[gate] A-fragment addressing: 8192 slots, all placed per the GEMM's documented layout
[state] 2520, 27.94, 42, Not Active   shape s27 TILE_M=128

  gate_proj k=5120 n=17408 M=2048  [2475, 64.66, 58, Active]
    A side   shipped   0.094  fused   0.040  paired  2.37x [2.32-2.48]
    full     shipped   1.178  fused   1.152  paired 1.040x [0.997-1.068]
    vs legB    1.711  ->  shipped 1.452x   FUSED 1.485x     [err vs fp32: shipped 1.343e-02 fused 1.343e-02]

  gate_proj k=5120 n=17408 M=8192  [1965, 445.20, 63, Active]
    A side   shipped   0.552  fused   0.155  paired  3.56x [3.52-3.59]
    full     shipped   5.216  fused   4.432  paired 1.177x [1.098-1.224]
    vs legB    6.234  ->  shipped 1.195x   FUSED 1.407x     [err vs fp32: shipped 1.350e-02 fused 1.350e-02]

  gate_proj k=5120 n=17408 M=16384  [2415, 446.08, 65, Active]
    A side   shipped   1.209  fused   0.307  paired  3.94x [3.90-4.06]
    full     shipped  10.161  fused   8.921  paired 1.141x [1.098-1.190]
    vs legB   12.758  ->  shipped 1.256x   FUSED 1.430x     [err vs fp32: shipped 1.347e-02 fused 1.347e-02]

  out_proj k=6144 n=5120 M=2048  [2415, 446.08, 65, Active]
    A side   shipped   0.106  fused   0.041  paired  2.60x [2.52-2.63]
    full     shipped   0.487  fused   0.420  paired 1.154x [1.140-1.200]
    vs legB    0.619  ->  shipped 1.271x   FUSED 1.474x     [err vs fp32: shipped 1.363e-02 fused 1.364e-02]

  out_proj k=6144 n=5120 M=8192  [2400, 391.96, 63, Active]
    A side   shipped   0.682  fused   0.174  paired  3.91x [3.90-3.94]
    full     shipped   2.255  fused   1.691  paired 1.322x [1.305-1.396]
    vs legB    2.111  ->  shipped 0.936x   FUSED 1.248x     [err vs fp32: shipped 1.369e-02 fused 1.369e-02]

  out_proj k=6144 n=5120 M=16384  [2580, 426.28, 66, Active]
    A side   shipped   1.439  fused   0.334  paired  4.31x [4.26-4.35]
    full     shipped   4.646  fused   3.236  paired 1.386x [1.356-1.448]
    vs legB    4.140  ->  shipped 0.891x   FUSED 1.279x     [err vs fp32: shipped 1.366e-02 fused 1.366e-02]

  down_proj k=17408 n=5120 M=2048  [2520, 388.24, 61, Active]
    A side   shipped   0.459  fused   0.120  paired  3.83x [3.80-3.85]
    full     shipped   1.447  fused   1.122  paired 1.314x [1.285-1.338]
    vs legB    1.587  ->  shipped 1.097x   FUSED 1.415x     [err vs fp32: shipped 1.392e-02 fused 1.392e-02]

  down_proj k=17408 n=5120 M=8192  [2460, 443.48, 68, Active]
    A side   shipped   2.031  fused   0.462  paired  4.39x [4.37-4.40]
    full     shipped   5.931  fused   4.211  paired 1.408x [1.355-1.456]
    vs legB    5.509  ->  shipped 0.929x   FUSED 1.308x     [err vs fp32: shipped 1.395e-02 fused 1.395e-02]

  down_proj k=17408 n=5120 M=16384  [2400, 444.60, 66, Active]
    A side   shipped   4.064  fused   0.920  paired  4.41x [4.41-4.42]
    full     shipped  11.671  fused   8.528  paired 1.369x [1.345-1.390]
    vs legB   10.885  ->  shipped 0.933x   FUSED 1.276x     [err vs fp32: shipped 1.391e-02 fused 1.391e-02]

  in_proj_qkv k=5120 n=10240 M=2048  [2295, 398.42, 63, Active]
    A side   shipped   0.095  fused   0.040  paired  2.37x [2.33-2.44]
    full     shipped   0.748  fused   0.720  paired 1.055x [1.000-1.076]
    vs legB    1.025  ->  shipped 1.370x   FUSED 1.424x     [err vs fp32: shipped 1.413e-02 fused 1.414e-02]

  in_proj_qkv k=5120 n=10240 M=8192  [2415, 436.39, 66, Active]
    A side   shipped   0.552  fused   0.156  paired  3.54x [3.51-3.59]
    full     shipped   3.363  fused   2.726  paired 1.229x [1.149-1.281]
    vs legB    3.568  ->  shipped 1.061x   FUSED 1.309x     [err vs fp32: shipped 1.415e-02 fused 1.415e-02]

  in_proj_qkv k=5120 n=10240 M=16384  [2355, 439.32, 68, Active]
    A side   shipped   1.210  fused   0.302  paired  4.00x [3.96-4.06]
    full     shipped   6.608  fused   5.285  paired 1.250x [1.176-1.265]
    vs legB    7.124  ->  shipped 1.078x   FUSED 1.348x     [err vs fp32: shipped 1.412e-02 fused 1.412e-02]
[gate] A-fragment addressing: 8192 slots, all placed per the GEMM's documented layout
[state] 210, 28.43, 45, Not Active   shape s27 TILE_M=128
[state] 210, 28.43, 45, Not Active   grouped, TILE_M=128
  gate|up G=2 k=5120 M=2048  separate   0.072  shared-read   0.063  paired 1.154x [1.114-1.158]
  gate|up G=2 k=5120 M=8192  separate   0.302  shared-read   0.248  paired 1.216x [1.207-1.219]
  gate|up G=2 k=5120 M=16384  separate   0.593  shared-read   0.489  paired 1.212x [1.210-1.214]
  qkv|z G=2 k=5120 M=2048  separate   0.067  shared-read   0.058  paired 1.154x [1.148-1.156]
  qkv|z G=2 k=5120 M=8192  separate   0.291  shared-read   0.233  paired 1.252x [1.246-1.254]
  qkv|z G=2 k=5120 M=16384  separate   0.572  shared-read   0.462  paired 1.241x [1.232-1.242]
