{ "prediction_1": { "claim": "residuals are numerical-zero independent of n,r", "sweep": [ { "n": 16, "r": 2, "iterations": 4, "max_row_residual": 1.1102230246251565e-16, "max_shared_column_residual": 3.552713678800501e-15, "projected_KL": 22.349774808788002, "reference_row_normalized_KL": 21.970317896801408, "hessian_null_norm": 4.782987168225453e-15 }, { "n": 16, "r": 4, "iterations": 4, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 3.552713678800501e-15, "projected_KL": 81.4472146488221, "reference_row_normalized_KL": 81.23143408788488, "hessian_null_norm": 1.7199501139797033e-15 }, { "n": 16, "r": 8, "iterations": 4, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 7.105427357601002e-15, "projected_KL": 270.3327018111314, "reference_row_normalized_KL": 270.0767076299735, "hessian_null_norm": 1.961044852356119e-15 }, { "n": 16, "r": 16, "iterations": 4, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 7.405187574249794e-13, "projected_KL": 639.6280424342364, "reference_row_normalized_KL": 638.849357586517, "hessian_null_norm": 9.930136612989092e-16 }, { "n": 32, "r": 2, "iterations": 4, "max_row_residual": 1.1102230246251565e-16, "max_shared_column_residual": 5.329070518200751e-15, "projected_KL": 76.88858258234636, "reference_row_normalized_KL": 76.47797845540806, "hessian_null_norm": 5.617333549722722e-15 }, { "n": 32, "r": 4, "iterations": 4, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 1.7763568394002505e-15, "projected_KL": 267.0935436685017, "reference_row_normalized_KL": 267.0568060243071, "hessian_null_norm": 2.175583928816829e-15 }, { "n": 32, "r": 8, "iterations": 4, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 4.574118861455645e-13, "projected_KL": 648.1912323747913, "reference_row_normalized_KL": 647.3843260564644, "hessian_null_norm": 2.945754565036118e-15 }, { "n": 32, "r": 16, "iterations": 4, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 1.7763568394002505e-15, "projected_KL": 1432.9330750053655, "reference_row_normalized_KL": 1432.5761948821998, "hessian_null_norm": 2.5023156146241385e-15 }, { "n": 64, "r": 2, "iterations": 4, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 1.7763568394002505e-14, "projected_KL": 200.50463638080583, "reference_row_normalized_KL": 200.00809649415328, "hessian_null_norm": 2.5618982671915014e-14 }, { "n": 64, "r": 4, "iterations": 4, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 1.2434497875801753e-14, "projected_KL": 459.5302941648396, "reference_row_normalized_KL": 459.18127872487105, "hessian_null_norm": 1.0509068785146553e-14 }, { "n": 64, "r": 8, "iterations": 4, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 1.1546319456101628e-14, "projected_KL": 1233.0106208627142, "reference_row_normalized_KL": 1232.1658981278708, "hessian_null_norm": 1.3440578383968888e-14 }, { "n": 64, "r": 16, "iterations": 4, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 7.105427357601002e-15, "projected_KL": 2791.8952543271416, "reference_row_normalized_KL": 2791.3472049464085, "hessian_null_norm": 6.993523870246153e-15 }, { "n": 128, "r": 2, "iterations": 4, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 0.0, "projected_KL": 294.05135467779064, "reference_row_normalized_KL": 293.9327992195989, "hessian_null_norm": 4.5496934130773415e-14 }, { "n": 128, "r": 4, "iterations": 4, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 2.842170943040401e-14, "projected_KL": 947.790297888946, "reference_row_normalized_KL": 947.2156128648421, "hessian_null_norm": 2.2748467065386707e-14 }, { "n": 128, "r": 8, "iterations": 4, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 2.1316282072803006e-14, "projected_KL": 2548.2812386598325, "reference_row_normalized_KL": 2547.8086204853953, "hessian_null_norm": 2.869792300935272e-14 }, { "n": 128, "r": 16, "iterations": 4, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 7.105427357601002e-15, "projected_KL": 5689.807386944296, "reference_row_normalized_KL": 5689.261505681804, "hessian_null_norm": 1.703821625409574e-14 }, { "n": 256, "r": 2, "iterations": 3, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 2.984279490192421e-13, "projected_KL": 690.3801290623901, "reference_row_normalized_KL": 690.2604154191021, "hessian_null_norm": 1.2224639458285405e-13 }, { "n": 256, "r": 4, "iterations": 4, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 7.105427357601002e-14, "projected_KL": 1957.818301241046, "reference_row_normalized_KL": 1957.7563239924232, "hessian_null_norm": 6.665481691674458e-14 }, { "n": 256, "r": 8, "iterations": 4, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 4.973799150320701e-14, "projected_KL": 4785.057366239143, "reference_row_normalized_KL": 4784.625678593483, "hessian_null_norm": 5.3290705182007514e-14 }, { "n": 256, "r": 16, "iterations": 4, "max_row_residual": 2.220446049250313e-16, "max_shared_column_residual": 4.618527782440651e-14, "projected_KL": 11319.745588977632, "reference_row_normalized_KL": 11319.284375240084, "hessian_null_norm": 3.940150075155706e-14 } ] }, "prediction_2": { "claim": "covariance-sum HVP is exact and reduced curvature is positive", "sweep": [ { "r": 2, "predicted_relative_error": 0.0, "observed_relative_error": 1.1423217667413144e-13, "predicted_positive_gauge_curvature": "> 0", "observed_min_reduced_eigenvalue": 21.503171812151656 }, { "r": 4, "predicted_relative_error": 0.0, "observed_relative_error": 2.0700800335795706e-16, "predicted_positive_gauge_curvature": "> 0", "observed_min_reduced_eigenvalue": 5.994879936143085 }, { "r": 8, "predicted_relative_error": 0.0, "observed_relative_error": 3.9327182635266995e-16, "predicted_positive_gauge_curvature": "> 0", "observed_min_reduced_eigenvalue": 1.6945364246549077 }, { "r": 12, "predicted_relative_error": 0.0, "observed_relative_error": 3.9198222699432094e-16, "predicted_positive_gauge_curvature": "> 0", "observed_min_reduced_eigenvalue": 0.7834050345856374 }, { "r": 16, "predicted_relative_error": 0.0, "observed_relative_error": 2.9076393011614695e-16, "predicted_positive_gauge_curvature": "> 0", "observed_min_reduced_eigenvalue": 0.4840439077177008 } ] }, "prediction_3": { "claim": "factor storage scales as n^2/(2nr+r), and thin equals dense factor application", "sweep": [ { "n": 32, "r": 4, "predicted_output_relative_error": 0.0, "observed_output_relative_error": 2.936797370715141e-16, "predicted_storage_ratio": 3.9384615384615387, "observed_storage_ratio": 3.9384615384615387, "dense_entries": 1024, "factor_entries": 260 }, { "n": 64, "r": 4, "predicted_output_relative_error": 0.0, "observed_output_relative_error": 3.386865835672514e-16, "predicted_storage_ratio": 7.937984496124031, "observed_storage_ratio": 7.937984496124031, "dense_entries": 4096, "factor_entries": 516 }, { "n": 128, "r": 8, "predicted_output_relative_error": 0.0, "observed_output_relative_error": 4.951231020480598e-16, "predicted_storage_ratio": 7.968871595330739, "observed_storage_ratio": 7.968871595330739, "dense_entries": 16384, "factor_entries": 2056 }, { "n": 256, "r": 8, "predicted_output_relative_error": 0.0, "observed_output_relative_error": 6.776688632570342e-16, "predicted_storage_ratio": 15.968810916179336, "observed_storage_ratio": 15.968810916179336, "dense_entries": 65536, "factor_entries": 4104 }, { "n": 512, "r": 16, "predicted_output_relative_error": 0.0, "observed_output_relative_error": 6.422168098100112e-16, "predicted_storage_ratio": 15.984390243902439, "observed_storage_ratio": 15.984390243902439, "dense_entries": 262144, "factor_entries": 16400 } ] }, "mini_experiment": { "n": 512, "r": 16, "d": 32, "dense_apply_median_ms": 1.2821340060327202, "thin_apply_median_ms": 0.24867449974408373, "speedup": 5.15587246521936, "projection_shared_column_residual": 7.460698725481052e-14, "dense_attention_state_entries": 262144, "idea_factor_entries": 16400 } }