr top-k decode: add block_count_ge_multi + R0 knobs (inert)

longcheng-nv
and
claude
committed
[None][perf] gvr top-k decode: add P1b histogram-quantile rung placement (inert)

longcheng-nv
and
claude
committed
[None][perf] gvr top-k decode: R0 single-pass admission in _run_phases (secant fallback)

longcheng-nv
and
claude
committed
[None][perf] gvr top-k decode: inline log-falsi R1 shot on R0 miss

longcheng-nv
and
claude
committed
[None][perf] gvr top-k decode: fb_fix bounded log-falsi fallback (replaces secant under R0)

longcheng-nv
and
claude
committed
[None][perf] gvr top-k decode: cluster-aware R0 (cs>1 long-row branch)

longcheng-nv
and
claude
committed
[None][perf] gvr top-k decode: R0 dispatch defaults ((0.85,0.35) qfracs + kC-diet + p1b_cache dtype-gate)

longcheng-nv
and
claude
committed
[None][perf] gvr top-k decode: p1b_cache path (SMEM gather-cache, 16-bit)

longcheng-nv
and
claude
committed
[None][perf] gvr top-k decode: fix cluster-R0 admission/refine handoff to Phase 3/4

longcheng-nv
and
claude
committed
[None][perf] gvr top-k decode: seeded cluster R0-miss refine (fold R1, no re-measure)

longcheng-nv
and
claude
committed
[None][perf] gvr top-k decode: add fused rank-and-scatter P4 (inert flag)
longcheng-nv
committed
[None][perf] gvr top-k decode: enable rank-scatter P4 by default under R0
longcheng-nv
committed
[None][perf] gvr top-k decode: p1b_cache ON for all dtypes at cluster_size>1
longcheng-nv
committed
[None][chore] GVR top-K decode: ruff-format the R0 additions
longcheng-nv
committed
[None][perf] GVR top-K decode: enable R0 admission by default + equivalence test
longcheng-nv
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[None][perf] GVR top-K decode: enable R0 histogram-ladder admission by default#16457

[None][perf] GVR top-K decode: enable R0 histogram-ladder admission by default#16457
longcheng-nv wants to merge 20 commits into
NVIDIA:mainfrom
longcheng-nv:perf/gvr-topk-r0-histogram-ladder

Commits on Jul 15, 2026

[None][perf] gvr top-k decode: add block_count_ge_multi + R0 knobs (inert)

[None][perf] gvr top-k decode: add P1b histogram-quantile rung placement (inert)

[None][perf] gvr top-k decode: R0 single-pass admission in _run_phases (secant fallback)

[None][perf] gvr top-k decode: inline log-falsi R1 shot on R0 miss

[None][perf] gvr top-k decode: fb_fix bounded log-falsi fallback (replaces secant under R0)

[None][perf] gvr top-k decode: cluster-aware R0 (cs>1 long-row branch)

[None][perf] gvr top-k decode: R0 dispatch defaults ((0.85,0.35) qfracs + kC-diet + p1b_cache dtype-gate)

[None][perf] gvr top-k decode: p1b_cache path (SMEM gather-cache, 16-bit)

[None][perf] gvr top-k decode: fix cluster-R0 admission/refine handoff to Phase 3/4

[None][perf] gvr top-k decode: seeded cluster R0-miss refine (fold R1, no re-measure)

[None][perf] gvr top-k decode: add fused rank-and-scatter P4 (inert flag)

[None][perf] gvr top-k decode: enable rank-scatter P4 by default under R0

[None][perf] gvr top-k decode: p1b_cache ON for all dtypes at cluster_size>1

[None][chore] GVR top-K decode: ruff-format the R0 additions

[None][perf] GVR top-K decode: enable R0 admission by default + equivalence test

Commits on Jul 16, 2026

[None][perf] GVR top-K decode: pick_config launch-shape policy + launch helper

[None][perf] GVR top-K decode: virtual seed rung (r0_vseed) + per-K rung defaults

[None][fix] GVR top-K decode: exact tie resolution in the P4 straddling fine bin

Commits on Jul 17, 2026

[None][chore] Fix ruff-format pre-commit violations

[None][test] Preserve -1 encoding for true index 0 in R0 pre_idx helper