21 lines
6.5 KiB
JSON
21 lines
6.5 KiB
JSON
{"repeat": 32, "batch_size": 16384, "dim": 1024, "attn": 2.1414458751678467, "ln": 1.6365647315979004, "ln_compiled": 1.799367368221283, "gelu": 1.0930374264717102, "gelu_compiled": 1.094818115234375, "standard": 4.159651696681976, "my_standard": 4.696495831012726, "standard_compiled": 3.675594925880432, "sb": 4.1465312242507935}
|
|
{"repeat": 32, "batch_size": 32768, "dim": 1024, "attn": 4.100345075130463, "ln": 3.1594187021255493, "ln_compiled": 3.437422215938568, "gelu": 2.109348773956299, "gelu_compiled": 2.11450457572937, "standard": 7.706902921199799, "my_standard": 8.799396455287933, "standard_compiled": 6.735652685165405, "sb": 7.66376405954361}
|
|
{"repeat": 32, "batch_size": 65536, "dim": 1024, "attn": 7.953710854053497, "ln": 6.236426532268524, "ln_compiled": 6.746955215930939, "gelu": 4.164382815361023, "gelu_compiled": 4.171714186668396, "standard": 14.894917607307434, "my_standard": 17.042435705661774, "standard_compiled": 12.985721230506897, "sb": 14.6140456199646}
|
|
{"repeat": 32, "batch_size": 131072, "dim": 1024, "attn": 15.638880431652069, "ln": 12.333884835243225, "ln_compiled": 13.272866606712341, "gelu": 8.228793740272522, "gelu_compiled": 8.243747055530548, "standard": 29.425136744976044, "my_standard": 35.08377820253372, "standard_compiled": 25.69487690925598, "sb": 28.760001063346863}
|
|
{"repeat": 32, "batch_size": 16384, "dim": 1280, "attn": 2.627238631248474, "ln": 2.0098239183425903, "ln_compiled": 2.4197474122047424, "gelu": 1.3455823063850403, "gelu_compiled": 1.35069340467453, "standard": 5.554787814617157, "my_standard": 6.2290579080581665, "standard_compiled": 5.132324993610382, "sb": 5.4178386926651}
|
|
{"repeat": 32, "batch_size": 32768, "dim": 1280, "attn": 5.0596073269844055, "ln": 3.903590142726898, "ln_compiled": 4.719957709312439, "gelu": 2.6203468441963196, "gelu_compiled": 2.627365291118622, "standard": 10.546617209911346, "my_standard": 11.850126087665558, "standard_compiled": 9.685918688774109, "sb": 10.088451206684113}
|
|
{"repeat": 32, "batch_size": 65536, "dim": 1280, "attn": 9.845800697803497, "ln": 7.711298763751984, "ln_compiled": 9.292080998420715, "gelu": 5.172915756702423, "gelu_compiled": 5.180932581424713, "standard": 21.371990442276, "my_standard": 23.921720683574677, "standard_compiled": 19.669152796268463, "sb": 20.267993211746216}
|
|
{"repeat": 32, "batch_size": 131072, "dim": 1280, "attn": 19.375711679458618, "ln": 15.333592891693115, "ln_compiled": 18.245264887809753, "gelu": 10.264746844768524, "gelu_compiled": 10.283775627613068, "standard": 41.79700464010239, "my_standard": 45.84744572639465, "standard_compiled": 38.35208714008331, "sb": 38.35364431142807}
|
|
{"repeat": 32, "batch_size": 16384, "dim": 1408, "attn": 2.9110386967658997, "ln": 2.1998360753059387, "ln_compiled": 2.581551671028137, "gelu": 1.4731436967849731, "gelu_compiled": 1.478634774684906, "standard": 6.764143705368042, "my_standard": 7.331632077693939, "standard_compiled": 6.24605268239975, "sb": 6.325609982013702}
|
|
{"repeat": 32, "batch_size": 32768, "dim": 1408, "attn": 5.542516708374023, "ln": 4.289716482162476, "ln_compiled": 5.065307021141052, "gelu": 2.8742849826812744, "gelu_compiled": 2.882353961467743, "standard": 12.749537825584412, "my_standard": 13.79828155040741, "standard_compiled": 11.728867888450623, "sb": 11.642806231975555}
|
|
{"repeat": 32, "batch_size": 65536, "dim": 1408, "attn": 10.80312579870224, "ln": 8.471302688121796, "ln_compiled": 9.96796041727066, "gelu": 5.681410431861877, "gelu_compiled": 5.6905597448349, "standard": 25.19702911376953, "my_standard": 27.226239442825317, "standard_compiled": 23.22910726070404, "sb": 22.682294249534607}
|
|
{"repeat": 32, "batch_size": 131072, "dim": 1408, "attn": 21.284908056259155, "ln": 16.85701310634613, "ln_compiled": 19.643358886241913, "gelu": 11.292420327663422, "gelu_compiled": 11.314474046230316, "standard": 50.06787180900574, "my_standard": 54.29378151893616, "standard_compiled": 44.58653926849365, "sb": 45.359253883361816}
|
|
{"repeat": 32, "batch_size": 16384, "dim": 1664, "attn": 3.382459282875061, "ln": 2.6206374168395996, "ln_compiled": 2.9666870832443237, "gelu": 1.7263293266296387, "gelu_compiled": 1.7317384481430054, "standard": 8.414775133132935, "my_standard": 9.117811918258667, "standard_compiled": 7.7542513608932495, "sb": 7.70898163318634}
|
|
{"repeat": 32, "batch_size": 32768, "dim": 1664, "attn": 6.468378007411957, "ln": 5.125559866428375, "ln_compiled": 5.791269242763519, "gelu": 3.3864825963974, "gelu_compiled": 3.3920034766197205, "standard": 16.016244888305664, "my_standard": 17.25083589553833, "standard_compiled": 14.60808515548706, "sb": 14.347739517688751}
|
|
{"repeat": 32, "batch_size": 65536, "dim": 1664, "attn": 12.645229697227478, "ln": 10.13532280921936, "ln_compiled": 11.427387595176697, "gelu": 6.6957250237464905, "gelu_compiled": 6.711684167385101, "standard": 31.792201101779938, "my_standard": 34.31189805269241, "standard_compiled": 29.10037338733673, "sb": 28.3128023147583}
|
|
{"repeat": 32, "batch_size": 131072, "dim": 1664, "attn": 24.970605969429016, "ln": 20.182937383651733, "ln_compiled": 22.7489173412323, "gelu": 13.326868414878845, "gelu_compiled": 13.345755636692047, "standard": 63.46555054187775, "my_standard": 70.19880414009094, "standard_compiled": 56.40875548124313, "sb": 56.22846633195877}
|
|
{"repeat": 32, "batch_size": 16384, "dim": 2048, "attn": 4.080049693584442, "ln": 3.2655522227287292, "ln_compiled": 3.3329352736473083, "gelu": 2.108432352542877, "gelu_compiled": 2.114713191986084, "standard": 11.370822787284851, "my_standard": 12.234866619110107, "standard_compiled": 10.377615690231323, "sb": 10.209612548351288}
|
|
{"repeat": 32, "batch_size": 32768, "dim": 2048, "attn": 7.74645060300827, "ln": 6.418220698833466, "ln_compiled": 6.55733048915863, "gelu": 4.163652658462524, "gelu_compiled": 4.171028733253479, "standard": 21.39316499233246, "my_standard": 23.04024249315262, "standard_compiled": 19.431106746196747, "sb": 18.732361495494843}
|
|
{"repeat": 32, "batch_size": 65536, "dim": 2048, "attn": 15.235155820846558, "ln": 12.684382498264313, "ln_compiled": 12.895286083221436, "gelu": 8.228868246078491, "gelu_compiled": 8.242718875408173, "standard": 42.55136102437973, "my_standard": 45.82635313272476, "standard_compiled": 38.663335144519806, "sb": 36.76284849643707}
|
|
{"repeat": 32, "batch_size": 131072, "dim": 2048, "attn": 30.24454414844513, "ln": 25.25731921195984, "ln_compiled": 25.67601203918457, "gelu": 16.384944319725037, "gelu_compiled": 16.409948468208313, "standard": 84.26841348409653, "my_standard": 91.10662341117859, "standard_compiled": 76.89539343118668, "sb": 71.73164188861847}
|