[state] 210, 34.44, 38, Not Active

  gate_proj k=5120 n=17408 M=2048  legB median 1.628 ms (224 TFLOPS, burst-inflated)  [2295, 443.66, 57, Active]
    s27  128x128x128 1x4w st2 sw8  gemm   0.887 full   1.125  paired legB/full = 1.462x [1.413-1.527]
    s12  128x128x64  1x4w st3 sw8  gemm   0.951 full   1.158  paired legB/full = 1.391x [1.318-1.493]
    s10  128x128x64  1x4w st3 sw1  gemm   0.967 full   1.199  paired legB/full = 1.368x [1.339-1.427]
    s11  128x256x64  1x8w st3 sw4  gemm   1.088 full   1.311  paired legB/full = 1.248x [1.184-1.315]
    s24  128x128x128 2x4w st2 sw8  gemm   1.076 full   1.309  paired legB/full = 1.242x [1.229-1.308]
    s7   128x128x128 2x4w st2 sw1  gemm   1.095 full   1.315  paired legB/full = 1.237x [1.127-1.281]
    s26  128x128x128 2x4w st2 sw4  gemm   1.087 full   1.318  paired legB/full = 1.234x [1.204-1.308]
    s8   128x256x64  1x8w st3 sw1  gemm   1.105 full   1.321  paired legB/full = 1.225x [1.205-1.266]
    s21  128x128x64  2x4w st3 sw8  gemm   1.106 full   1.331  paired legB/full = 1.215x [1.172-1.253]
    s20  128x128x64  2x4w st3 sw4  gemm   1.126 full   1.359  paired legB/full = 1.214x [1.177-1.264]
    s23  256x128x64  2x4w st3 sw8  gemm   1.165 full   1.378  paired legB/full = 1.189x [1.128-1.246]
    s22  256x128x64  2x4w st3 sw4  gemm   1.175 full   1.375  paired legB/full = 1.178x [1.155-1.217]
    s0   128x128x64  2x4w st3 sw1  gemm   1.114 full   1.360  paired legB/full = 1.171x [1.037-1.225]
    s25  256x128x64  2x4w st3 sw16 gemm   1.162 full   1.414  paired legB/full = 1.159x [1.124-1.237]
    s1   256x128x64  2x4w st3 sw1  gemm   1.189 full   1.431  paired legB/full = 1.128x [1.062-1.182]

  gate_proj k=5120 n=17408 M=8192  legB median 6.345 ms (230 TFLOPS, burst-inflated)  [2745, 445.85, 57, Active]
    s27  128x128x128 1x4w st2 sw8  gemm   3.519 full   4.740  paired legB/full = 1.373x [1.226-1.472]
    s10  128x128x64  1x4w st3 sw1  gemm   3.645 full   4.973  paired legB/full = 1.318x [1.189-1.398]
    s12  128x128x64  1x4w st3 sw8  gemm   3.635 full   4.996  paired legB/full = 1.267x [1.178-1.424]
    s11  128x256x64  1x8w st3 sw4  gemm   4.002 full   5.120  paired legB/full = 1.210x [1.132-1.370]
    s8   128x256x64  1x8w st3 sw1  gemm   4.071 full   5.101  paired legB/full = 1.195x [1.135-1.380]
    s22  256x128x64  2x4w st3 sw4  gemm   4.457 full   5.371  paired legB/full = 1.188x [1.100-1.295]
    s0   128x128x64  2x4w st3 sw1  gemm   5.353 full   5.766  paired legB/full = 1.147x [1.009-1.205]
    s7   128x128x128 2x4w st2 sw1  gemm   4.395 full   5.692  paired legB/full = 1.144x [1.038-1.226]
    s24  128x128x128 2x4w st2 sw8  gemm   4.490 full   5.698  paired legB/full = 1.141x [1.038-1.197]
    s23  256x128x64  2x4w st3 sw8  gemm   4.088 full   5.316  paired legB/full = 1.125x [1.074-1.313]
    s21  128x128x64  2x4w st3 sw8  gemm   4.690 full   5.956  paired legB/full = 1.121x [0.991-1.170]
    s26  128x128x128 2x4w st2 sw4  gemm   4.416 full   5.708  paired legB/full = 1.121x [1.029-1.222]
    s1   256x128x64  2x4w st3 sw1  gemm   4.304 full   5.315  paired legB/full = 1.115x [1.040-1.315]
    s25  256x128x64  2x4w st3 sw16 gemm   4.328 full   5.343  paired legB/full = 1.115x [1.052-1.309]
    s20  128x128x64  2x4w st3 sw4  gemm   4.504 full   5.834  paired legB/full = 1.085x [1.028-1.196]

  out_proj k=6144 n=5120 M=2048  legB median 0.578 ms (223 TFLOPS, burst-inflated)  [2640, 428.80, 60, Active]
    s27  128x128x128 1x4w st2 sw8  gemm   0.308 full   0.448  paired legB/full = 1.289x [1.279-1.295]
    s12  128x128x64  1x4w st3 sw8  gemm   0.359 full   0.500  paired legB/full = 1.156x [1.148-1.167]
    s26  128x128x128 2x4w st2 sw4  gemm   0.382 full   0.523  paired legB/full = 1.105x [1.075-1.110]
    s24  128x128x128 2x4w st2 sw8  gemm   0.384 full   0.526  paired legB/full = 1.099x [1.071-1.106]
    s10  128x128x64  1x4w st3 sw1  gemm   0.391 full   0.528  paired legB/full = 1.094x [1.076-1.107]
    s20  128x128x64  2x4w st3 sw4  gemm   0.390 full   0.530  paired legB/full = 1.092x [1.085-1.094]
    s0   128x128x64  2x4w st3 sw1  gemm   0.390 full   0.531  paired legB/full = 1.090x [1.004-1.092]
    s21  128x128x64  2x4w st3 sw8  gemm   0.389 full   0.532  paired legB/full = 1.088x [1.080-1.096]
    s7   128x128x128 2x4w st2 sw1  gemm   0.392 full   0.537  paired legB/full = 1.077x [1.030-1.107]
    s11  128x256x64  1x8w st3 sw4  gemm   0.426 full   0.561  paired legB/full = 1.030x [1.005-1.045]
    s8   128x256x64  1x8w st3 sw1  gemm   0.428 full   0.568  paired legB/full = 1.019x [0.988-1.050]
    s25  256x128x64  2x4w st3 sw16 gemm   0.439 full   0.577  paired legB/full = 1.001x [0.969-1.020]
    s22  256x128x64  2x4w st3 sw4  gemm   0.438 full   0.583  paired legB/full = 0.990x [0.962-1.002]
    s23  256x128x64  2x4w st3 sw8  gemm   0.442 full   0.585  paired legB/full = 0.989x [0.966-1.002]
    s1   256x128x64  2x4w st3 sw1  gemm   0.454 full   0.595  paired legB/full = 0.974x [0.890-1.017]

  out_proj k=6144 n=5120 M=8192  legB median 1.979 ms (260 TFLOPS, burst-inflated)  [2760, 443.73, 57, Active]
    s27  128x128x128 1x4w st2 sw8  gemm   1.202 full   2.086  paired legB/full = 0.959x [0.938-1.017]
    s10  128x128x64  1x4w st3 sw1  gemm   1.259 full   2.120  paired legB/full = 0.934x [0.916-1.014]
    s12  128x128x64  1x4w st3 sw8  gemm   1.252 full   2.125  paired legB/full = 0.925x [0.880-1.016]
    s11  128x256x64  1x8w st3 sw4  gemm   1.402 full   2.234  paired legB/full = 0.879x [0.844-0.966]
    s8   128x256x64  1x8w st3 sw1  gemm   1.426 full   2.264  paired legB/full = 0.875x [0.860-0.947]
    s23  256x128x64  2x4w st3 sw8  gemm   1.474 full   2.352  paired legB/full = 0.862x [0.825-0.943]
    s22  256x128x64  2x4w st3 sw4  gemm   1.515 full   2.333  paired legB/full = 0.861x [0.842-0.918]
    s7   128x128x128 2x4w st2 sw1  gemm   1.499 full   2.367  paired legB/full = 0.840x [0.824-0.896]
    s24  128x128x128 2x4w st2 sw8  gemm   1.499 full   2.359  paired legB/full = 0.840x [0.829-0.900]
    s25  256x128x64  2x4w st3 sw16 gemm   1.502 full   2.364  paired legB/full = 0.839x [0.806-0.908]
    s26  128x128x128 2x4w st2 sw4  gemm   1.501 full   2.390  paired legB/full = 0.837x [0.822-0.898]
    s1   256x128x64  2x4w st3 sw1  gemm   1.575 full   2.376  paired legB/full = 0.833x [0.819-0.903]
    s21  128x128x64  2x4w st3 sw8  gemm   1.549 full   2.415  paired legB/full = 0.827x [0.810-0.879]
    s20  128x128x64  2x4w st3 sw4  gemm   1.540 full   2.450  paired legB/full = 0.813x [0.788-0.887]
    s0   128x128x64  2x4w st3 sw1  gemm   1.688 full   2.584  paired legB/full = 0.795x [0.760-0.828]
