{"repetition": 0, "order": ["sdpa", "triton"], "seq_len": 1024, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.06253215789794922, "triton_event_ms": 0.04827370643615723, "speedup": 1.295366826258701}
{"repetition": 1, "order": ["triton", "sdpa"], "seq_len": 1024, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.06256671905517579, "triton_event_ms": 0.048174721399943034, "speedup": 1.2987458409100374}
{"repetition": 2, "order": ["sdpa", "triton"], "seq_len": 1024, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.06257247924804688, "triton_event_ms": 0.048352320988972984, "speedup": 1.2940946363736474}
{"repetition": 3, "order": ["triton", "sdpa"], "seq_len": 1024, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.06253365198771159, "triton_event_ms": 0.04813151995340983, "speedup": 1.2992245424254767}
{"repetition": 4, "order": ["sdpa", "triton"], "seq_len": 1024, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.06251338958740234, "triton_event_ms": 0.048249918619791665, "speedup": 1.2956164772008534}
{"repetition": 5, "order": ["triton", "sdpa"], "seq_len": 1024, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.06255861282348633, "triton_event_ms": 0.04837674776713053, "speedup": 1.293154577579761}
{"repetition": 6, "order": ["sdpa", "triton"], "seq_len": 1024, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.0625047492980957, "triton_event_ms": 0.04834570566813151, "speedup": 1.2928707614107189}
{"repetition": 7, "order": ["triton", "sdpa"], "seq_len": 1024, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.057854932149251304, "triton_event_ms": 0.04599808057149251, "speedup": 1.2577683988211266}
{"repetition": 8, "order": ["sdpa", "triton"], "seq_len": 1024, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.057929280598958334, "triton_event_ms": 0.044848426183064775, "speedup": 1.291668081339118}
{"repetition": 9, "order": ["triton", "sdpa"], "seq_len": 1024, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.057881495157877605, "triton_event_ms": 0.04476458549499512, "speedup": 1.2930197949526199}
{"repetition": 0, "order": ["sdpa", "triton"], "seq_len": 2048, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.14671744028727213, "triton_event_ms": 0.15172939300537108, "speedup": 0.9669678193603428}
{"repetition": 1, "order": ["triton", "sdpa"], "seq_len": 2048, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.14656202952067057, "triton_event_ms": 0.15197279612223308, "speedup": 0.9643964792408598}
{"repetition": 2, "order": ["sdpa", "triton"], "seq_len": 2048, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.1467746098836263, "triton_event_ms": 0.15003584543863932, "speedup": 0.9782636239661356}
{"repetition": 3, "order": ["triton", "sdpa"], "seq_len": 2048, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.13672799428304036, "triton_event_ms": 0.14129642486572266, "speedup": 0.9676677553092813}
{"repetition": 4, "order": ["sdpa", "triton"], "seq_len": 2048, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.13662741343180337, "triton_event_ms": 0.14160074869791667, "speedup": 0.9648777615101235}
{"repetition": 5, "order": ["triton", "sdpa"], "seq_len": 2048, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.1365948740641276, "triton_event_ms": 0.14193354288736978, "speedup": 0.9623861370988347}
{"repetition": 6, "order": ["sdpa", "triton"], "seq_len": 2048, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.13701098124186198, "triton_event_ms": 0.14175338745117189, "speedup": 0.9665446710333926}
{"repetition": 7, "order": ["triton", "sdpa"], "seq_len": 2048, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.13741525014241537, "triton_event_ms": 0.14199488321940104, "speedup": 0.9677479006767482}
{"repetition": 8, "order": ["sdpa", "triton"], "seq_len": 2048, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.13689205169677734, "triton_event_ms": 0.14169333140055337, "speedup": 0.9661149917478948}
{"repetition": 9, "order": ["triton", "sdpa"], "seq_len": 2048, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.13639957427978516, "triton_event_ms": 0.1422372309366862, "speedup": 0.958958307761914}
{"repetition": 0, "order": ["sdpa", "triton"], "seq_len": 4096, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.46475667317708336, "triton_event_ms": 0.5049159240722656, "speedup": 0.9204634891066844}
{"repetition": 1, "order": ["triton", "sdpa"], "seq_len": 4096, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.43656702677408854, "triton_event_ms": 0.507698974609375, "speedup": 0.8598934577521739}
{"repetition": 2, "order": ["sdpa", "triton"], "seq_len": 4096, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.43430892944335936, "triton_event_ms": 0.5084651692708333, "speedup": 0.8541566968415595}
{"repetition": 3, "order": ["triton", "sdpa"], "seq_len": 4096, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.4369736735026042, "triton_event_ms": 0.5079674784342448, "speedup": 0.8602394681831376}
{"repetition": 4, "order": ["sdpa", "triton"], "seq_len": 4096, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.4358643086751302, "triton_event_ms": 0.5095476277669271, "speedup": 0.8553946381524506}
{"repetition": 5, "order": ["triton", "sdpa"], "seq_len": 4096, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.437459716796875, "triton_event_ms": 0.5086007690429688, "speedup": 0.8601239782237068}
{"repetition": 6, "order": ["sdpa", "triton"], "seq_len": 4096, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.43573099772135415, "triton_event_ms": 0.5084250895182292, "speedup": 0.8570210375224439}
{"repetition": 7, "order": ["triton", "sdpa"], "seq_len": 4096, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.43709879557291664, "triton_event_ms": 0.5099888102213541, "speedup": 0.8570752667753622}
{"repetition": 8, "order": ["sdpa", "triton"], "seq_len": 4096, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.4362843831380208, "triton_event_ms": 0.5094790649414063, "speedup": 0.85633426996298}
{"repetition": 9, "order": ["triton", "sdpa"], "seq_len": 4096, "batch": 2, "heads": 12, "head_dim": 64, "iterations": 300, "dtype": "float16", "gpu": "NVIDIA GeForce RTX 4090", "max_abs_error": 0.00048828125, "sdpa_event_ms": 0.43750539143880207, "triton_event_ms": 0.5085294087727864, "speedup": 0.860334493721054}
