FP32 mapped graph
authors: cosine=1.000000000 clipped=0
siglip: cosine=1.000000000 clipped=0
caffeine: cosine=1.000000000 clipped=0
robosign: cosine=1.000000000 clipped=0
fried_fish: cosine=1.000000000 clipped=0
cow_beach2: cosine=1.000000000 clipped=0
Equalized FP32 graph
authors: cosine=1.000000000 clipped=0
siglip: cosine=1.000000000 clipped=0
caffeine: cosine=1.000000000 clipped=0
robosign: cosine=1.000000000 clipped=0
fried_fish: cosine=1.000000000 clipped=0
cow_beach2: cosine=1.000000000 clipped=0
Tensor scales, nearest weights
authors: cosine=0.984671701 clipped=0
siglip: cosine=0.984611895 clipped=1
caffeine: cosine=0.991943868 clipped=1
robosign: cosine=0.981991425 clipped=14
fried_fish: cosine=0.983353749 clipped=44
cow_beach2: cosine=0.964912135 clipped=115
GPTQ embedding: 0.04545 weighted error ratio
GPTQ encoder.layer0.attention.qkv: 0.85174 weighted error ratio
GPTQ encoder.layer0.attention.output: 0.18091 weighted error ratio
GPTQ encoder.layer0.mlp.up: 0.84690 weighted error ratio
GPTQ encoder.layer0.mlp.down: 0.79659 weighted error ratio
GPTQ encoder.layer1.attention.qkv: 0.78068 weighted error ratio
GPTQ encoder.layer1.attention.output: 0.15878 weighted error ratio
GPTQ encoder.layer1.mlp.up: 0.76547 weighted error ratio
GPTQ encoder.layer1.mlp.down: 0.60102 weighted error ratio
GPTQ encoder.layer2.attention.qkv: 0.85973 weighted error ratio
GPTQ encoder.layer2.attention.output: 0.41378 weighted error ratio
GPTQ encoder.layer2.mlp.up: 0.84448 weighted error ratio
GPTQ encoder.layer2.mlp.down: 0.59522 weighted error ratio
GPTQ encoder.layer3.attention.qkv: 0.87929 weighted error ratio
GPTQ encoder.layer3.attention.output: 0.51084 weighted error ratio
GPTQ encoder.layer3.mlp.up: 0.80978 weighted error ratio
GPTQ encoder.layer3.mlp.down: 0.56302 weighted error ratio
GPTQ encoder.layer4.attention.qkv: 0.83634 weighted error ratio
GPTQ encoder.layer4.attention.output: 0.46678 weighted error ratio
GPTQ encoder.layer4.mlp.up: 0.76478 weighted error ratio
GPTQ encoder.layer4.mlp.down: 0.57484 weighted error ratio
GPTQ encoder.layer5.attention.qkv: 0.82972 weighted error ratio
GPTQ encoder.layer5.attention.output: 0.48863 weighted error ratio
GPTQ encoder.layer5.mlp.up: 0.76482 weighted error ratio
GPTQ encoder.layer5.mlp.down: 0.58455 weighted error ratio
GPTQ encoder.layer6.attention.qkv: 0.80546 weighted error ratio
GPTQ encoder.layer6.attention.output: 0.44571 weighted error ratio
GPTQ encoder.layer6.mlp.up: 0.75163 weighted error ratio
GPTQ encoder.layer6.mlp.down: 0.64796 weighted error ratio
GPTQ encoder.layer7.attention.qkv: 0.77634 weighted error ratio
GPTQ encoder.layer7.attention.output: 0.45982 weighted error ratio
GPTQ encoder.layer7.mlp.up: 0.75109 weighted error ratio
GPTQ encoder.layer7.mlp.down: 0.57839 weighted error ratio
GPTQ encoder.layer8.attention.qkv: 0.76324 weighted error ratio
GPTQ encoder.layer8.attention.output: 0.42414 weighted error ratio
GPTQ encoder.layer8.mlp.up: 0.75372 weighted error ratio
GPTQ encoder.layer8.mlp.down: 0.65282 weighted error ratio
GPTQ encoder.layer9.attention.qkv: 0.75168 weighted error ratio
GPTQ encoder.layer9.attention.output: 0.43574 weighted error ratio
GPTQ encoder.layer9.mlp.up: 0.77489 weighted error ratio
GPTQ encoder.layer9.mlp.down: 0.65482 weighted error ratio
GPTQ encoder.layer10.attention.qkv: 0.76228 weighted error ratio
GPTQ encoder.layer10.attention.output: 0.44491 weighted error ratio
GPTQ encoder.layer10.mlp.up: 0.76155 weighted error ratio
GPTQ encoder.layer10.mlp.down: 0.65682 weighted error ratio
GPTQ encoder.layer11.attention.qkv: 0.74258 weighted error ratio
GPTQ encoder.layer11.attention.output: 0.37619 weighted error ratio
GPTQ encoder.layer11.mlp.up: 0.74972 weighted error ratio
GPTQ encoder.layer11.mlp.down: 0.65363 weighted error ratio
GPTQ encoder.layer12.attention.qkv: 0.75466 weighted error ratio
GPTQ encoder.layer12.attention.output: 0.44553 weighted error ratio
GPTQ encoder.layer12.mlp.up: 0.71996 weighted error ratio
GPTQ encoder.layer12.mlp.down: 0.63100 weighted error ratio
GPTQ encoder.layer13.attention.qkv: 0.74842 weighted error ratio
GPTQ encoder.layer13.attention.output: 0.43893 weighted error ratio
GPTQ encoder.layer13.mlp.up: 0.71503 weighted error ratio
GPTQ encoder.layer13.mlp.down: 0.64946 weighted error ratio
GPTQ encoder.layer14.attention.qkv: 0.75508 weighted error ratio
GPTQ encoder.layer14.attention.output: 0.39102 weighted error ratio
GPTQ encoder.layer14.mlp.up: 0.71314 weighted error ratio
GPTQ encoder.layer14.mlp.down: 0.65897 weighted error ratio
GPTQ encoder.layer15.attention.qkv: 0.74468 weighted error ratio
GPTQ encoder.layer15.attention.output: 0.33622 weighted error ratio
GPTQ encoder.layer15.mlp.up: 0.60444 weighted error ratio
GPTQ encoder.layer15.mlp.down: 0.73127 weighted error ratio
GPTQ encoder.layer16.attention.qkv: 0.73214 weighted error ratio
GPTQ encoder.layer16.attention.output: 0.31962 weighted error ratio
GPTQ encoder.layer16.mlp.up: 0.44768 weighted error ratio
GPTQ encoder.layer16.mlp.down: 0.86710 weighted error ratio
GPTQ encoder.layer17.attention.qkv: 0.72280 weighted error ratio
GPTQ encoder.layer17.attention.output: 0.22744 weighted error ratio
GPTQ encoder.layer17.mlp.up: 0.34973 weighted error ratio
GPTQ encoder.layer17.mlp.down: 0.87600 weighted error ratio
GPTQ encoder.layer18.attention.qkv: 0.70189 weighted error ratio
GPTQ encoder.layer18.attention.output: 0.22486 weighted error ratio
GPTQ encoder.layer18.mlp.up: 0.37137 weighted error ratio
GPTQ encoder.layer18.mlp.down: 0.88315 weighted error ratio
GPTQ encoder.layer19.attention.qkv: 0.70547 weighted error ratio
GPTQ encoder.layer19.attention.output: 0.22913 weighted error ratio
GPTQ encoder.layer19.mlp.up: 0.31295 weighted error ratio
GPTQ encoder.layer19.mlp.down: 0.92805 weighted error ratio
GPTQ encoder.layer20.attention.qkv: 0.69029 weighted error ratio
GPTQ encoder.layer20.attention.output: 0.20517 weighted error ratio
GPTQ encoder.layer20.mlp.up: 0.28757 weighted error ratio
GPTQ encoder.layer20.mlp.down: 0.96194 weighted error ratio
GPTQ encoder.layer21.attention.qkv: 0.68404 weighted error ratio
GPTQ encoder.layer21.attention.output: 0.19830 weighted error ratio
GPTQ encoder.layer21.mlp.up: 0.26373 weighted error ratio
GPTQ encoder.layer21.mlp.down: 0.97674 weighted error ratio
GPTQ encoder.layer22.attention.qkv: 0.66529 weighted error ratio
GPTQ encoder.layer22.attention.output: 0.19613 weighted error ratio
GPTQ encoder.layer22.mlp.up: 0.26460 weighted error ratio
GPTQ encoder.layer22.mlp.down: 0.97545 weighted error ratio
GPTQ encoder.layer23.attention.qkv: 0.64868 weighted error ratio
GPTQ encoder.layer23.attention.output: 0.20788 weighted error ratio
GPTQ encoder.layer23.mlp.up: 0.26369 weighted error ratio
GPTQ encoder.layer23.mlp.down: 0.96558 weighted error ratio
GPTQ encoder.layer24.attention.qkv: 0.63468 weighted error ratio
GPTQ encoder.layer24.attention.output: 0.16593 weighted error ratio
GPTQ encoder.layer24.mlp.up: 0.29251 weighted error ratio
GPTQ encoder.layer24.mlp.down: 0.97196 weighted error ratio
GPTQ encoder.layer25.attention.qkv: 0.60947 weighted error ratio
GPTQ encoder.layer25.attention.output: 0.14492 weighted error ratio
GPTQ encoder.layer25.mlp.up: 0.36167 weighted error ratio
GPTQ encoder.layer25.mlp.down: 0.91258 weighted error ratio
GPTQ encoder.layer26.attention.qkv: 0.59607 weighted error ratio
GPTQ encoder.layer26.attention.output: 0.11493 weighted error ratio
GPTQ encoder.layer26.mlp.up: 0.52445 weighted error ratio
GPTQ encoder.layer26.mlp.down: 0.83664 weighted error ratio
GPTQ map.attention.query: 0.02027 weighted error ratio
GPTQ map.attention.kv: 0.50968 weighted error ratio
GPTQ map.attention.output: 0.02820 weighted error ratio
GPTQ map.mlp.up: 0.02461 weighted error ratio
GPTQ map.mlp.down: 0.05305 weighted error ratio
Tensor scales, GPTQ weights and bias correction
authors: cosine=0.985422278 clipped=0
siglip: cosine=0.987191279 clipped=1
caffeine: cosine=0.989431021 clipped=0
robosign: cosine=0.983603028 clipped=10
fried_fish: cosine=0.985032662 clipped=46
cow_beach2: cosine=0.959466363 clipped=114