# table from NVIDIA GeForce RTX 4090 (/models/Qwen3.8-27B-exl3-4.0bpw)
# clocks: start {'sm_mhz': 210, 'max_sm_mhz': 3165, 'util_pct': 0, 'suspect_pin': False}  end {'sm_mhz': 2685, 'max_sm_mhz': 3165, 'util_pct': 100, 'suspect_pin': False}
# sole tenancy for the whole run: True
# decode-class rows reachable at max_batch=32 K=7:
#   1..256, 60 values, 28 above 32

## 17408x5120xK4  (model.language_model.layers.1.mlp.down_proj; reconstruct threshold 280 rows)
  select_rows shipped(1,16,32)   -> K32N128    (K16N128=0.2642  K16N512=0.3125  K32N128=0.2406  K32N256=0.2877)
  select_rows corners            -> K16N128    (K16N128=0.7578  K16N512=1.1419  K32N128=0.7711  K32N256=1.0181)
  select_rows reachable-all      -> K16N128    (K16N128=5.5007  K16N512=9.3235  K32N128=6.2189  K32N256=8.1347)
      M    K16N128    K16N512    K32N128    K32N256   K16N128 vs K32N128
      1     0.0799     0.0829     0.0778     0.0778    0.974x
      2     0.0799     0.0829     0.0776     0.0771    0.972x
      3     0.0799     0.0840     0.0778     0.0778    0.974x
      4     0.0799     0.0840     0.0778     0.0778    0.974x
      6     0.0809     0.0860     0.0787     0.0789    0.972x
      8     0.0809     0.0881     0.0789     0.0789    0.975x
     12     0.0820     0.0949     0.0799     0.0809    0.973x
     16     0.0819     0.1014     0.0799     0.0827    0.975x
     24     0.1014     0.1147     0.0821     0.1239    0.810x
     32     0.1024     0.1282     0.0829     0.1272    0.810x
     40     0.1126     0.1946     0.1434     0.1669    1.273x
     48     0.1137     0.2069     0.1434     0.1700    1.261x
     56     0.1290     0.2181     0.1455     0.2089    1.127x
     64     0.1311     0.2314     0.1464     0.2130    1.117x
     72     0.1471     0.2959     0.2069     0.2519    1.406x
     80     0.1485     0.3082     0.2072     0.2560    1.395x
     88     0.1700     0.3205     0.2095     0.2959    1.233x
     96     0.1720     0.3331     0.2099     0.2990    1.220x
    112     0.2499     0.4106     0.2724     0.3441    1.090x
    128     0.2529     0.4352     0.2755     0.3871    1.089x
    144     0.2867     0.5128     0.3350     0.4303    1.168x
    160     0.2898     0.5369     0.3389     0.4733    1.170x
    176     0.3328     0.6144     0.4003     0.5164    1.203x
    192     0.3369     0.6400     0.4024     0.5610    1.195x
    208     0.4241     0.7166     0.4632     0.6062    1.092x
    224     0.4289     0.7414     0.4669     0.6472    1.089x
    240     0.4311     0.8172     0.5274     0.6905    1.223x
    256     0.4946     0.8427     0.5315     0.7342    1.075x
  BAND K16N128 vs K32N128: M<=32 0.939x   M>32 1.187x

## 5120x10240xK4  (model.language_model.layers.1.linear_attn.in_proj_qkv; reconstruct threshold 280 rows)
  select_rows shipped(1,16,32)   -> K32N128    (K16N128=0.1714  K16N512=0.1854  K32N128=0.1568  K32N256=0.1762)
  select_rows corners            -> K16N128    (K16N128=0.4572  K16N512=0.6523  K32N128=0.4696  K32N256=0.6066)
  select_rows reachable-all      -> K16N128    (K16N128=3.2708  K16N512=5.2977  K32N128=3.7814  K32N256=4.8422)
      M    K16N128    K16N512    K32N128    K32N256   K16N128 vs K32N128
      1     0.0522     0.0512     0.0504     0.0484    0.966x
      2     0.0522     0.0511     0.0502     0.0481    0.961x
      3     0.0522     0.0512     0.0505     0.0484    0.967x
      4     0.0522     0.0512     0.0502     0.0482    0.961x
      6     0.0532     0.0522     0.0512     0.0491    0.962x
      8     0.0524     0.0532     0.0505     0.0491    0.964x
     12     0.0532     0.0573     0.0512     0.0502    0.962x
     16     0.0532     0.0595     0.0521     0.0512    0.979x
     24     0.0655     0.0676     0.0535     0.0740    0.817x
     32     0.0659     0.0747     0.0543     0.0767    0.823x
     40     0.0707     0.1126     0.0901     0.1003    1.275x
     48     0.0705     0.1183     0.0901     0.1023    1.279x
     56     0.0778     0.1249     0.0911     0.1260    1.171x
     64     0.0778     0.1321     0.0919     0.1280    1.181x
     72     0.0860     0.1681     0.1260     0.1505    1.464x
     80     0.0881     0.1761     0.1280     0.1536    1.454x
     88     0.0993     0.1812     0.1280     0.1761    1.289x
     96     0.1003     0.1884     0.1288     0.1784    1.284x
    112     0.1465     0.2314     0.1638     0.2038    1.118x
    128     0.1485     0.2450     0.1659     0.2294    1.117x
    144     0.1659     0.2898     0.2013     0.2550    1.214x
    160     0.1690     0.3030     0.2037     0.2806    1.206x
    176     0.1935     0.3461     0.2386     0.3055    1.233x
    192     0.1956     0.3594     0.2402     0.3317    1.228x
    208     0.2458     0.4031     0.2755     0.3564    1.121x
    224     0.2478     0.4159     0.2775     0.3822    1.120x
    240     0.2485     0.4598     0.3123     0.4065    1.257x
    256     0.2867     0.4731     0.3144     0.4324    1.096x
  BAND K16N128 vs K32N128: M<=32 0.934x   M>32 1.224x

## 5120x1024xK4  (model.language_model.layers.3.self_attn.k_proj; reconstruct threshold 64 rows)
  select_rows shipped(1,16,32)   -> K32N128    (K16N128=0.0773  K16N512=0.4714  K32N128=0.0769  K32N256=0.1669)
  select_rows corners            -> K32N128    (K16N128=0.0753  K16N512=0.4058  K32N128=0.0739  K32N256=0.1544)
  select_rows reachable-all      -> K16N128    (K16N128=0.3975  K16N512=2.8303  K32N128=0.4338  K32N256=1.0029)
      M    K16N128    K16N512    K32N128    K32N256   K16N128 vs K32N128
      1     0.0205     0.0601     0.0206     0.0307    1.007x
      2     0.0207     0.0604     0.0215     0.0317    1.038x
      3     0.0205     0.0614     0.0215     0.0308    1.050x
      4     0.0215     0.0635     0.0215     0.0319    1.000x
      6     0.0205     0.0696     0.0215     0.0329    1.047x
      8     0.0222     0.0787     0.0225     0.0357    1.013x
     12     0.0225     0.1085     0.0236     0.0399    1.045x
     16     0.0242     0.1444     0.0256     0.0481    1.056x
     24     0.0287     0.2017     0.0271     0.0728    0.944x
     32     0.0325     0.2669     0.0307     0.0881    0.944x
     40     0.0358     0.3523     0.0461     0.1147    1.288x
     48     0.0399     0.4022     0.0481     0.1219    1.205x
     56     0.0420     0.4475     0.0502     0.1546    1.195x
     64     0.0459     0.5130     0.0533     0.1690    1.163x
  BAND K16N128 vs K32N128: M<=32 1.014x   M>32 1.212x

## 5120x12288xK4  (model.language_model.layers.3.self_attn.q_proj; reconstruct threshold 280 rows)
  select_rows shipped(1,16,32)   -> K32N128    (K16N128=0.1948  K16N512=0.1997  K32N128=0.1762  K32N256=0.2017)
  select_rows corners            -> K16N128    (K16N128=0.5236  K16N512=0.6738  K32N128=0.5400  K32N256=0.6810)
  select_rows reachable-all      -> K16N128    (K16N128=3.7562  K16N512=5.4467  K32N128=4.3558  K32N256=5.4235)
      M    K16N128    K16N512    K32N128    K32N256   K16N128 vs K32N128
      1     0.0604     0.0584     0.0574     0.0563    0.950x
      2     0.0594     0.0584     0.0573     0.0563    0.965x
      3     0.0602     0.0584     0.0573     0.0563    0.952x
      4     0.0602     0.0584     0.0573     0.0563    0.952x
      6     0.0604     0.0594     0.0580     0.0571    0.960x
      8     0.0604     0.0604     0.0581     0.0573    0.962x
     12     0.0604     0.0625     0.0584     0.0575    0.966x
     16     0.0604     0.0645     0.0584     0.0583    0.966x
     24     0.0737     0.0717     0.0603     0.0858    0.818x
     32     0.0740     0.0768     0.0604     0.0870    0.817x
     40     0.0799     0.1188     0.1031     0.1137    1.291x
     48     0.0799     0.1229     0.1034     0.1150    1.295x
     56     0.0891     0.1290     0.1044     0.1423    1.172x
     64     0.0891     0.1341     0.1045     0.1434    1.172x
     72     0.0995     0.1759     0.1464     0.1700    1.472x
     80     0.1003     0.1802     0.1464     0.1716    1.459x
     88     0.1146     0.1864     0.1475     0.1986    1.287x
     96     0.1147     0.1915     0.1475     0.1997    1.286x
    112     0.1700     0.2376     0.1905     0.2283    1.120x
    128     0.1704     0.2488     0.1908     0.2560    1.120x
    144     0.1926     0.2951     0.2327     0.2837    1.208x
    160     0.1935     0.3052     0.2338     0.3113    1.208x
    176     0.2215     0.3523     0.2765     0.3400    1.248x
    192     0.2232     0.3625     0.2767     0.3675    1.240x
    208     0.2847     0.4096     0.3195     0.3963    1.122x
    224     0.2867     0.4219     0.3215     0.4250    1.121x
    240     0.2881     0.4680     0.3635     0.4526    1.262x
    256     0.3288     0.4782     0.3640     0.4803    1.107x
  BAND K16N128 vs K32N128: M<=32 0.929x   M>32 1.229x

## 5120x17408xK4  (model.language_model.layers.1.mlp.gate_proj; reconstruct threshold 280 rows)
  select_rows shipped(1,16,32)   -> K32N128    (K16N128=0.2601  K16N512=0.2461  K32N128=0.2386  K32N256=0.2692)
  select_rows corners            -> K16N128    (K16N128=0.7158  K16N512=0.7793  K32N128=0.7403  K32N256=0.8938)
  select_rows reachable-all      -> K16N128    (K16N128=5.1553  K16N512=6.2736  K32N128=5.9709  K32N256=7.1061)
      M    K16N128    K16N512    K32N128    K32N256   K16N128 vs K32N128
      1     0.0799     0.0758     0.0778     0.0758    0.974x
      2     0.0799     0.0758     0.0778     0.0758    0.974x
      3     0.0799     0.0758     0.0785     0.0758    0.983x
      4     0.0799     0.0758     0.0778     0.0758    0.974x
      6     0.0800     0.0768     0.0789     0.0759    0.986x
      8     0.0799     0.0768     0.0778     0.0766    0.974x
     12     0.0809     0.0789     0.0790     0.0768    0.976x
     16     0.0809     0.0802     0.0789     0.0777    0.975x
     24     0.0983     0.0860     0.0809     0.1148    0.823x
     32     0.0993     0.0901     0.0819     0.1157    0.825x
     40     0.1068     0.1423     0.1403     0.1515    1.313x
     48     0.1075     0.1461     0.1406     0.1526    1.308x
     56     0.1208     0.1497     0.1413     0.1876    1.169x
     64     0.1219     0.1536     0.1423     0.1887    1.168x
     72     0.1372     0.2058     0.2001     0.2243    1.458x
     80     0.1382     0.2090     0.2007     0.2253    1.452x
     88     0.1567     0.2140     0.2017     0.2611    1.288x
     96     0.1577     0.2171     0.2018     0.2611    1.280x
    112     0.2335     0.2734     0.2607     0.2980    1.117x
    128     0.2353     0.2806     0.2621     0.3348    1.114x
    144     0.2673     0.3359     0.3205     0.3697    1.199x
    160     0.2683     0.3441     0.3215     0.4065    1.198x
    176     0.3063     0.3994     0.3799     0.4424    1.240x
    192     0.3084     0.4079     0.3819     0.4791    1.238x
    208     0.3963     0.4649     0.4403     0.5153    1.111x
    224     0.3983     0.4731     0.4424     0.5530    1.111x
    240     0.3994     0.5282     0.5007     0.5888    1.254x
    256     0.4567     0.5366     0.5028     0.6257    1.101x
  BAND K16N128 vs K32N128: M<=32 0.944x   M>32 1.224x

## 5120x248320xK6  (lm_head; reconstruct threshold 280 rows)
  select_rows shipped(1,16,32)   -> K16N512    (K16N128=3.6701  K16N512=3.3014  K32N128=3.3489  K32N256=4.3520)
  select_rows corners            -> K16N128    (K16N128=10.2485  K16N512=11.9357  K32N128=12.0803  K32N256=21.2572)
  select_rows reachable-all      -> K16N128    (K16N128=73.1989  K16N512=97.6055  K32N128=98.7898  K32N256=173.1377)
      M    K16N128    K16N512    K32N128    K32N256   K16N128 vs K32N128
      1     1.1684     1.0732     1.0926     1.0803    0.935x
      2     1.1694     1.0750     1.0957     1.0833    0.937x
      3     1.1725     1.0783     1.0977     1.0865    0.936x
      4     1.1735     1.0803     1.0997     1.0881    0.937x
      6     1.1745     1.0854     1.1039     1.0926    0.940x
      8     1.1766     1.0875     1.1072     1.0977    0.941x
     12     1.1817     1.0947     1.1131     1.1029    0.942x
     16     1.1837     1.1008     1.1176     1.1090    0.944x
     24     1.3107     1.1162     1.1325     2.1504    0.864x
     32     1.3180     1.1274     1.1387     2.1627    0.864x
     40     1.4454     2.1671     2.2006     3.2051    1.522x
     48     1.4531     2.1811     2.2123     3.2165    1.523x
     56     1.6312     2.1924     2.2191     4.2587    1.360x
     64     1.6411     2.1985     2.2252     4.2732    1.356x
     72     1.8986     3.2410     3.2870     5.3115    1.731x
     80     1.9117     3.2573     3.2983     5.3238    1.725x
     88     2.1784     3.2652     3.3055     6.3662    1.517x
     96     2.1823     3.2758     3.3126     6.3775    1.518x
    112     3.2840     4.3285     4.3837     7.4331    1.335x
    128     3.2969     4.3515     4.3991     8.4890    1.334x
    144     3.8236     5.4047     5.4692     9.5380    1.430x
    160     3.8350     5.4252     5.4835    10.5925    1.430x
    176     4.3786     6.4799     6.5555    11.6480    1.497x
    192     4.3945     6.5014     6.5710    12.7034    1.495x
    208     5.7324     7.5592     7.6416    13.7564    1.333x
    224     5.7395     7.5730     7.6565    14.8101    1.334x
    240     5.7580     8.6374     8.7286    15.8648    1.516x
    256     6.5855     8.6477     8.7419    16.9165    1.327x
  BAND K16N128 vs K32N128: M<=32 0.924x   M>32 1.455x

## 6144x5120xK4  (model.language_model.layers.1.linear_attn.out_proj; reconstruct threshold 280 rows)
  select_rows shipped(1,16,32)   -> K32N128    (K16N128=0.1198  K16N512=0.1840  K32N128=0.1085  K32N256=0.1355)
  select_rows corners            -> K16N128    (K16N128=0.3256  K16N512=0.7493  K32N128=0.3316  K32N256=0.5031)
  select_rows reachable-all      -> K16N128    (K16N128=2.3304  K16N512=6.1468  K32N128=2.6646  K32N256=4.0440)
      M    K16N128    K16N512    K32N128    K32N256   K16N128 vs K32N128
      1     0.0358     0.0406     0.0348     0.0351    0.972x
      2     0.0348     0.0399     0.0338     0.0348    0.970x
      3     0.0357     0.0410     0.0348     0.0357    0.974x
      4     0.0348     0.0410     0.0338     0.0356    0.970x
      6     0.0358     0.0430     0.0348     0.0359    0.972x
      8     0.0358     0.0452     0.0348     0.0358    0.972x
     12     0.0360     0.0522     0.0358     0.0379    0.996x
     16     0.0359     0.0584     0.0358     0.0389    0.999x
     24     0.0471     0.0717     0.0379     0.0584    0.804x
     32     0.0481     0.0850     0.0379     0.0614    0.787x
     40     0.0512     0.1198     0.0623     0.0799    1.218x
     48     0.0521     0.1313     0.0625     0.0840    1.199x
     56     0.0553     0.1423     0.0645     0.1024    1.167x
     64     0.0563     0.1546     0.0645     0.1057    1.145x
     72     0.0632     0.1905     0.0891     0.1247    1.410x
     80     0.0635     0.2017     0.0891     0.1278    1.403x
     88     0.0737     0.2130     0.0907     0.1464    1.231x
     96     0.0747     0.2255     0.0922     0.1505    1.233x
    112     0.1034     0.2724     0.1157     0.1720    1.119x
    128     0.1045     0.2949     0.1178     0.1942    1.127x
    144     0.1180     0.3430     0.1423     0.2161    1.207x
    160     0.1208     0.3666     0.1444     0.2386    1.195x
    176     0.1400     0.4127     0.1681     0.2601    1.201x
    192     0.1423     0.4358     0.1707     0.2826    1.199x
    208     0.1731     0.4833     0.1946     0.3039    1.124x
    224     0.1751     0.5079     0.1966     0.3267    1.123x
    240     0.1774     0.5550     0.2212     0.3481    1.247x
    256     0.2058     0.5786     0.2242     0.3707    1.089x
  BAND K16N128 vs K32N128: M<=32 0.939x   M>32 1.199x

## roll-up (UNWEIGHTED geometric mean over geometries — each shape counts
   once, NOT once per linear that carries it; both bands are DECODE-class)
  decode<=32   K16N128 vs K32N128: 0.946x  (LOSS)
  verify>32    K16N128 vs K32N128: 1.244x  (GAIN)

  Both bands are per-token. A decode step runs M = B; a verify step runs
  M = B*(K+1) and is part of the same token loop. Neither band is prefill.
