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

Merged

lfr-0531 merged 50 commits into

NVIDIA:mainfrom

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

Jul 28, 2026

Commits on Jul 15, 2026

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

longcheng-nv
and
claude
committed
ad3f9c4
[None][perf] gvr top-k decode: add P1b histogram-quantile rung placement (inert)

longcheng-nv
and
claude
committed
9878e1a
[None][perf] gvr top-k decode: R0 single-pass admission in _run_phases (secant fallback)

longcheng-nv
and
claude
committed
d2252ee
[None][perf] gvr top-k decode: inline log-falsi R1 shot on R0 miss

longcheng-nv
and
claude
committed
af5bd95
[None][perf] gvr top-k decode: fb_fix bounded log-falsi fallback (replaces secant under R0)

longcheng-nv
and
claude
committed
b28b884
[None][perf] gvr top-k decode: cluster-aware R0 (cs>1 long-row branch)

longcheng-nv
and
claude
committed
a8c69a7
[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
4d00c8a
[None][perf] gvr top-k decode: p1b_cache path (SMEM gather-cache, 16-bit)

longcheng-nv
and
claude
committed
f926cdf
[None][perf] gvr top-k decode: fix cluster-R0 admission/refine handoff to Phase 3/4

longcheng-nv
and
claude
committed
a489d80
[None][perf] gvr top-k decode: seeded cluster R0-miss refine (fold R1, no re-measure)

longcheng-nv
and
claude
committed
1ba38ec
[perf] GVR P3/P4: hoist SMEM window bases; single-pass P4 writeback on converged snap

siyidNV
and
claude
committed
ee10ccf
[perf] GVR P4 level-2 histogram refinement

siyidNV
and
claude
committed
4e69280
[perf] GVR P4 multi-level refinement + lane-parallel bin-search step1

siyidNV
and
claude
committed
d29d6f7
[perf] GVR P4: bracket-window histogram + zero-atomic converged writeback

siyidNV
and
claude
committed
ff03622
[fix] GVR P4: clamp histogram range overflow to keep threshold ordered

siyidNV
and
claude
committed
1a44fff
[fix] GVR: parity-buffer the cluster all-reduce DSMEM count exchange

siyidNV
and
claude
committed
ac1c5f8
[None][perf] gvr top-k decode: add fused rank-and-scatter P4 (inert flag)
longcheng-nv
committed
c528253
[None][perf] gvr top-k decode: enable rank-scatter P4 by default under R0
longcheng-nv
committed
6002551
[None][perf] gvr top-k decode: p1b_cache ON for all dtypes at cluster_size>1
longcheng-nv
committed
39f17b2
[perf] GVR P2/P4: redundant-warp sync reduction

siyidNV
and
claude
committed
b8b616b
[None][chore] Address review: float-domain bin clamp; document knobs

siyidNV
and
claude
committed
7b31f3c
[None][chore] GVR top-K decode: ruff-format the R0 additions
longcheng-nv
committed
3e20622
[None][perf] GVR top-K decode: enable R0 admission by default + equivalence test
longcheng-nv
committed
95a8388

Commits on Jul 16, 2026

Commits on Jul 17, 2026

Merge branch 'main' into perf/gvr-kernel-opts
siyidNV
authored
0560d0c
[None][chore] Fix ruff-format pre-commit violations

longcheng-nv
and
claude
committed
af7351c
[None][test] Preserve -1 encoding for true index 0 in R0 pre_idx helper

longcheng-nv
and
claude
committed
7408f3b

Commits on Jul 19, 2026

[None][perf] gvr_topk_decode: K2048 R0 low rung 0.85 -> 0.6 (real-content recalibration)

longcheng-nv
and
claude
committed
0d6fc4f

Commits on Jul 20, 2026

[None][perf] gvr_topk_decode: K2048 R0 Phase-4 histogram diet 2048 -> 512 bins

longcheng-nv
and
claude
committed
1128c05
[None][perf] gvr_topk_decode: tiny-tie fast path for p4_exact_tail

longcheng-nv
and
claude
committed
e6fdbfa
[None][chore] Address review: tidy comments, parameterize SMEM scalar helpers

siyidNV
and
claude
committed
6fa4b5c
Merge branch 'main' into perf/gvr-kernel-opts
siyidNV
authored
d693f33
Merge branch 'main' into perf/gvr-kernel-opts
siyidNV
committed
9b7f0bf

Commits on Jul 21, 2026

Merge branch 'main' into perf/gvr-kernel-opts
siyidNV
authored
f485738
Merge branch 'main' into perf/gvr-kernel-opts
siyidNV
authored
9269219
Merge PR #16424 (perf/gvr-kernel-opts) into perf/gvr-topk-r0-histogram-ladder

longcheng-nv
and
claude
committed
6140078
[None][chore] gvr_topk_decode: ruff-format cleanup (fixes CI Release-Check)

longcheng-nv
and
claude
committed
b14ec40
[None][fix] gvr_topk: equalize cs=1/cs>1 SMEM layouts for the LB hybrid kernel

longcheng-nv
and
claude
committed
a707cfe

Commits on Jul 22, 2026

Commits on Jul 23, 2026

Merge remote-tracking branch 'origin/main' into perf/gvr-topk-r0-histogram-ladder
longcheng-nv
committed
dc2bdc8