Kai: the speed table from stage a6 beside Laya on dbc, Metal and CPU.

Claude-Session: https://claude.ai/code/session_018Gt5TvrE2rxXuZikWSauv3

kai/tables/capability.tex

@@ -1,4 +1,4 @@
-% capability.tex: written by `bench tex` (hanzoai/decision); Kai runs/a6, weights a211cc70, cpu f32; harness a1a4a73; benchmarks e418767
+% capability.tex: written by `bench tex` (hanzoai/decision); Kai runs/a6, weights a211cc70, cpu f32; harness a1a4a73; benchmarks efd4d91
 % Kai: joint rows are its refinement over three passes (projected: its own selection and projection); Laya's joint rows are its typed-decisions checkpoint; media is Laya as routed. Rejected: questions a backend refuses, over all asked.
 \begin{tabular}{@{}lccc@{}}
 \toprule
@@ -41,5 +41,5 @@ Mean set size & 2.498 & 6.801 & 1.304 \\
 Refresh: questions asked / full & 6 / 101 & 100 / 100 & 100 / 100 \\
 Runs offline & yes & yes & no \\
 \bottomrule
-\multicolumn{4}{@{}l@{}}{\footnotesize\begin{tabular}[t]{@{}l@{}}Kai runs/a6, weights a211cc70, cpu f32 \\ harness a1a4a73, benchmarks e418767\end{tabular}} \\
+\multicolumn{4}{@{}l@{}}{\footnotesize\begin{tabular}[t]{@{}l@{}}Kai runs/a6, weights a211cc70, cpu f32 \\ harness a1a4a73, benchmarks efd4d91\end{tabular}} \\
 \end{tabular}

kai/tables/harness.tex

@@ -1,4 +1,4 @@
-% harness.tex: written by `bench tex` (hanzoai/decision); Kai runs/a6, weights a211cc70, cpu f32; harness a1a4a73; benchmarks e418767
+% harness.tex: written by `bench tex` (hanzoai/decision); Kai runs/a6, weights a211cc70, cpu f32; harness a1a4a73; benchmarks efd4d91
 % Laya: the checkpoint its router picks per case; typed decisions on laya-agent (kai-1-agent), the typed-decisions checkpoint Laya ships and its router does not pick.
 \begin{tabular}{@{}l@{\quad}ccc@{\quad}ccc@{\quad}ccc@{}}
 \toprule
@@ -20,5 +20,5 @@ MASSIVE (51 languages, macro) & 0.910 & 0.382 & 0.890 & 0.136 & 0.987 & 0.161 &
 \midrule
 Macro, 11 suites & 0.850 & 0.702 & 0.770 & 0.241 & 0.444 & 0.359 & 0.099 & 0.164 & 0.145 \\
 \bottomrule
-\multicolumn{10}{@{}l@{}}{\footnotesize\begin{tabular}[t]{@{}l@{}}Kai runs/a6, weights a211cc70, cpu f32 \\ harness a1a4a73, benchmarks e418767\end{tabular}} \\
+\multicolumn{10}{@{}l@{}}{\footnotesize\begin{tabular}[t]{@{}l@{}}Kai runs/a6, weights a211cc70, cpu f32 \\ harness a1a4a73, benchmarks efd4d91\end{tabular}} \\
 \end{tabular}

kai/tables/languages.tex

@@ -1,4 +1,4 @@
-% languages.tex: written by `bench tex` (hanzoai/decision); Kai runs/a6, weights a211cc70, cpu f32; harness a1a4a73; benchmarks e418767
+% languages.tex: written by `bench tex` (hanzoai/decision); Kai runs/a6, weights a211cc70, cpu f32; harness a1a4a73; benchmarks efd4d91
 % MASSIVE intent accuracy by language, 20 options, 100 utterances each; Laya as its router picks.
 \begin{tabular}{@{}lccc@{\qquad}lccc@{}}
 \toprule
@@ -31,5 +31,5 @@ ka & 0.830 & 0.150 & 0.880 & zh-CN & 0.950 & 0.650 & 0.940 \\
 km & 0.850 & 0.200 & 0.810 & zh-TW & 0.960 & 0.610 & 0.930 \\
 kn & 0.900 & 0.300 & 0.870 &  & & & \\
 \bottomrule
-\multicolumn{8}{@{}l@{}}{\footnotesize\begin{tabular}[t]{@{}l@{}}Kai runs/a6, weights a211cc70, cpu f32 \\ harness a1a4a73, benchmarks e418767\end{tabular}} \\
+\multicolumn{8}{@{}l@{}}{\footnotesize\begin{tabular}[t]{@{}l@{}}Kai runs/a6, weights a211cc70, cpu f32 \\ harness a1a4a73, benchmarks efd4d91\end{tabular}} \\
 \end{tabular}

kai/tables/speed.tex

@@ -1,4 +1,4 @@
-% speed.tex: written by `bench tex` (hanzoai/decision); Kai runs/a, encoder 768 wide, 22 layers; timed F32 on Metal and on the CPU, Kai's CPU encoder projections Q8_0; harness a1a4a73; benchmarks e418767
+% speed.tex: written by `bench tex` (hanzoai/decision); Kai runs/a6, encoder 768 wide, 22 layers; timed F32 on Metal and on the CPU, Kai's CPU encoder projections Q8_0; harness a1a4a73; benchmarks efd4d91
 % Kai and Laya on this runtime (hanzo-ml), same device and dtype (F32; on the CPU Kai's encoder projections are Q8_0, as Decider::load serves them), their calls interleaved; the last row names each checkpoint and its encoder. Jev is one network run, the same cells in both blocks.
 \begin{tabular}{@{}r@{\quad}cc@{\quad}cc@{\quad}cc@{\quad}ccc@{}}
 \toprule
@@ -6,20 +6,20 @@
 $N$ & p50 & p95 & p50 & p95 & p50 & p95 & Kai & Laya & Jev \\
 \midrule
 \multicolumn{10}{@{}l}{\emph{Metal}} \\
-1 & 31.9 & 46.0 & 28.9 & 133 & 133 & 546 & 27.5 & 26.5 & 5.72 \\
-10 & 129 & 205 & 214 & 314 & 155 & 221 & 74.5 & 44.5 & 62.9 \\
-50 & 555 & 750 & 1{,}097 & 1{,}246 & 202 & 275 & 85.9 & 44.3 & 233 \\
-100 & 1{,}214 & 1{,}403 & 2{,}362 & 2{,}587 & 343 & 556 & 80.4 & 42.4 & 267 \\
-500 & 5{,}235 & 6{,}079 & 10{,}133 & 10{,}828 & 515 & 2{,}193 & 93.0 & 49.4 & 726 \\
-1{,}000 & 10{,}239 & 11{,}433 & 19{,}616 & 20{,}714 & 602 & 1{,}077 & 94.7 & 50.7 & 1{,}489 \\
+1 & 17.8 & 24.7 & 15.6 & 16.5 & 133 & 546 & 49.3 & 66.9 & 5.72 \\
+10 & 45.6 & 84.8 & 122 & 175 & 155 & 221 & 206 & 78.3 & 62.9 \\
+50 & 135 & 227 & 604 & 617 & 202 & 275 & 315 & 83.4 & 233 \\
+100 & 351 & 376 & 1{,}263 & 1{,}324 & 343 & 556 & 293 & 79.4 & 267 \\
+500 & 1{,}264 & 1{,}394 & 5{,}370 & 5{,}756 & 515 & 2{,}193 & 395 & 93.1 & 726 \\
+1{,}000 & 2{,}440 & 2{,}502 & 10{,}315 & 10{,}920 & 602 & 1{,}077 & 411 & 95.8 & 1{,}489 \\
 \midrule
 \multicolumn{10}{@{}l}{\emph{CPU}} \\
-1 & 162 & 179 & 387 & 409 & 133 & 546 & 6.92 & 2.96 & 5.72 \\
-10 & 1{,}024 & 1{,}409 & 4{,}337 & 6{,}282 & 155 & 221 & 9.04 & 2.01 & 62.9 \\
-50 & 4{,}778 & 4{,}862 & 20{,}607 & 21{,}073 & 202 & 275 & 10.6 & 2.44 & 233 \\
-100 & 8{,}974 & 9{,}379 & 39{,}938 & 42{,}672 & 343 & 556 & 11.2 & 2.46 & 267 \\
+1 & 139 & 147 & 214 & 219 & 133 & 546 & 8.32 & 4.99 & 5.72 \\
+10 & 365 & 399 & 1{,}622 & 2{,}025 & 155 & 221 & 28.5 & 5.83 & 62.9 \\
+50 & 920 & 980 & 7{,}492 & 7{,}818 & 202 & 275 & 53.5 & 6.58 & 233 \\
+100 & 1{,}720 & 1{,}767 & 15{,}391 & 16{,}170 & 343 & 556 & 57.6 & 6.40 & 267 \\
 500 & 43{,}152 & 44{,}820 & 178{,}544 & 182{,}739 & 515 & 2{,}193 & 11.7 & 2.80 & 726 \\
 1{,}000 & 80{,}165 & 80{,}165 & 329{,}783 & 329{,}783 & 602 & 1{,}077 & 12.5 & 3.03 & 1{,}489 \\
 \bottomrule
-\multicolumn{10}{@{}l@{}}{\footnotesize\begin{tabular}[t]{@{}l@{}}Kai runs/a, encoder 768 wide, 22 layers \\ harness a1a4a73, benchmarks e418767 \\ timed F32 on Metal and on the CPU, Kai's CPU encoder projections Q8\_0 \\ Laya laya-multilingual: hanzoai/kai-1-multilingual@3119843b, encoder 768 wide, 22 layers \\ timed calls per N: Metal 10; CPU 3, 1 at N = 1,000 \\ dgx (NVIDIA GB10, CUDA)\end{tabular}} \\
+\multicolumn{10}{@{}l@{}}{\footnotesize\begin{tabular}[t]{@{}l@{}}Kai runs/a6, encoder 768 wide, 22 layers \\ harness a1a4a73, benchmarks efd4d91 \\ timed F32 on Metal and on the CPU, Kai's CPU encoder projections Q8\_0 \\ Laya laya-multilingual: hanzoai/kai-1-multilingual@3119843b, encoder 768 wide, 22 layers \\ timed calls per N: Metal 10; CPU 3, 1 at N = 1,000 \\ dbc (Apple M4 Max, 128 GB)\end{tabular}} \\
 \end{tabular}