causal prefix error = 0.0

round 1: full input ids = [0, 1, 2]
H0 shape = (3, 4) last row = [-0.2455 -0.267   0.253   0.0263]
layer 1: Q/K/V (3, 2), attention (3, 3), H (3, 4)
  attention last row = [0.4434 0.2687 0.288 ]
  H last row = [-0.3924 -0.2914  0.0515 -0.1443]
layer 2: Q/K/V (3, 2), attention (3, 3), H (3, 4)
  attention last row = [0.3699 0.3198 0.3103]
  H last row = [-0.0973 -0.1626 -0.1454 -0.0003]
logits = [-0.3701 -0.6962  0.0472  0.162   0.373  -0.6302]
probabilities = [0.128  0.0923 0.1942 0.2178 0.269  0.0986]
selected id = 4

round 2: full input ids = [0, 1, 2, 4]
H0 shape = (4, 4) last row = [-0.4065 -0.0488 -0.6287 -0.398 ]
layer 1: Q/K/V (4, 2), attention (4, 4), H (4, 4)
  attention last row = [0.4245 0.0853 0.1917 0.2985]
  H last row = [-0.5065 -0.5458 -0.8496 -0.4123]
layer 2: Q/K/V (4, 2), attention (4, 4), H (4, 4)
  attention last row = [0.2526 0.2478 0.2415 0.2581]
  H last row = [-0.0295 -1.0339 -0.6432 -0.3141]
logits = [-0.2638 -1.0283 -0.1541  0.0775  0.8219 -0.6144]
probabilities = [0.1307 0.0608 0.1458 0.1838 0.3869 0.092 ]
selected id = 4

round 3: full input ids = [0, 1, 2, 4, 4]
H0 shape = (5, 4) last row = [-0.3922 -0.1309 -0.6929 -0.3792]
layer 1: Q/K/V (5, 2), attention (5, 5), H (5, 4)
  attention last row = [0.3355 0.0658 0.1521 0.2375 0.2091]
  H last row = [-0.5356 -0.5761 -0.9774 -0.4379]
layer 2: Q/K/V (5, 2), attention (5, 5), H (5, 4)
  attention last row = [0.1997 0.1979 0.1931 0.2047 0.2046]
  H last row = [-0.0312 -0.975  -0.7795 -0.3523]
logits = [-0.3131 -1.0132 -0.2332  0.0479  0.7339 -0.5271]
probabilities = [0.1304 0.0647 0.1412 0.187  0.3714 0.1053]
selected id = 4

total input positions = 12
CHECK: PASS
