[None][perf] Add BSX multi-tier CuTe DSL top-k decode kernels (stacked on #16457) by longcheng-nv · Pull Request #16877 · NVIDIA/TensorRT-LLM

longcheng-nv · 2026-07-26T07:25:49Z

Description

Rebased onto main (#16457 merged as f3d4c85461). Six commits: the four reviewed bsx-tier commits (e92519d226 tier port, a196684b8a tp admission fast path, a8f4899bcf acceptance-bound fix, 7c9cd2a4a3 MTP/compress_ratio=1) plus the op43-pr6 campaign result: 59a019bb70 lean-pivot admission + mask-hoisted streaming loops, 78df9cbb90 measured fallback band table + tests.

Adds the BSX multi-tier CuTe DSL top-K decode kernel family as a guarded fast path inside the existing trtllm::cute_dsl_gvr_topk_decode custom op (op signature unchanged, no production call-site change — DSL-path activation remains the deferred follow-up, same as #16457):

gvr_topk_decode_bsx_direct.py — short-row (npad <= 12288) single-CTA exact radix tier
gvr_topk_decode_bsx_reg.py — register-resident tier (one-time row load, rank-scatter P3, packed-u64 candidates)
gvr_topk_decode_bsx_tp.py — throughput tier (sampled pivot + multi-rung secant + fused count/collect)
gvr_topk_decode_bsx_dispatch.py — route table + v1 guard; anything outside the guard falls back to the existing [None][perf] GVR top-K decode: enable R0 histogram-ladder admission by default #16457 kernel unchanged

Dispatcher guard: fp32, no load-balance (order_row/counters), K in {512, 1024, 2048}, npad <= 262144, cluster size within _query_max_cluster_size(). MTP (next_n >= 1) and compress_ratio in {1, 4} are fully supported (4th commit): hint-row sharing row // next_n, cr=1 temporal hint offset, per-row N_eff — all mirroring the in-tree kernel's semantics line-by-line; next_n==1/cr==4 hot paths are constexpr-guarded to compile to exactly the previous code. MTP correctness is fully unit-tested (differential vs the in-tree kernel per row); MTP performance validation is deferred to e2e integration — the full-grid re-measure below confirms the next_n=1 envelope is unchanged (9490/9515 cases flat within 5pp vs the pre-MTP head).

Production-contract adaptations (vs the op-bench origin): ragged per-row N_eff lane masking (poisoned-tail safe), pre_idx out-of-range clamping (all-zero cold start safe), in-kernel degenerate-row (N_eff <= K) identity emit with -1 padding, cluster-size clamp with dispatcher fallback.

Evidence (local, B200; text tables only — no profiler artifacts)

Benchmark: 865 real DSv4/V3.2 decode-capture cells x 11 batch sizes = 9515 paired same-rep cold-L2 nsys cases, 0 skips (every cell falls inside the v1 guard), 0 exactness mismatches.

Direct A/B: this PR's head (pr6: 4 tier commits + lean-pivot + band table) vs the #16457 kernel now on main — full 865x11 real decode-capture grid (9515 same-rep cold-L2 nsys pairs), re-measured 2026-07-28:

slice	bare kernels gm	shipped operator gm (band table on)
overall	1.4170x (win 88.0%)	1.3966x, worst case capped at 1.10x
v32 (K=2048)	1.4397x	1.4068x
pro (K=1024)	1.4053x	1.4000x
flash (K=512)	1.3858x	1.3703x

Bare-kernel loss tail vs the in-tree head: <0.909 cases 636 (pr4) -> 517; the shipped operator routes the 27 (npad, bs) buckets containing them (30.9% of grid cases, the L2-resident mid-N regime) back to the in-tree kernel — the operator is a strict Pareto improvement: routed buckets run at parity, everything else keeps the bsx win, floor 0.9125. The pr6 campaign additionally falsified (measured: 16-subset ablation + component fingerprints): exact-count admission as a wholesale replacement (gain and harm co-sourced), K-scaled kC diet (pure harm), ladder-quantile re-spread, CS occupancy cuts, and multi-pass straggler hypotheses (the residual band is 1-pass; the gap is candidate-set fatness — hence the lean pivot).

fp32 gm by model x batch size (bare kernels; the shipped-operator numbers differ only in the routed mid-N columns by 2-5pp):

model \ BS	1	2	4	8	16	32	64	128	256	512	1024
v32 (K=2048, cr=1)	1.80	1.78	1.76	1.56	1.28	1.27	1.27	1.29	1.32	1.34	1.33
flash (K=512, cr=4)	1.64	1.63	1.62	1.49	1.29	1.29	1.27	1.30	1.26	1.28	1.26
pro (K=1024, cr=4)	1.69	1.67	1.65	1.53	1.30	1.31	1.28	1.30	1.27	1.29	1.28

Layer dimension (per-layer gm across that layer's own 9515-grid slice — the loss tail is a per-(layer, shape) phenomenon, not a per-layer one): all 109 production layers win at layer granularity, none below 1.28x — v32: 58 layers, worst L26 = 1.286 / median 1.444 / best L52 = 1.544; flash: 21 layers, worst L06 = 1.297 / median 1.380 / best L24 = 1.490; pro: 30 layers, worst L16 = 1.303 / median 1.388 / best L10 = 1.594.

dtype scope. All numbers above are fp32 — the production DSA indexer logits path, and the bsx guard intentionally restricts the fast path to fp32: bf16/fp16 calls flow to the in-tree kernel unchanged (functional support only in this PR, verified by the dispatcher-fallback test). Non-fp32 bsx tiers are future work if a production 16-bit indexer materializes.

Test Coverage

New tests/unittest/_torch/attention/sparse/test_cute_dsl_bsx_topk_decode.py (143 cases): all 14 reg launch-table instances, fallback band-table routing (bucket membership, nearest-pow2 resolution, kill-switch), CUDA-graph capture/replay on both sides of the band table, admission fast path (high/low hit-rate, tie plateau at the admission threshold, overflow fallback, mixed admit/fallback ragged batch), MTP axes (next_n {2,3,4} x cr {1,4} x all tiers x 4 input classes, each with dual oracles: host-emulated torch.topk AND per-row differential against the in-tree kernel), MTP hardening (all-zero/garbage hints) and degenerate-row boundaries, direct/tp incl. trivial path, ragged varlen with poisoned tails (+1e30), quantized-tie plateaus, degenerate mixed batches, all-zero/OOR pre_idx, route-table + env-knob asserts, bf16 / next_n=2 fallback-to-in-tree checks. Registered in l0_b300.yml.
Existing test_cute_dsl_gvr_topk_decode.py: full suite green on this branch (671 passed / 144 skipped) — fp32/K2048/cr4 cells of that suite now exercise the bsx route.

Draft checklist (before ready-for-review)

Direct A/B perf table: adapted bsx tiers vs [None][perf] GVR top-K decode: enable R0 histogram-ladder admission by default #16457 kernel (uniform axis, full 865x11 grid) — ragged-mix axis still pending
CUDA-graph capture/replay test (both band-table sides, replay over rewritten inputs)
Rebase onto main ([None][perf] GVR top-K decode: enable R0 histogram-ladder admission by default #16457 merged; zero-conflict rebase 2026-07-28)

PR Checklist

PR title follows the required format
Description explains what and why
Test cases added and passing locally
DCO sign-off on all commits

🤖 Generated with Claude Code

…GVR decode op Adds the op43 bsx CuTe DSL kernel family as a guarded fast path inside trtllm::cute_dsl_gvr_topk_decode (op signature unchanged): - gvr_topk_decode_bsx_tp.py: throughput GVR tier (gvr_topk_tp port) with ragged-N lane masking, pre_idx clamping, in-kernel degenerate emit and the mandatory cluster exit rendezvous. - gvr_topk_decode_bsx_reg.py: register-resident tier (gvr_topk_reg port); ragged-N handled by extending the OOR-lane -FLT_MAX idiom at the single register load; packed u64 candidate pushes and aligned cluster barriers preserved from the op43 convergence rounds. - gvr_topk_decode_bsx_direct.py: short-row (npad <= 12288) exact radix tier with ragged-N key substitution and degenerate emit. - gvr_topk_decode_bsx_dispatch.py: route table transcribed from op42 gvr_bsx.cu; env knobs renamed GVR_BSX_* -> TRTLLM_BSX_*; v1 guard (fp32, next_n=1, cr=4, no order_row/counters, K in {512,1024,2048}, npad <= 262144, npad % 64 == 0) plus a hardware cluster-size cap check that falls back to the in-tree kernel instead of degrading silently. - CI-sized exactness test covering every reg launch-table instance once, direct/tp tiers, ragged rows with poisoned tails, quantized ties, degenerate rows, pre_idx hardening, route-table asserts and dispatcher fallback; registered in l0_b300.yml. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: longcheng-nv <243710427+longcheng-nv@users.noreply.github.com>

…gh-hit-rate rows) Port the in-tree R0 histogram-ladder ADMISSION concept (op#26, PR NVIDIA#16457) into the bsx throughput (tp) tier's pivot selection, closing the pr1 full-grid losses against the in-tree GVR kernel on real-capture rows at BS >= 16 while keeping the fused one-pass structure and every existing exactness invariant. Mechanism (measured on pr1 real-capture cells): the old 3-stage pivot pick targets ~3K sampled candidates inside a narrow [1.5K, 0.6kC] band. On real rows this either (a) picks a FAT rung (2-4x more P3 pushes and P4 candidates than needed - the v32/pro BS>=16 band), or (b) trusts a clustering-inflated sampled estimate and undershoots K, which fails the fused-pass reuse check and re-streams / secant-loops the full row (the flash_512k 1.6-1.8x losses; spatially clustered rows inflate a float4-sampled estimate up to 2.5x over the true count). Changes (gvr_topk_decode_bsx_tp.py only): * P2a stage-0 ADMISSION pick (R0 parity): accept the TIGHTEST ladder rung whose sampled-count confidence interval sits inside the [K, kC] acceptance window - the same "smallest exact count in [K, kC]" rule as the in-tree R0 admission, applied to the pre-pass estimates. The legacy 3-stage pick is unchanged as the fallback when no rung qualifies (cold-start / degenerate ladders take exactly the old path). * Clustering-aware sigma: sample_count now also tracks per-rung float4 OCCUPANCY, packed into the same per-thread accumulator (occ << 16 | cnt), giving the compound-Poisson sigma cnt/sqrt(occ) (equals the classic Poisson sigma on IID rows). Packing keeps registers, SMEM and the exchange at their pre-change sizes - an unpacked occ implementation measured a 14-25% whole-kernel regression (register spill in the streaming loops; A/B/bisect on v32_64k_L20). No field overflow: bsx guards npad <= 262144 => cluster-total cnt <= 8192. * Lower margin 2-sigma; K2048 uses 1.5-sigma (its [K, 4K] window is too narrow for 2-sigma to fire tight) backed by a RESCUE rung: the fused pass's second count column is now the next-fatter ladder rung instead of hmin (identical cost), so a pivot undershoot is caught with ONE collect re-stream instead of the multi-pass secant loop. Exactness machinery (accept window, P4 + tie tickets, plateau descent, ragged-N masking, degenerate emit, exit rendezvous) is untouched; admission never drops a top-K element because acceptance still requires the EXACT fused count in [K, kC]. Perf (nsys cold-L2 paired, same-rep A/B vs in-tree GVR kernel, real pr1 capture cells, B200; ratio = bsx/in-tree, lower is better): target cells before after flash_512k_L34 BS16 1.57 1.13 flash_512k_L34 BS64 1.72 1.11 flash_512k_L34 BS128 1.83 1.04 flash_512k_L34 BS256 1.54 0.73 flash_512k_L34 BS512 1.64 0.74 flash_512k_L34 BS1024 1.58 0.74 v32_32k_L50 BS16 1.40 1.21 v32_32k_L50 BS64 1.54 1.30 v32_32k_L50 BS256-1024 1.29-1.32 1.10-1.11 pro_128k_L54 BS16 1.27 1.19 pro_128k_L54 BS64 1.35 1.26 pro_128k_L54 BS128-1024 1.21-1.32 1.12-1.25 no-regression cells before after flash_512k_L36 BS1-128 0.55-0.94 0.56-0.92 flash_512k_L36 BS256-1024 0.67-0.68 0.66-0.67 v32_64k_L20 BS1-8 (reg) 0.63-0.65 0.63 v32_64k_L20 BS16-128 1.09-1.19 1.00-1.10 v32_64k_L20 BS256-1024 1.00-1.05 0.93-0.97 v32_32k_L04 BS16-1024 0.88-1.06 0.90-1.10 pro_256k_L30 BS16-1024 0.89-1.07 0.91-1.12 flash_16k_L26 BS1-1024 0.61-0.86 0.62-0.86 (reg/direct tiers untouched; the three +3-5% readings - pro_256k_L30 BS16/64, v32_32k_L04 BS64 - are inside the +-4.5% session noise floor measured on the fully untouched direct tier, e.g. flash_16k_L26 BS1 0.607 -> 0.634 with zero code change) The remaining v32_32k/pro_128k BS16-128 gap (1.19-1.30) is NOT the admission/pass-count mechanism: with admission the reuse check fires (one fused pass, admitted set slimmed 6678->2917 on v32_32k_L50) and a CS in {1,2,4} launch-shape sweep moves <= 7%. On those L2-resident shapes the in-tree kernel's whole-row-per-CTA structure is simply faster than the tp cluster split; closing it needs a tier-structure change, out of scope for this admission port. Exactness: bsx suite 56/56 (incl. new admission cases: hit-rate extremes on all three production shapes, tie plateau AT the admission threshold, count>kC overflow fallback, mixed admit/fallback ragged batch), in-tree gvr suite 671 passed / 144 skipped, 11/11 targeted screen (clustered/ties/zeros/oor/uniform/trivial), all pr1 A/B runs value-set-exact. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: longcheng-nv <243710427+longcheng-nv@users.noreply.github.com>

… target Fix round for the hint-ladder admission fast path (previous commit): the full-grid pr2-vs-pr1 A/B (9515 cases) showed the admission closed the deep tail (flash_512k_L34 1.83 -> 1.05) but regressed 396 cases by >5pp and flipped 47 former wins, netting the overall gm to a wash. Verified root cause (in-kernel decision probe + host-side P2a emulation on the regression rows) - TWO distinct failure modes, both in the PICK: 1. Fat admit: the [K, kC] acceptance window let the tightest CI-passing rung admit a candidate set 2.5x the legacy ~3K pivot-band target (pro_64k_L06: admitted 4134 vs legacy 1671; pro_32k_L28 4047 vs 1554; flash_128k_L10 1970 vs 788), inflating P3 pushes + P4 scan. 2. Undershoot admit: the K2048 1.5-sigma lower margin accepted rungs whose true count lands under K on clustered rows (v32_32k_L23 rung3 est 2720, sigma 374, TRUE count 1959 < 2048) - the rescue re-stream then costs more than the fat-but-valid legacy pick it displaced (22.6 -> 28.2us at BS64). The improved twin v32_32k_L50 has nearly identical rung3 stats (est 2784, sigma 382) with true count 2917: the pivot margin (+1.80 vs +1.92 sigma) is the only discriminant. Changes (pick logic only; fused pass, rescue and all exactness machinery are untouched from the admission commit): * Admission upper bound tightened from kC to the legacy pivot-band hi (0.6*kC), so admission only fires when the rung is genuinely tight. * K2048 lower margin raised 1.5 -> 1.85 sigma: rejects the L23-class undershoots (+1.80 sigma) while keeping the genuine tight admits (L50 +1.92 sigma). K512/K1024 stay at 2.0 sigma. * Stage 0b: the legacy band pick (min |est-tgt| in [1.5K, 0.6kC], the exact pr1 stage-1 rule) now OVERRIDES the admitted rung when it is strictly leaner and safe by its own clustering-aware 1.5-sigma lower CI; with no admission it is taken as-is (pr1 parity). * The pick's margin tests are now sqrt/div-free (squared comparisons): the pick is a thread-0 serial section between two CTA barriers, and the previous sqrt+fdiv chain measured 2-5% whole-kernel on L2-resident accept-path rows. Perf (nsys cold-L2 paired vs in-tree GVR kernel, real capture cells, B200; ratio = bsx/in-tree, lower is better; bar = pr1 ratio + 3% for the regression cells, absolute for the retention cells): regression cells pr1 pr2 fixed bar v32_32k_L23 BS16 1.202 1.521 1.198 <=1.238 PASS v32_32k_L23 BS64 1.334 1.662 1.325-1.352 <=1.374 PASS v32_32k_L23 BS1024 1.131 1.413 1.123 <=1.165 PASS pro_64k_L06 BS256 1.062 1.352 1.061 <=1.094 PASS pro_64k_L06 BS1024 1.099 1.414 1.121 <=1.132 PASS pro_32k_L28 BS1024 1.059 1.295 1.084 <=1.091 PASS flash_128k_L10 BS64 1.100 1.279 1.094 <=1.133 PASS pro_32k_L02 BS64 0.652 0.659 0.652 <=0.672 PASS retention cells pr1 pr2 fixed bar flash_512k_L34 BS128 1.831 1.052 1.020 <=1.08 PASS flash_512k_L34 BS256 1.537 0.730 0.725 <=0.80 PASS v32_32k_L50 BS64 1.537 1.290 1.267 <=1.35 PASS v32_32k_L50 BS16/1024 1.401/1.316 1.199/1.130 1.177/1.112 no-regression spots pr1 pr2 fixed v32_32k_L04 BS64 1.057 1.086 1.032-1.068 (3 reps) flash_16k_L26 BS1 (direct) 0.607 0.622 0.600 v32_64k_L20 BS8 (reg) 0.633 0.630 0.624 flash_512k_L36 BS64/256 0.944/0.674 0.898/0.656 0.883/0.655 pro_128k_L54 BS64/256 1.351/1.209 1.245/1.122 1.234/1.093 Known give-back: v32_64k_L20 BS16-128 (rung margin +1.53 sigma, below the new 1.85 K2048 cut) reverts from the pr2 admission pick to the legacy band pick; BS256-1024 keep the tight rung via stage 0b. Falsified along the way (kept out): stashing the [rescue, pivot) band in the candidate-buffer tail during the fused pass (per-element atomic, per-warp stripes, and warp-aggregated ballot variants) - the fused streaming loop cannot absorb ANY extra per-element code within the +-3% bar (measured +4-10% whole-kernel on accept-path rows for all three variants). Exactness: bsx suite 56/56 (admission cases included; acceptance-rule comment updated), in-tree gvr suite K2048 subset green, all A/B runs value-set-exact (exact=True on every measured cell x BS x arm). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: longcheng-nv <243710427+longcheng-nv@users.noreply.github.com>

…guard narrowed Extend the op43 bsx CuTe DSL tier family (direct/reg/tp) to the full in-tree run_one_row contract: next_n >= 1 (MTP speculative-decode rows) and compress_ratio in {1, 4}, and drop the next_n==1 / cr==4 conditions from the dispatcher guard. Semantics mirror the in-tree kernel line-for-line: - request-level hint sharing: pre_idx row = row // next_n (pre_idx and seq_lens are [num_rows // next_n, ...], validated by the guard); - cr==1 temporal hint shift (row % next_n) + 1, with shifted-OOR hints falling into the tiers' existing clamp hardening; - per-row N_eff = (seq_lens[req] - next_n + row % next_n + 1) // cr (already general in _row_n_eff). next_n / cr are ctor constexpr: the next_n==1 / cr==4 hot path traces identically to the v1 port (const_expr branches; no offset value is even computed on cr>1 builds). All compile / dispatch / runner caches key on (next_n, cr). Validation (umbriel-b200-074, torch nv26.05): - bsx suite extended with the MTP axis (next_n in {2,3,4} x cr in {1,4} x three tiers x {random-garbage, noised-hint, cold-start-zeros, ties}), each case checked against torch.topk host N_eff/offset simulation AND a differential in-tree arm (order_row-forced, per-row value-multiset equality): 140/140 passed. - full in-tree GVR suite (fp32 MTP/cr=1 cells now auto-route to bsx): 671 passed, 144 skipped. - next_n=1 perf anchors (ab_pr smoke, paired same-GPU back-to-back vs stashed baseline): bsxd/gvrpr ratio deltas within +/-3% on flash_16k_L26 BS1, v32_64k_L20 BS8, flash_512k_L34 BS128/256, v32_32k_L23 BS64, pro_256k_L30 BS256; all exact. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: longcheng-nv <243710427+longcheng-nv@users.noreply.github.com>

…-K interpolated tpush) and mask-hoisted streaming loops Ships the op43-pr6 campaign candidate (measured head 1a88ede77d) as a clean patch on the pr4 head: - stage 0c LEAN-PIVOT OVERRIDE (pick-only): on stage-0 CI-admission miss with npad in [16384, 262144], a strictly leaner ladder rung whose clustering-aware 1.5-sigma CI lands in [K, kC] becomes the pivot; undershoot costs one rescue re-stream (pr4 economics preserved). - per-K interpolated lean tpush (log2-count interpolation between pick and next-tighter rung; target 1.5K, 2.0K for K=512 sampling noise; npad <= 98304 gate) with original-pick rescue. - mask-hoisted streaming loops in count/max-below/fused-count-collect/ collect (unmasked main loop + masked vec-tail) - uniform micro-gain, independently exonerated of the pr5 regression by the C5 fingerprint ablation. - candidate budget pinned at the flat pr4 kC (8192 for K>=2048, else 6144); wide pr4 ladder quantiles pinned. Full-grid evidence (865 real decode cells x 11 BS, same-rep cold-L2 paired vs the in-tree kernel): gm 1.4170 (pr4 head 1.4120), win 88.0%, <0.909 cases 517 -> from 636. Exactness: 140-test production suite + adversarial + synthetic gates all PASS; measured module and this file are AST-identical modulo dead scaffolding removal. Signed-off-by: longcheng-nv <243710427+longcheng-nv@users.noreply.github.com>

…e kernel Routes the (npad, bs) buckets where the op43-pr6 full-grid verdict found at least one production layer >1.10x slower on every bsx tier than the in-tree kernel (L2-resident mid-N regime: the in-tree exact-count ladder admits a leaner candidate set and its row-slice cluster split keeps all CTAs busy through P4). With the table the operator is a strict Pareto improvement over the in-tree kernel: routed buckets run at parity, everything else keeps the bsx win (full-grid gm 1.40 vs the in-tree head, worst case capped at 1.10x). npad keys resolve by nearest power of two; TRTLLM_BSX_FALLBACK_BANDS=0 disables the table. Band data: 865 real decode cells x 11 BS same-rep cold-L2 nsys pairs, 2026-07-28. Tests: kernel-contract cases pin the table off via an autouse fixture (they exist to exercise the bsx tiers); the table itself is covered by test_bsx_fallback_band_table (bucket membership, nearest-pow2 resolution, neighbour non-routing, kill-switch), and CUDA-graph capture/replay by test_bsx_cuda_graph_capture_replay on both sides of the table (host-side dispatch bakes at capture; replay stays exact over in-place rewritten inputs). Suite: 143 passed; sibling test_cute_dsl_gvr_topk_decode suite: 671 passed / 144 skipped. Signed-off-by: longcheng-nv <243710427+longcheng-nv@users.noreply.github.com>

github-actions Bot assigned longcheng-nv Jul 26, 2026

longcheng-nv and others added 4 commits July 28, 2026 13:51

longcheng-nv force-pushed the perf/gvr-topk-bsx-cutedsl-tiers branch from 1f4bdff to 4f95451 Compare July 28, 2026 14:18

longcheng-nv added 2 commits July 28, 2026 14:33

longcheng-nv force-pushed the perf/gvr-topk-bsx-cutedsl-tiers branch from 4f95451 to 78df9cb Compare July 28, 2026 14:34

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[None][perf] Add BSX multi-tier CuTe DSL top-k decode kernels (stacked on #16457) - #16877

[None][perf] Add BSX multi-tier CuTe DSL top-k decode kernels (stacked on #16457)#16877
longcheng-nv wants to merge 6 commits into
NVIDIA:mainfrom
longcheng-nv:perf/gvr-topk-bsx-cutedsl-tiers

longcheng-nv commented Jul 26, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

longcheng-nv commented Jul 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Evidence (local, B200; text tables only — no profiler artifacts)

Test Coverage

Draft checklist (before ready-for-review)

PR Checklist

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

longcheng-nv commented Jul 26, 2026 •

edited

Loading