checkpoint 0 max 0.00048828125 rms 6.430571e-05 outside tolerance 0
checkpoint 1 max 0.00048828125 rms 6.2104096e-05 outside tolerance 0
checkpoint 2 max 0.005859375 rms 0.0004562542 outside tolerance 35
checkpoint 3 max 0.00269413 rms 0.0002301382 outside tolerance 13
GELU applied to actual GEMM1 max 0.00048828125 rms 5.1055486e-05 outside tolerance 0
FP8 threshold crossings 26 rows [ 2  3  8  9 13 18 19 28 29 30 31 32 34 36 37 40 42 44 49 51 60 63]
crossing indices [[2, 35], [3, 3], [8, 3], [8, 41], [9, 19], [9, 52], [13, 6], [18, 26], [19, 23], [28, 36], [29, 22], [30, 25], [31, 47], [31, 58], [32, 49], [34, 54], [36, 2], [37, 54], [40, 40], [40, 52], [42, 51], [44, 63], [49, 45], [51, 7], [60, 35], [63, 37]]
GEMM2 using actual GELU1, quantized max 0.00024414062 rms 3.3769986e-05 outside tolerance 0
final using actual GELU1, quantized max 0.00024414062 rms 3.1249303e-05 outside tolerance 0
final GELU using actual GEMM2 max 0.00024414062 rms 2.5981306e-05 outside tolerance 0
final mismatch rows [13 28 36]
cosine 0.9999943
minimum row cosine 0.9998975