{ "metadata": { "width": 1344, "height": 768, "frames": 124, "steps": 12, "sampler_step": 1, "seed": 440420, "text_tokens": 100, "block_index": 24, "attention": "sage2", "hidden_shape": [ 37810, 5376 ], "h_msa_shape": [ 37810, 5376 ], "segments": [ [ 0, 100, 1 ], [ 100, 514, 2 ], [ 514, 37810, 0 ] ] }, "q_shape": [ 1, 37810, 56, 128 ], "k_shape": [ 1, 37810, 56, 128 ], "v_shape": [ 1, 37810, 56, 128 ], "warmup": 3, "iterations": 10, "phase_timings": { "k_mean_and_smoothing_preparation": { "count": 10, "mean_s": 0.002420537567138672, "p50_s": 0.0023651360273361207, "p90_s": 0.0025749952793121334, "p95_s": 0.0026686816215515134, "p99_s": 0.0027436306953430175, "min_s": 0.002331295967102051, "max_s": 0.0027623679637908934 }, "qk_int8_quantization": { "count": 10, "mean_s": 0.007748214387893677, "p50_s": 0.0076711041927337645, "p90_s": 0.007954323053359986, "p95_s": 0.007966649413108826, "p99_s": 0.007976510500907898, "min_s": 0.007604640007019043, "max_s": 0.007978975772857666 }, "q_int8_quantization": { "count": 10, "mean_s": 0.0037990527868270877, "p50_s": 0.0037993919849395753, "p90_s": 0.0038029504060745237, "p95_s": 0.0038038431525230405, "p99_s": 0.003804557349681854, "min_s": 0.003791712045669556, "max_s": 0.0038047358989715574 }, "k_int8_subtract_mean_quantization": { "count": 10, "mean_s": 0.0038265248060226436, "p50_s": 0.0038259359598159792, "p90_s": 0.0038348320722579954, "p95_s": 0.0038351920247077943, "p99_s": 0.003835479986667633, "min_s": 0.003819711923599243, "max_s": 0.0038355519771575927 }, "v_fp8_transpose_scale_quantization": { "count": 10, "mean_s": 0.011513513278961182, "p50_s": 0.010691807746887208, "p90_s": 0.011705465221405026, "p95_s": 0.015195276212692253, "p99_s": 0.017987125005722047, "min_s": 0.010680031776428223, "max_s": 0.018685087203979493 }, "v_transpose_pad_permute": { "count": 10, "mean_s": 0.0051148447513580325, "p50_s": 0.005081151962280274, "p90_s": 0.005127699327468873, "p95_s": 0.005270201659202576, "p99_s": 0.005384203524589539, "min_s": 0.005073599815368653, "max_s": 0.00541270399093628 }, "v_scale_fp8_quantization": { "count": 10, "mean_s": 0.005551833629608155, "p50_s": 0.0055316162109375, "p90_s": 0.0055659454822540285, "p95_s": 0.005644108700752258, "p99_s": 0.005706639275550842, "min_s": 0.00552732801437378, "max_s": 0.005722271919250488 }, "fused_mainloop": { "count": 10, "mean_s": 0.23669681091308595, "p50_s": 0.23708939361572268, "p90_s": 0.23775543670654295, "p95_s": 0.2378476737976074, "p99_s": 0.23792146347045898, "min_s": 0.23529600524902344, "max_s": 0.23793991088867186 } }, "fused_mainloop_phases": { "int8_qk": "fused inside qk_int_sv_f8_attn_kernel", "scale_application": "fused inside qk_int_sv_f8_attn_kernel", "online_softmax": "fused inside qk_int_sv_f8_attn_kernel", "pv_accumulation": "fused inside qk_int_sv_f8_attn_kernel", "final_normalization_and_output": "fused inside qk_int_sv_f8_attn_kernel", "timing_policy": "Do not assign independent wall time without changing the exact kernel schedule; use source-correlated hardware counters." }, "scheduler": { "cta_q": 128, "cta_k": 64, "warp_q": 32, "warp_k": 64, "warps_per_cta": 4, "threads_per_cta": 128, "dynamic_shared_memory_bytes": 32768, "q_ctas_per_head": 296, "heads": 56, "grid_ctas": 16576, "k_iterations_per_cta": 591, "q_tail_rows": 50, "k_tail_rows": 50, "explicit_pipeline_stages": 2 }, "manual_decomposition_vs_public_sage2": { "equal": true, "max_abs": 0.0, "mean_abs": 0.0 }, "manual_checksum": 9190064.0, "reference_checksum": 9190064.0, "tail_study": [ { "sweep": "q_tail_fixed_kv", "q_len": 37760, "kv_len": 37810, "q_ctas_per_head": 295, "k_iterations_per_cta": 591, "q_tail_rows": 0, "k_tail_rows": 50, "scheduled_q_rows": 37760, "q_row_efficiency": 1.0, "mainloop": { "count": 5, "mean_s": 0.2392295654296875, "p50_s": 0.23901161193847656, "p90_s": 0.24024520874023436, "p95_s": 0.24033575439453123, "p99_s": 0.24040819091796872, "min_s": 0.23829190063476563, "max_s": 0.24042630004882812 }, "checksum": 9183940.0 }, { "sweep": "q_tail_fixed_kv", "q_len": 37761, "kv_len": 37810, "q_ctas_per_head": 296, "k_iterations_per_cta": 591, "q_tail_rows": 1, "k_tail_rows": 50, "scheduled_q_rows": 37888, "q_row_efficiency": 0.9966480152027027, "mainloop": { "count": 5, "mean_s": 0.23963267211914063, "p50_s": 0.23961180114746095, "p90_s": 0.23997588500976563, "p95_s": 0.24002579956054687, "p99_s": 0.2400657312011719, "min_s": 0.23915267944335938, "max_s": 0.24007571411132814 }, "checksum": 9183946.0 }, { "sweep": "q_tail_fixed_kv", "q_len": 37810, "kv_len": 37810, "q_ctas_per_head": 296, "k_iterations_per_cta": 591, "q_tail_rows": 50, "k_tail_rows": 50, "scheduled_q_rows": 37888, "q_row_efficiency": 0.9979413006756757, "mainloop": { "count": 5, "mean_s": 0.23999984436035157, "p50_s": 0.23978463745117187, "p90_s": 0.24042403869628906, "p95_s": 0.24055045166015623, "p99_s": 0.24065158203124998, "min_s": 0.2397151641845703, "max_s": 0.24067686462402343 }, "checksum": 9190064.0 }, { "sweep": "kv_tail_fixed_q", "q_len": 37810, "kv_len": 37760, "q_ctas_per_head": 296, "k_iterations_per_cta": 590, "q_tail_rows": 50, "k_tail_rows": 0, "scheduled_q_rows": 37888, "q_row_efficiency": 0.9979413006756757, "mainloop": { "count": 5, "mean_s": 0.23982306518554686, "p50_s": 0.23983763122558593, "p90_s": 0.2401393035888672, "p95_s": 0.24022885437011718, "p99_s": 0.2403004949951172, "min_s": 0.2395267791748047, "max_s": 0.24031840515136718 }, "checksum": 9175197.0 }, { "sweep": "kv_tail_fixed_q", "q_len": 37810, "kv_len": 37761, "q_ctas_per_head": 296, "k_iterations_per_cta": 591, "q_tail_rows": 50, "k_tail_rows": 1, "scheduled_q_rows": 37888, "q_row_efficiency": 0.9979413006756757, "mainloop": { "count": 5, "mean_s": 0.2405945220947266, "p50_s": 0.24039689636230469, "p90_s": 0.24119658203125, "p95_s": 0.24143297729492186, "p99_s": 0.24162209350585936, "min_s": 0.2401432342529297, "max_s": 0.24166937255859375 }, "checksum": 9176321.0 }, { "sweep": "kv_tail_fixed_q", "q_len": 37810, "kv_len": 37810, "q_ctas_per_head": 296, "k_iterations_per_cta": 591, "q_tail_rows": 50, "k_tail_rows": 50, "scheduled_q_rows": 37888, "q_row_efficiency": 0.9979413006756757, "mainloop": { "count": 5, "mean_s": 0.24062385864257813, "p50_s": 0.24038169860839845, "p90_s": 0.24108528442382812, "p95_s": 0.2411941619873047, "p99_s": 0.24128126403808592, "min_s": 0.2403317108154297, "max_s": 0.24130303955078125 }, "checksum": 9190064.0 } ] }