Image-2.1-Calibrated-NVFP4 / reports /quantization_summary.json
ProCreations's picture
Release calibrated Image2.1 NVFP4 transformer with dynamic scaling and BF16 rank correction, native SM120 runtime, quality evidence and real-time demo
1961af5 verified
Raw History Blame Contribute Delete
715 Bytes
{
"modules": 224,
"transformer_bytes": 4870924032,
"mean_diagnostic_nrmse": 0.057958053558000495,
"max_diagnostic_nrmse": 0.1279429793357849,
"naive_mean_diagnostic_nrmse": 0.07584479178019267,
"protocol": "64 original BF16 trajectories;384 fit and384 disjoint diagnostic rows per layer, stratified across prompts and timesteps. Five smoothing exponents, two weight-scale methods. Rank128 randomized SVD (seed1234,q160,niter3); optional activation-weighted15-point block-scale MSE search. Native W4A4 kernels with dynamic activation amax used for all calibrated selection/diagnostic output errors. Plain reference uses original static range. No held-out image evaluation prompts used for calibration."
}