perf(ArcKernels/ArcAttention/ArcQuant): session-10 measured default flips — GEMV_WIDE ON, QK_FUSED_COHORT ON, CUBLAS_MIN_M 512→64 (stacked: B=256 575→758 tok/s, b=1 40.8→43.8)

[heydryft]

Parent systems: ArcKernels, ArcQuant, ArcInfer/ArcAttention. Session-10 box run, box arc-s10-ledger (H200, exclusive, PTX-JIT gate passed, provenance asserted per leg), candidate 4d03b9e25 = release/openrouter-ready after the six wave-1 lanes. One binary, env-toggle A/Bs, seeded canaries (greedy banned, D4), produced-token rates with a >=256/seq floor.

The three flips, each measured this session

commit	flag	evidence
8c81489	ARC_FP8_GEMV_WIDE default ON	b=1 43.80 vs 40.76 tok/s (+7.5%), engagement path=gemv_wide, canary coherent (f32 re-association bound). The ledger's ~1.6x prediction did NOT materialise end-to-end.
0ef138b	ARC_QK_FUSED_COHORT default ON	ENGAGED at B=256 (default arm DECLINED to eager), seeded canaries byte-identical, 599.35 vs 575.12 tok/s. +4.2% is inside the measured ~±5% inter-leg variance — the flip stands on engaged + bit-identical + not-slower, the exact bar the capability entry set.
e570d30	ARC_FP8_CUBLAS_MIN_M 512 → 64	The three-arm sweep this constant's own doc demanded (fp8_threshold_sweep.sh, M∈{8,64,256}, >=1000-token floor). Clean rows: M=64 cublaslt 552.4 vs wmma 328.2 vs tiled 223.1; M=256 715.9 vs 620.4 vs 285.8. First clean crossover: 64. M=8 rows VOID on the floor ⇒ 5..63 stays native (a void row makes no claim). M<=4 stays fp8_gemv_warp.

Stacked verification (all three ON together, same binary, env-equivalent to these defaults)

• B=256 decode aggregate 757.54 tok/s (5 reps 748.8–761.4, effective_B=256 verified, 0/1280 errors), per-user p50 3.35 tok/s, TTFT p50 4.87 s, $1.80/Mtok @ $4.92/hr.
• b=1 steady state 43.77 tok/s (5 runs, spread 0.04).
• Ladder context: pre-wave (5516036) B=256 was 355.4; candidate baseline 575.1; stacked 757.5 — 2.13× in one session, with the wave-1 merges and these flips both contributing.

Also decided by the same ladder (no code change here)

• ARC_NO_FP8_WMMA=1 control: tiled is 46% slower at B=256 with an identical canary — the never-run WMMA kernel is confirmed as the correct default arm (its "silently wrong GEMM" fear is REFUTED on hardware).
• Ragged pair, ARC_SAMPLE_ON_DEVICE (sample share 7.4%→57% of step — prediction inverted, measured loudly), TCFRAG (init OOM beside a loaded V4), capture/device-loop (verify_failed=true at replay perf(stage2): absorbed-MLA V4 decode + GPU radix top-k/top-p sampler #10 — the +1-step divergence is still open): all measured FAIL/unengaged, all stay OFF, evidence in the session results.
• ARC_V4_FP8_KV: canaries byte-identical and not slower, but the lane has NO engagement instrument at all (silent-success class) — no flip on unproven engagement; needs an engagement line first.

Verification

• Box: every number above from arc-s10-ledger, exclusivity-bracketed (compute-apps == server pid before AND after each leg), running-revision assertion per server, PTX-JIT gate (kernel returned 42 through a JIT-only path) at bootstrap.
• Local: cargo check -p mistralrs-quant -p mistralrs-core clean; polarity tests updated both ways (only_literal_zero_disables, only_literal_zero_disables_the_cohort, the_batched_cohort_is_on_by_default_since_the_s10_measurement); capability_reachability 11/11 with both registry entries updated (QK cohort entry text now records the measurement; new ArcKernels entry for gemv_wide).

🤖 Generated with Claude Code

https://claude.ai/code/session_01WK8VgocBUrr5djjE1ZfCNf

[github-actions]

Code Metrics Report

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Language              Files        Lines         Code     Comments       Blanks
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 C Header                  5          305          210           52           43
 CSS                       2         1181         1036           34          111
 CUDA                     81        29475        20161         6275         3039
 Dockerfile                1           39           22            8            9
 JavaScript               16         3546         2676          482          388
 Jinja2                    7          694          656            5           33
 JSON                     75         4896         4893            0            3
 Makefile                  1            6            5            0            1
 Metal Shading Lan|       33        12224         9431         1142         1651
 PowerShell                1          300          227           30           43
 Python                  147        15371        12677          824         1870
 Shell                    42        10304         6858         2769          677
 Plain Text                4         3801            0         2479         1322
 TOML                     33         1498         1294           54          150
 YAML                      3           25           23            2            0
─────────────────────────────────────────────────────────────────────────────────
 HTML                      4         2687         2604           43           40
 |- CSS                    2          543          479           37           27
 |- JavaScript             1         1233         1215           12            6
 (Total)                             4463         4298           92           73
─────────────────────────────────────────────────────────────────────────────────
 Jupyter Notebooks         4          122           83           23           16
 |- Markdown               1           60           30           22            8
 |- Python                 1          122          113            1            8
 (Total)                              304          226           46           32
─────────────────────────────────────────────────────────────────────────────────
 Markdown                213        48932            0        38081        10851
 |- BASH                  72         1655         1203          331          121
 |- C                      4           19           19            0            0
 |- CUDA                   2           84           56           16           12
 |- JSON                  19          779          779            0            0
 |- PowerShell             1            1            1            0            0
 |- Python                23         1008          787          113          108
 |- Rust                  68         2063         1727           78          258
 |- TOML                   6          207          164            0           43
 |- YAML                   5           41           36            5            0
 (Total)                            54789         4772        38624        11393
─────────────────────────────────────────────────────────────────────────────────
 Rust                    681       340988       293252        18096        29640
 |- Markdown             504        32562          471        28063         4028
 (Total)                           373550       293723        46159        33668
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Total                  1353       516771       363188        99077        54506
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

[heydryft]

Fourth flip added: 450208d — ARC_V4_FP8_KV default ON. Engagement was proven by nsys kernel capture (arc_kv_fp8_quantize_kernel x16,297 + arc_kv_fp8_dequantize_kernel x32,594 in one b=1 run) since the lane has no engagement log line (silent-success class; follow-up named). 4-gate stacked re-verification, one binary @4d03b9e25: B=256 744.7 tok/s (2 reps, within the ±5% band of the 3-gate 757.5), b=1 44.68 tok/s (session best), canaries IDENTICAL through the full stack, 0 errors. Ladder totals vs pre-wave 5516036: B=256 355.4 → 744.7–757.5; b=1 39.3 → 44.7.