From 66991e5d5490f5092e142e82a97fa7496bcaad96 Mon Sep 17 00:00:00 2001
From: "github-actions[bot]"
 <41898282+github-actions[bot]@users.noreply.github.com>
Date: Wed, 30 Sep 2026 12:59:40 +0000
Subject: [PATCH] ci: refresh .test_durations (ops test set changed: +52/-0)

---
 .github/.test_durations | 5706 ++++++++++++++++++++-------------------
 1 file changed, 2879 insertions(+), 2827 deletions(-)

diff --git a/.github/.test_durations b/.github/.test_durations
index f9af879e..3cf23095 100644
--- a/.github/.test_durations
+++ b/.github/.test_durations
@@ -1,2829 +1,2881 @@
 {
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-0-torch.float16-none]": 0.6053893400000163,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-0-torch.float16-tanh]": 0.5217558470000085,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-0-torch.float32-none]": 0.5697794060000092,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-0-torch.float32-tanh]": 0.4884153109999829,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-1-torch.float16-none]": 0.5851004509999882,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-1-torch.float16-tanh]": 0.5272629639999877,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-1-torch.float32-none]": 0.4900146950000135,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-1-torch.float32-tanh]": 0.5190763009999841,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-0-torch.float16-none]": 0.5113377390000267,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-0-torch.float16-tanh]": 0.5095946259999664,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-0-torch.float32-none]": 0.48080111199999465,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-0-torch.float32-tanh]": 0.5045792950000134,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-1-torch.float16-none]": 0.5875845080000204,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-1-torch.float16-tanh]": 0.5364971090000097,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-1-torch.float32-none]": 0.5551290349999931,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-1-torch.float32-tanh]": 0.5276502820000246,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-2-torch.float16-none]": 0.13090156800001296,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-2-torch.float16-tanh]": 0.1513705680000328,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-2-torch.float32-none]": 0.19814306200001397,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-2-torch.float32-tanh]": 0.15289546799996856,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-3-torch.float16-none]": 0.1369768449999924,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-3-torch.float16-tanh]": 0.15586383999999498,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-3-torch.float32-none]": 0.15484163899998293,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-3-torch.float32-tanh]": 0.16293802900000287,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(4,)-0-torch.float16-none]": 0.5291022429999828,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(4,)-0-torch.float16-tanh]": 0.48907463100002246,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(4,)-0-torch.float32-none]": 1.3312764530000436,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(4,)-0-torch.float32-tanh]": 0.48822427999996876,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-0-torch.float16-none]": 0.49782704099999364,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-0-torch.float16-tanh]": 0.5265960219999783,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-0-torch.float32-none]": 0.5762739099999692,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-0-torch.float32-tanh]": 0.5633892890000141,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-1-torch.float16-none]": 0.5730634530000032,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-1-torch.float16-tanh]": 0.5195544579999876,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-1-torch.float32-none]": 0.585620667000029,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-1-torch.float32-tanh]": 0.528796916000033,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-2-torch.float16-none]": 0.1481872089999854,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-2-torch.float16-tanh]": 0.14009204600000658,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-2-torch.float32-none]": 0.17472349199999826,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-2-torch.float32-tanh]": 0.13782638099999645,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-0-torch.float16-none]": 1.3695475310000234,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-0-torch.float16-tanh]": 1.2718758800000103,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-0-torch.float32-none]": 1.9200039150000237,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-0-torch.float32-tanh]": 1.3164392250000105,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-1-torch.float16-none]": 0.15638716599997338,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-1-torch.float16-tanh]": 0.16440650800004164,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-1-torch.float32-none]": 0.182130098999977,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-1-torch.float32-tanh]": 0.15466278599998873,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-0-torch.float16-none]": 0.15841635100002804,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-0-torch.float16-tanh]": 0.13055932299999995,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-0-torch.float32-none]": 0.13191395100000136,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-0-torch.float32-tanh]": 0.1451212739999619,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-1-torch.float16-none]": 0.15345455700000343,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-1-torch.float16-tanh]": 0.11921909499997696,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-1-torch.float32-none]": 0.1529555569999843,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-1-torch.float32-tanh]": 0.14497002000001658,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-2-torch.float16-none]": 0.1568328200000053,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-2-torch.float16-tanh]": 0.1298434529999497,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-2-torch.float32-none]": 0.14802270000001272,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-2-torch.float32-tanh]": 0.15729278099999533,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-3-torch.float16-none]": 0.1344588200000203,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-3-torch.float16-tanh]": 0.15868141800001467,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-3-torch.float32-none]": 0.1437860050000097,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-3-torch.float32-tanh]": 0.16478507599998693,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(4,)-0-torch.float16-none]": 0.15141545799997402,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(4,)-0-torch.float16-tanh]": 0.12630204900000308,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(4,)-0-torch.float32-none]": 0.15569636399999354,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(4,)-0-torch.float32-tanh]": 0.13952790599998366,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-0-torch.float16-none]": 0.15886676400000965,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-0-torch.float16-tanh]": 0.143185729999999,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-0-torch.float32-none]": 0.14544333099999562,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-0-torch.float32-tanh]": 0.1375360119999982,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-1-torch.float16-none]": 0.14287017999998852,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-1-torch.float16-tanh]": 0.15629679600002078,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-1-torch.float32-none]": 0.15022380300001714,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-1-torch.float32-tanh]": 0.16042277100004299,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-2-torch.float16-none]": 0.1523508929999764,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-2-torch.float16-tanh]": 0.1515706999999793,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-2-torch.float32-none]": 0.14884779300001583,
-"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-2-torch.float32-tanh]": 0.15090108699996563,
-"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[cutile-256-2048-none-torch.float16]": 0.29477589100000046,
-"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[cutile-256-2048-none-torch.float32]": 0.35649147500001277,
-"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[cutile-256-2048-tanh-torch.float16]": 0.27562762100001237,
-"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[cutile-256-2048-tanh-torch.float32]": 0.30212622400003397,
-"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[tilecpp-256-2048-none-torch.float16]": 0.7208230049999997,
-"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[tilecpp-256-2048-none-torch.float32]": 0.6941223689999845,
-"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[tilecpp-256-2048-tanh-torch.float16]": 0.7051868750000096,
-"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[tilecpp-256-2048-tanh-torch.float32]": 0.6978630599999747,
-"tests/ops/activation/test_relu.py::Test_ReLU::test_op[cutile-256-2048-torch.float32]": 0.41355173499999864,
-"tests/ops/activation/test_relu.py::Test_ReLU::test_op[cutile-256-256-torch.float16]": 0.41637999099998524,
-"tests/ops/activation/test_relu.py::Test_ReLU::test_op[cutile-256-32768-torch.float16]": 0.49259253699997885,
-"tests/ops/activation/test_relu.py::Test_ReLU::test_op[cutile-256-64-torch.float16]": 1.1168847789999745,
-"tests/ops/activation/test_relu.py::Test_ReLU::test_op[tilecpp-256-2048-torch.float32]": 0.7487660670000196,
-"tests/ops/activation/test_relu.py::Test_ReLU::test_op[tilecpp-256-256-torch.float16]": 0.3167616729999736,
-"tests/ops/activation/test_relu.py::Test_ReLU::test_op[tilecpp-256-32768-torch.float16]": 0.337343001000022,
-"tests/ops/activation/test_relu.py::Test_ReLU::test_op[tilecpp-256-64-torch.float16]": 0.8585800490000111,
-"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_backward_corner_cases[cutile-1-1-64-64-True-torch.float16]": 14.23651766100005,
-"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_backward_corner_cases[cutile-1-32-128-64-True-torch.float16]": 14.291485359000006,
-"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_backward_irregular_shapes[cutile-1-4-127-64-True-torch.float16]": 14.430805042000003,
-"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_backward_irregular_shapes[cutile-2-8-2047-128-True-torch.bfloat16]": 11.848035576000001,
-"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_backward_regular_shapes[cutile-2-8-256-64-True-torch.float16]": 14.080979548000045,
-"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_backward_regular_shapes[cutile-2-8-512-128-False-torch.bfloat16]": 10.687304627999964,
-"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_forward_with_lse_matches_reference[cutile-1-8-256-128-True-torch.float16]": 1.4178816179999956,
-"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_forward_with_lse_matches_reference[cutile-2-8-512-64-False-torch.bfloat16]": 0.843168841999983,
-"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_functional_api_inference_mode[cutile]": 1.70386251299999,
-"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_functional_api_training_mode[cutile]": 14.156464977000013,
-"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward_GQA::test_op_backward_gqa[cutile-1-8-1-128-64-True-torch.float16]": 14.83967427999994,
-"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward_GQA::test_op_backward_gqa[cutile-1-8-4-128-64-True-torch.float16]": 14.582536329999982,
-"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward_Numerical::test_op_numerical_gradient_check[cutile]": 0.19474330100007364,
-"tests/ops/experimental/test_fused_linear_cross_entropy.py::TestFusedLinearCrossEntropy::test_op_chunk_size_consistency[cutile]": 0.7825574620000566,
-"tests/ops/experimental/test_fused_linear_cross_entropy.py::TestFusedLinearCrossEntropy::test_op_forward_matches_pytorch[cutile-1-256-384-4096-torch.float16-sum]": 0.3961107460000335,
-"tests/ops/experimental/test_fused_linear_cross_entropy.py::TestFusedLinearCrossEntropy::test_op_forward_matches_pytorch[cutile-2-128-256-2048-torch.float16-mean]": 0.6678163599999607,
-"tests/ops/experimental/test_fused_linear_cross_entropy.py::TestFusedLinearCrossEntropy::test_op_forward_matches_pytorch[cutile-2-64-256-2048-torch.bfloat16-mean]": 0.5628469829999858,
-"tests/ops/experimental/test_fused_linear_cross_entropy.py::TestFusedLinearCrossEntropy::test_op_peak_memory_less_than_pytorch[cutile]": 1.2144154339999886,
-"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op[cutile-torch.float32-2-512-1024-4096]": 3.175445532000026,
-"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op[cutile-torch.float32-2-64-256-512]": 2.0277287100000194,
-"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op[cutile-torch.float32-4-128-512-1024]": 2.601571251999985,
-"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op[cutile-torch.float32-8-256-512-2048]": 2.6691556300000343,
-"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-11-127-750-1500]": 0.6332476280000492,
-"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-13-128-512-1024]": 0.14578609700004108,
-"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-13-333-500-1000]": 0.6153365129999884,
-"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-4-100-256-512]": 0.13399814500002094,
-"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-4-128-300-600]": 3.092563860999917,
-"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-4-128-500-1000]": 3.0359421699999416,
-"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-4-128-750-1500]": 3.203327553000065,
-"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-4-333-512-1024]": 0.4911071189999916,
-"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-7-100-300-600]": 0.6098597609999956,
-"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-7-64-256-512]": 0.17596039799997243,
-"tests/ops/experimental/test_mhc.py::Test_MHC::test_op_mhc_apply_residual[cutile-128-2-1024]": 0.37365006899997866,
-"tests/ops/experimental/test_mhc.py::Test_MHC::test_op_mhc_apply_residual[cutile-128-4-1024]": 0.46702334600001905,
-"tests/ops/experimental/test_mhc.py::Test_MHC::test_op_mhc_apply_residual[cutile-64-8-2048]": 0.4025114750000114,
-"tests/ops/experimental/test_mhc.py::Test_MHC::test_op_mhc_gemm_rms_scale_bf16_precision[cutile-128-1000-8]": 22.003300546000048,
-"tests/ops/experimental/test_mhc.py::Test_MHC::test_op_mhc_gemm_rms_scale_bf16_precision[cutile-128-1024-4]": 22.04024993600001,
-"tests/ops/experimental/test_mhc.py::Test_MHC::test_op_mhc_gemm_rms_scale_bf16_precision[cutile-128-1024-8]": 22.057073146999983,
-"tests/ops/experimental/test_mhc.py::Test_MHC::test_op_mhc_sinkhorn[cutile-256-4-torch.float32]": 0.3531446859999505,
-"tests/ops/experimental/test_mhc.py::Test_MHC::test_op_mhc_sinkhorn[cutile-256-8-torch.float32]": 1.1213951590000022,
-"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_input_dtypes[cutile-torch.bfloat16]": 0.00034673999999768057,
-"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_input_dtypes[cutile-torch.float16]": 0.00036934300001689735,
-"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_input_dtypes[cutile-torch.float32]": 0.0003653859999701581,
-"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_invalid_inputs[cutile]": 0.000394014999983483,
-"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-1-64]": 0.00035464800001250296,
-"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-128-64]": 0.0015886700000180554,
-"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-128-80]": 0.0004480249999971875,
-"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-160-80]": 0.00037714500001584383,
-"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-17-64]": 0.000384698999965849,
-"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-192-64]": 0.0004533419999859234,
-"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-256-48]": 0.00047013299999321134,
-"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-33-48]": 0.0007257580000157304,
-"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-512-256]": 0.0005753579999918657,
-"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-64-64]": 0.00047972900000559093,
-"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-None-256-256-torch.float16]": 0.3227062299999943,
-"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-None-256-256-torch.float32]": 0.29871442600000364,
-"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-None-4096-256-torch.bfloat16]": 0.41247341700000106,
-"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-multi_wave_cached-256-256-torch.float16]": 0.270056980000021,
-"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-multi_wave_cached-256-256-torch.float32]": 0.31469059899998797,
-"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-multi_wave_cached-4096-256-torch.bfloat16]": 0.3079473750000261,
-"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-multi_wave_reload-256-256-torch.float16]": 0.1830220050000264,
-"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-multi_wave_reload-256-256-torch.float32]": 0.1866813860000036,
-"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-multi_wave_reload-4096-256-torch.bfloat16]": 0.27457359400003156,
-"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-static_persistent-256-256-torch.float16]": 0.4125913969999999,
-"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-static_persistent-256-256-torch.float32]": 1.1573738550000314,
-"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-static_persistent-4096-256-torch.bfloat16]": 0.18547258800001032,
-"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormBackward::test_op[cutile-2003-2001-torch.float16]": 1.2192593559999807,
-"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormBackward::test_op[cutile-256-256-torch.float16]": 1.5960999100000208,
-"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormBackward::test_op[cutile-256-256-torch.float32]": 0.44227652899999725,
-"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormBackward::test_op[cutile-31072-4096-torch.bfloat16]": 5.277308180000006,
-"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormBackward::test_op[cutile-4096-256-torch.bfloat16]": 1.4690393539999889,
-"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_basic[cutile-torch.bfloat16-1-16-64-128-128-64-1-64]": 23.987745812000014,
-"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_basic[cutile-torch.bfloat16-1-64-128-256-128-64-1-256]": 72.71789477399997,
-"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_basic[cutile-torch.bfloat16-2-32-128-256-128-64-1-128]": 54.19948703700001,
-"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_gqa[cutile-torch.bfloat16-1-16-64-128-128-64-4-64]": 11.969101440999964,
-"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_gqa[cutile-torch.bfloat16-1-32-64-128-128-64-8-64]": 11.99379472000004,
-"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_invalid_kernel_configs[cutile]": 0.17762501099991823,
-"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_irregular_shapes[cutile-torch.bfloat16-1-16-1-128-128-64-1-64]": 23.37311749600002,
-"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_irregular_shapes[cutile-torch.bfloat16-1-16-33-77-128-64-1-64]": 23.572224214000016,
-"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_tile_h_configs[cutile-torch.bfloat16-16-1-16-32-TILE_H with small TILE_N]": 0.163957962999973,
-"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_tile_h_configs[cutile-torch.bfloat16-16-1-4-64-small TILE_H]": 0.21193477500003155,
-"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_tile_h_configs[cutile-torch.bfloat16-16-16-1-64-1:1 mapping fallback]": 0.8605218060000084,
-"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_tile_h_configs[cutile-torch.bfloat16-16-4-4-64-GQA group_size=4]": 0.20039560100002518,
-"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_tile_h_configs[cutile-torch.bfloat16-18-6-1-64-odd prime group_size=3 fallback]": 0.8395221980000542,
-"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_tile_h_configs[cutile-torch.bfloat16-24-4-2-64-non_pow2 group_size=6, TILE_H=2]": 0.8913572380000119,
-"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_tile_h_configs[cutile-torch.bfloat16-24-8-1-64-non_pow2 group_size=3 fallback]": 0.8122334290000026,
-"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_tile_h_configs[cutile-torch.bfloat16-64-1-16-64-large TILE_H]": 0.19307529399992518,
-"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_topk_equals_skv[cutile-torch.bfloat16-1-16-64-128-64-1]": 0.16808514699994248,
-"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_first_tokens[cutile]": 0.15300257099994496,
-"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_gqa[cutile]": 0.14020915700001524,
-"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_long_context_mistral[cutile]": 0.1464283339999497,
-"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_multi_head[cutile]": 0.14853086900001244,
-"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_seq_not_divisible_by_tile[cutile]": 0.12686812300006522,
-"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_small_window[cutile]": 0.17203821200001812,
-"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_various_configs[cutile-1024-512]": 0.12733472200005735,
-"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_various_configs[cutile-2048-1024]": 0.13733208999991575,
-"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_various_configs[cutile-4096-2048]": 0.15960417999991705,
-"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_various_configs[cutile-4096-4096]": 0.15222652100004552,
-"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_various_configs[cutile-512-256]": 0.12223583100001179,
-"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_very_small_seq[cutile]": 0.15322240099999362,
-"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_window_equals_seq[cutile]": 1.0128446389998658,
-"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_window_equals_tile[cutile]": 0.14546149500000638,
-"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_window_of_one[cutile]": 0.15153256199999987,
-"tests/ops/test_attention.py::Test_FMHA::test_op[cutile-1-1-9-128-False-torch.bfloat16]": 0.004221206000124766,
-"tests/ops/test_attention.py::Test_FMHA::test_op[cutile-1-32-2047-128-True-torch.float16]": 0.0032436640000810257,
-"tests/ops/test_attention.py::Test_FMHA::test_op[cutile-2-32-4095-128-True-torch.bfloat16]": 0.0009857249999640771,
-"tests/ops/test_attention.py::Test_FMHA::test_op[cutile-2-32-4095-128-True-torch.float8_e5m2]": 0.00111708299999691,
-"tests/ops/test_attention.py::Test_FMHA::test_op[tilecpp-1-1-9-128-False-torch.bfloat16]": 0.0013015129999871533,
-"tests/ops/test_attention.py::Test_FMHA::test_op[tilecpp-1-32-2047-128-True-torch.float16]": 0.0010286569999493622,
-"tests/ops/test_attention.py::Test_FMHA::test_op[tilecpp-2-32-4095-128-True-torch.bfloat16]": 0.0009819090000178221,
-"tests/ops/test_attention.py::Test_FMHA::test_op[tilecpp-2-32-4095-128-True-torch.float8_e5m2]": 0.0011111019999816563,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-128-0.125-64-8-8-128-1-1]": 27.348949671000014,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-128-0.125-64-8-8-128-1-2]": 26.96431093399997,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-128-0.125-64-8-8-128-128-1]": 20.408174393000024,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-128-0.125-64-8-8-128-128-2]": 1.1380520039999737,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-128-0.125-64-8-8-32-1-1]": 23.83852512599998,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-128-0.125-64-8-8-32-1-2]": 23.619622453000034,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-128-0.125-64-8-8-32-128-1]": 0.0011951109999586151,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-128-0.125-64-8-8-32-128-2]": 0.0012297370000737828,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-None-0.125-64-8-8-128-1-1]": 1.012083682000025,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-None-0.125-64-8-8-128-1-2]": 5.632853217000047,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-None-0.125-64-8-8-128-128-1]": 1.2442101940000043,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-None-0.125-64-8-8-128-128-2]": 18.965424010999982,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-None-0.125-64-8-8-32-1-1]": 22.763951412999972,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-None-0.125-64-8-8-32-1-2]": 22.60411602299999,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-None-0.125-64-8-8-32-128-1]": 0.0011114230000259795,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-None-0.125-64-8-8-32-128-2]": 0.0034237760000053186,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-128-0.125-64-8-8-128-1-1]": 0.136886469999979,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-128-0.125-64-8-8-128-1-2]": 0.1331104299999879,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-128-0.125-64-8-8-128-128-1]": 0.14555783199995176,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-128-0.125-64-8-8-128-128-2]": 0.13650307299997166,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-128-0.125-64-8-8-32-1-1]": 0.11658188499995958,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-128-0.125-64-8-8-32-1-2]": 0.1350320189999934,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-128-0.125-64-8-8-32-128-1]": 0.0012107629999604796,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-128-0.125-64-8-8-32-128-2]": 0.0010797049999951014,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-None-0.125-64-8-8-128-1-1]": 25.24413892199999,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-None-0.125-64-8-8-128-1-2]": 24.936055907999958,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-None-0.125-64-8-8-128-128-1]": 1.319809111999973,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-None-0.125-64-8-8-128-128-2]": 0.17510492800005295,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-None-0.125-64-8-8-32-1-1]": 0.14420343400007596,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-None-0.125-64-8-8-32-1-2]": 0.17893401700001732,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-None-0.125-64-8-8-32-128-1]": 0.0035821970000142755,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-None-0.125-64-8-8-32-128-2]": 0.002543734999960634,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-128-0.125-64-8-8-128-1-1]": 3.394933610999999,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-128-0.125-64-8-8-128-1-2]": 0.1747254450000355,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-128-0.125-64-8-8-128-128-1]": 0.14453283799997507,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-128-0.125-64-8-8-128-128-2]": 0.14144523599998138,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-128-0.125-64-8-8-32-1-1]": 0.163625106999973,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-128-0.125-64-8-8-32-1-2]": 0.16512914700001602,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-128-0.125-64-8-8-32-128-1]": 0.0014828639999677762,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-128-0.125-64-8-8-32-128-2]": 0.0011047170000324513,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-None-0.125-64-8-8-128-1-1]": 3.4024638259999733,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-None-0.125-64-8-8-128-1-2]": 0.174875596999982,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-None-0.125-64-8-8-128-128-1]": 0.15278321400001005,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-None-0.125-64-8-8-128-128-2]": 0.1453694570000721,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-None-0.125-64-8-8-32-1-1]": 0.13868214400002898,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-None-0.125-64-8-8-32-1-2]": 0.13999234599998545,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-None-0.125-64-8-8-32-128-1]": 0.0014033180000296852,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-None-0.125-64-8-8-32-128-2]": 0.0013720219999981964,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-128-0.125-64-8-8-128-1-1]": 0.16732041999995317,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-128-0.125-64-8-8-128-1-2]": 0.1407116359999918,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-128-0.125-64-8-8-128-128-1]": 0.16617098899996563,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-128-0.125-64-8-8-128-128-2]": 0.16272272100002283,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-128-0.125-64-8-8-32-1-1]": 0.14194329100007508,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-128-0.125-64-8-8-32-1-2]": 0.16295605700008764,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-128-0.125-64-8-8-32-128-1]": 0.0014254900000310045,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-128-0.125-64-8-8-32-128-2]": 0.001675059000035617,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-None-0.125-64-8-8-128-1-1]": 0.1538389490000327,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-None-0.125-64-8-8-128-1-2]": 0.13430227899999636,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-None-0.125-64-8-8-128-128-1]": 0.125279259000024,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-None-0.125-64-8-8-128-128-2]": 0.16857088199992631,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-None-0.125-64-8-8-32-1-1]": 0.1544212460000267,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-None-0.125-64-8-8-32-1-2]": 0.1802905389999978,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-None-0.125-64-8-8-32-128-1]": 0.0015040030000363913,
-"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-None-0.125-64-8-8-32-128-2]": 0.004789891000029911,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0012249860000110857,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.001277080000022579,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0016846270000030472,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0019856039999979203,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0021563799999739786,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0011818250000033004,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0012106460000325114,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0013097419999894555,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.0011458270000446191,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0013169810000022153,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0012522899999680703,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.002299813000007589,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.0011747180000440949,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0011767229999861684,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.001363480999998501,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.001243822999981603,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0012403679999977157,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0011249840000004951,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0012735560000010082,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.001344881999983727,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0012398700000346707,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0011874509999643124,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0011790859999791792,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0012679510000168648,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0012055659999816726,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0011971210000183419,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0013444040000081259,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0011981799999603027,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0012311420000230555,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0012075269999911598,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0011493140000027324,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.001324291000003086,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.0012184349999984079,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0012059040000167442,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0012031499999807238,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0013351089999957821,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.001277568000006113,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.001251026000005595,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.001227012999947874,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.001226530999986153,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0012258339999959844,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0012358350000170049,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0012469540000097368,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.001475760000033688,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0013604029999783052,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.001242395999980772,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0011876800000152343,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0012308840000514465,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0012528970000289519,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0012212540000007266,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.001339122999979736,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0012073649999706504,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0012260120000178176,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0012227849999817408,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0011998810000193316,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0013570360000301207,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.001210997000015368,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0012253280000038558,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.001199096000021882,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.001944166000015457,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.0012342219999936788,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0012214840000126514,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0014047550000100273,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.001198683999973582,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0011992549999888524,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0012142960000005587,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.001214010000012422,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.001406304000028058,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0012130349999779355,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.001221298999979581,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.001198805000029779,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0013644909999754873,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0013793170000155897,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0011990190000119583,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0013442689999862978,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0012212009999927886,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0012554539999598546,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0012476900000137903,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0012588540000137982,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0016444079999757832,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.001416295000012724,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.001313594999999168,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0012202260000151455,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0013515499999812164,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.001231043000018417,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0013289679999957116,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0015282319999982974,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0012241829999766196,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0012223170000140726,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0012533620000283463,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0012080859999628046,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0011357249999832675,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0012853540000321573,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0011010989999817866,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0012492049999934807,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0012474909999866668,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0012559500000008939,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.001036084999952891,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0011161980000053973,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0010372720000191293,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0010293269999976928,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0010430540000072597,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0010727540000345925,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0013458040000102756,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.0011964139999918189,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0011605280000139828,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0010281810000094538,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0015519880000169906,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.0011839270000280067,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.001144477999986293,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0013114120000068397,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0012018220000129531,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.00116403099997342,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0014092209999887473,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0012029420000203572,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0013497120000067753,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.001224745999991228,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0011756369999886829,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0011748660000137079,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.001298219000005929,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0012505879999764602,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0011890050000147312,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0013307489999760946,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0012004660000286549,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0012058000000081392,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.001218644999994467,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0012199050000560874,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0013535969999907138,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.001225003999962837,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0012116160000061882,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0012079359999859207,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0013190029999918806,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.001264886999990722,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0014282919999573096,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0012538299999960145,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0013301449999971737,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.001231585999988738,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0012268459999802417,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0012558369999737806,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0012420429999906446,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0014398410000069362,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0012451990000101887,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0013544710000417126,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0013642410000045402,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0012298100000123213,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0011461889999964114,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.001300023000027295,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.001172379999985651,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0012046779999934643,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0012292600000023413,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0011692150000044421,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.001346520999987888,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.0011490859999696568,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0011472369999978582,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0011283709999645453,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.001768787000003158,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.0011039170000231024,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0011028870000302504,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0013834559999850171,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0010420709999721112,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0010397099999863713,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0011849850000089646,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.001062157000035313,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0011484840000264285,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.001105675999980349,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0010810769999807235,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.001266018999984908,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0012927930000046217,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0011360709999905794,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0010648870000125044,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0011906509999732862,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0010772709999855579,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0010953529999540024,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.001075717999981407,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0011325430000113101,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.001109232999993992,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.001164309999978741,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.001208780000013121,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0010895070000458418,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.001171815999981618,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.0017541139999934785,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0011730309999791189,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0013095369999973627,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0011435040000264962,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0012369309999940015,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0012739649999957692,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0011637209999832976,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.001363667000020996,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0013343809999923906,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.001916794999999638,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0012037510000197926,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0011656379999749333,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.002706841000019722,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0010770069999921361,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0015104009999902246,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0011732410000320215,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0012192929999912394,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0011735149999765326,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0012410780000209343,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0015852309999786485,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.0015747199999793793,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.001570309999976871,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0011269440000205577,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.002075996999963081,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.0015982839999821863,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0017293190000202685,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0017831440000293242,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0016272349999724156,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0016002960000207622,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0016220539999949324,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0015812400000072557,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.001906456000028811,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.001593470999978308,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0016035229999715739,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.001598571000016591,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0018173630000148933,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.00163935699998774,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0016003230000194435,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0017877959999736959,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0015920700000151555,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0016101629999809575,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0015795660000037515,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.001630679000015789,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0017995719999817084,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.0016328730000054748,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0016460620000202653,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0017752999999913754,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.00178159099996833,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.0012657390000185842,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0017241619999879276,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0017402900000149657,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.001830498999993324,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0012359199999991688,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0016495679999479762,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0012185070000043652,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0012326169999710146,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0014197110000395696,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0013187829999878886,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.001678078000026062,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0016238230000169551,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0011894969999843852,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0015677420000201892,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.002670698000002858,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0028020279999907416,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0019111729999963245,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0012479589999827567,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.001202691000003142,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0013178989999858004,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.001177890999997544,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0012906999999984237,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0013581149999879472,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0021669919999851572,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.001341607000000522,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0012054139999975177,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0013934889999802635,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0013573570000176005,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0012620159999983116,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0012019029999805753,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0013845309999851452,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0014741100000321694,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.001174297999995133,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.00118904300001077,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0012987290000125995,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0012858960000130537,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0011945429999968837,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.001191157999983261,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0013498099999935675,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0012582480000276064,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0013959880000129488,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0012379089999967618,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0012308190000283048,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0015072800000268671,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.0011330950000001394,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.001137773999971614,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.00119684399999187,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0013671460000068691,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.0010883520000106728,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.001162041999975827,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0011869310000065525,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0011274160000027678,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0012082069999905798,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0012856719999945199,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0011206160000085674,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.001101029000011522,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0010731810000095265,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.00118261000000075,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0011600550000139265,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.001278530999996974,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.001480847000010499,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0010745730000110143,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.001156559000037305,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0012291280000056304,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.00125234300000443,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.001053498000004538,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0012609799999836468,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0015937420000113889,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.0011767619999716317,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0012309389999813902,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0010501730000100906,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0016672710000307234,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.0012380980000159525,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0013386299999922358,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0015080080000302587,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.001189137999972445,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0012227189999691745,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0014159850000226015,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0012585609999860026,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0013129159999891726,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0012240830000109781,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0011865529999681712,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0012229850000267106,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0014623030000393555,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.001249970999992911,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0012042920000112645,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0013749140000527404,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0012117609999791057,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0012044270000046708,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.001247704999968846,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0011968219999971552,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0013673199999857388,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.0013108949999889319,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0012481640000032712,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.001219470999984651,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.001351051000000325,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.001250247999990961,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0012046220000172525,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0012531719999913093,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0014473420000058468,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0012378760000046896,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0012432970000304522,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0012227859999711654,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0012530189999893082,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0011468680000064069,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0011659810000139714,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0012214700000185985,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0012103340000066964,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0014420949999873756,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0011911010000176248,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0013814739999702397,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0014417399999615554,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0013057820000028642,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.00119236099999398,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0016498999999896569,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.00157607000002713,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.001406660999975884,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.00138720000001058,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0011970849999727307,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0018694779999748334,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.0016207070000007207,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0015245310000011614,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0016891880000002857,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.001469799000005878,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0015042349999987437,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.001441154000019651,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0014791250000030232,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0016281220000280427,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0014964550000229337,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0015573129999779667,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0014506400000016129,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.001523161999983813,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.001216443000032541,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0016516119999891998,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.001358089999968115,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0012325490000080208,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0013899319999950421,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0014783100000101967,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0012021709999885388,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0013732900000036352,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.0012160789999882127,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.001247311000014406,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.001534810999999081,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0016276560000392237,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.0012562629999877117,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0012466990000064015,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0013157750000232227,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0012737210000466348,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0011984070000323754,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0014828390000047875,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.001213253999992503,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0012749879999773839,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0012500220000219997,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.001285789000036175,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0012898599999857652,
-"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0012362549999807015,
-"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-128-2064-128-False-torch.bfloat16-False]": 0.0011067229999639494,
-"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-128-2064-128-False-torch.bfloat16-True]": 0.0010529810000434736,
-"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-128-2064-128-True-torch.bfloat16-False]": 0.0010380090000410291,
-"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-128-2064-128-True-torch.bfloat16-True]": 0.001290371000010282,
-"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-128-496-128-False-torch.bfloat16-False]": 0.0010742119999918032,
-"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-128-496-128-False-torch.bfloat16-True]": 0.0010844230000088828,
-"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-128-496-128-True-torch.bfloat16-False]": 0.0010506309999982477,
-"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-128-496-128-True-torch.bfloat16-True]": 0.0010603520000245226,
-"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-257-2064-128-False-torch.bfloat16-False]": 0.0011151609999728862,
-"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-257-2064-128-False-torch.bfloat16-True]": 0.0011237330000142265,
-"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-257-2064-128-True-torch.bfloat16-False]": 0.001074563999992506,
-"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-257-2064-128-True-torch.bfloat16-True]": 0.00114174799998068,
-"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-257-496-128-False-torch.bfloat16-False]": 0.0010356679999858898,
-"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-257-496-128-False-torch.bfloat16-True]": 0.001038116000017908,
-"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-257-496-128-True-torch.bfloat16-False]": 0.0010587640000210286,
-"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-257-496-128-True-torch.bfloat16-True]": 0.0010558570000114287,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-1023-False-False-torch.float16]": 0.14352786399999218,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-1023-False-True-torch.float16]": 0.13792577999998912,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-1023-True-False-torch.float16]": 0.13160875599999144,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-1023-True-True-torch.float16]": 0.1291036629999951,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-1024-False-False-torch.float16]": 1.5504859629999999,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-1024-False-True-torch.float16]": 0.17888685799999848,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-1024-True-False-torch.float16]": 0.17936500999999794,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-1024-True-True-torch.float16]": 0.13670385400004648,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-511-False-False-torch.float16]": 0.13974770300001182,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-511-False-True-torch.float16]": 0.13276363099996047,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-511-True-False-torch.float16]": 0.13968390499999828,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-511-True-True-torch.float16]": 0.14015477399996712,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-512-False-False-torch.float16]": 0.1300619250000068,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-512-False-True-torch.float16]": 0.12800402500002406,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-512-True-False-torch.float16]": 0.1324981240000227,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-512-True-True-torch.float16]": 0.13686975599998163,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-1023-False-False-torch.float16]": 0.17366386499998043,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-1023-False-True-torch.float16]": 0.18035357200002977,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-1023-True-False-torch.float16]": 0.18021495200000004,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-1023-True-True-torch.float16]": 0.18195007500000315,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-1024-False-False-torch.float16]": 0.18540067900002555,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-1024-False-True-torch.float16]": 0.17270706699997618,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-1024-True-False-torch.float16]": 0.17785839000001147,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-1024-True-True-torch.float16]": 0.17578743400000008,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-511-False-False-torch.float16]": 0.2146395030000008,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-511-False-True-torch.float16]": 0.18322966799999563,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-511-True-False-torch.float16]": 0.18468860499996254,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-511-True-True-torch.float16]": 0.1964918559999944,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-512-False-False-torch.float16]": 0.17273569799996835,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-512-False-True-torch.float16]": 0.1783330809999768,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-512-True-False-torch.float16]": 0.16902746400000979,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-512-True-True-torch.float16]": 0.19645542599997157,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-1023-False-False-torch.float16]": 1.5848281750000126,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-1023-False-True-torch.float16]": 0.167801489999988,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-1023-True-False-torch.float16]": 0.17227979599996956,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-1023-True-True-torch.float16]": 0.17071512200001848,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-1024-False-False-torch.float16]": 0.20503201400001103,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-1024-False-True-torch.float16]": 0.18886806099999376,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-1024-True-False-torch.float16]": 0.17820494900001904,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-1024-True-True-torch.float16]": 0.19686035399999469,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-511-False-False-torch.float16]": 0.17286634900005993,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-511-False-True-torch.float16]": 0.1413663139999244,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-511-True-False-torch.float16]": 0.13930550199984282,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-511-True-True-torch.float16]": 0.19627024399983384,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-512-False-False-torch.float16]": 0.1753268620000199,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-512-False-True-torch.float16]": 0.13643812300028912,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-512-True-False-torch.float16]": 0.13015909199975795,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-512-True-True-torch.float16]": 0.11941991999992752,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-1023-False-False-torch.float16]": 0.17792729600003554,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-1023-False-True-torch.float16]": 0.18816490299997213,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-1023-True-False-torch.float16]": 0.17850339900002155,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-1023-True-True-torch.float16]": 0.18158796199995209,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-1024-False-False-torch.float16]": 0.14316911199998117,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-1024-False-True-torch.float16]": 0.158183202999993,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-1024-True-False-torch.float16]": 0.1674369750000153,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-1024-True-True-torch.float16]": 0.1654174270000226,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-511-False-False-torch.float16]": 0.18823592499998654,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-511-False-True-torch.float16]": 0.17398274499996091,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-511-True-False-torch.float16]": 0.1816077489999941,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-511-True-True-torch.float16]": 0.16726551600004314,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-512-False-False-torch.float16]": 0.1843474749999814,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-512-False-True-torch.float16]": 0.17326306999996177,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-512-True-False-torch.float16]": 0.1783643689999792,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-512-True-True-torch.float16]": 0.18119622099999333,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-1023-False-False-torch.float16]": 57.69836592699994,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-1023-False-True-torch.float16]": 291.49695931999986,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-1023-True-False-torch.float16]": 13.491787292000026,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-1023-True-True-torch.float16]": 23.311794528999997,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-1024-False-False-torch.float16]": 13.88605859200004,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-1024-False-True-torch.float16]": 13.837973902000044,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-1024-True-False-torch.float16]": 13.642192917999978,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-1024-True-True-torch.float16]": 13.770189781,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-511-False-False-torch.float16]": 132.77990358699998,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-511-False-True-torch.float16]": 157.78928541599998,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-511-True-False-torch.float16]": 16.965602879000016,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-511-True-True-torch.float16]": 26.072262103999947,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-512-False-False-torch.float16]": 17.71862301200008,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-512-False-True-torch.float16]": 17.353098821999993,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-512-True-False-torch.float16]": 17.32231065100001,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-512-True-True-torch.float16]": 17.51899144099997,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-1023-False-False-torch.float16]": 57.25240083100016,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-1023-False-True-torch.float16]": 298.6479218239999,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-1023-True-False-torch.float16]": 13.718209162999983,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-1023-True-True-torch.float16]": 23.832339372000206,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-1024-False-False-torch.float16]": 13.6509698389998,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-1024-False-True-torch.float16]": 14.16245389400001,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-1024-True-False-torch.float16]": 13.441948295999964,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-1024-True-True-torch.float16]": 13.882072858000129,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-511-False-False-torch.float16]": 133.77471671600006,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-511-False-True-torch.float16]": 170.41293349400007,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-511-True-False-torch.float16]": 17.643053216999988,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-511-True-True-torch.float16]": 25.50387371799991,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-512-False-False-torch.float16]": 17.70623671600015,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-512-False-True-torch.float16]": 17.95185384399997,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-512-True-False-torch.float16]": 17.820820709000145,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-512-True-True-torch.float16]": 17.81172918299967,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-1023-False-False-torch.float16]": 50.539672963000044,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-1023-False-True-torch.float16]": 295.674271277,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-1023-True-False-torch.float16]": 13.472480462000021,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-1023-True-True-torch.float16]": 23.802054028999862,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-1024-False-False-torch.float16]": 13.484827779000057,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-1024-False-True-torch.float16]": 13.53746117999998,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-1024-True-False-torch.float16]": 13.34516426500005,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-1024-True-True-torch.float16]": 13.67660106599999,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-511-False-False-torch.float16]": 134.01968161299988,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-511-False-True-torch.float16]": 168.03832814000006,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-511-True-False-torch.float16]": 17.01759517100001,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-511-True-True-torch.float16]": 26.271212453999993,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-512-False-False-torch.float16]": 17.045606676000034,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-512-False-True-torch.float16]": 17.492527484999982,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-512-True-False-torch.float16]": 16.717237677999947,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-512-True-True-torch.float16]": 16.894005321999884,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-1023-False-False-torch.float16]": 52.303189300999975,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-1023-False-True-torch.float16]": 289.81769851499996,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-1023-True-False-torch.float16]": 13.166050436000091,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-1023-True-True-torch.float16]": 23.095223638999983,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-1024-False-False-torch.float16]": 13.555932448000021,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-1024-False-True-torch.float16]": 13.659670687000016,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-1024-True-False-torch.float16]": 13.141814365999949,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-1024-True-True-torch.float16]": 13.18186683700003,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-511-False-False-torch.float16]": 129.578922627,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-511-False-True-torch.float16]": 164.28969659600006,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-511-True-False-torch.float16]": 16.99811207199997,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-511-True-True-torch.float16]": 25.444895737000024,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-512-False-False-torch.float16]": 16.970982986999843,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-512-False-True-torch.float16]": 16.9878909439999,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-512-True-False-torch.float16]": 17.09009328899981,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-512-True-True-torch.float16]": 17.11837285899992,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-1023-False-False-torch.float16]": 0.12972502400000963,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-1023-False-True-torch.float16]": 0.1471466419999956,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-1023-True-False-torch.float16]": 0.13393206500001043,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-1023-True-True-torch.float16]": 0.11622101799997608,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-1024-False-False-torch.float16]": 1.327832347999987,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-1024-False-True-torch.float16]": 1.9241082950000248,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-1024-True-False-torch.float16]": 0.16195460499997694,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-1024-True-True-torch.float16]": 0.1413496800000189,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-511-False-False-torch.float16]": 0.1204590440000004,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-511-False-True-torch.float16]": 0.1276232419999701,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-511-True-False-torch.float16]": 0.1202325430000144,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-511-True-True-torch.float16]": 0.12902814699998544,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-512-False-False-torch.float16]": 0.1098283259999846,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-512-False-True-torch.float16]": 0.10579896500004793,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-512-True-False-torch.float16]": 0.10737069700002166,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-512-True-True-torch.float16]": 0.11067168800005334,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-1023-False-False-torch.float16]": 0.1387717799999848,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-1023-False-True-torch.float16]": 0.11998391399998809,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-1023-True-False-torch.float16]": 0.15890873099996838,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-1023-True-True-torch.float16]": 0.1562839530000133,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-1024-False-False-torch.float16]": 0.13047116699996764,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-1024-False-True-torch.float16]": 0.1371181200000251,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-1024-True-False-torch.float16]": 0.1383077030000095,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-1024-True-True-torch.float16]": 0.14826274899999703,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-511-False-False-torch.float16]": 0.1620138529999906,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-511-False-True-torch.float16]": 0.1790892660000054,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-511-True-False-torch.float16]": 1.3597707250000042,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-511-True-True-torch.float16]": 0.1711673929999904,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-512-False-False-torch.float16]": 0.15053349000001504,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-512-False-True-torch.float16]": 0.16103127299999187,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-512-True-False-torch.float16]": 0.1536379620000048,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-512-True-True-torch.float16]": 0.15407123600002137,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-1023-False-False-torch.float16]": 0.12016235500004768,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-1023-False-True-torch.float16]": 0.13306764399999338,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-1023-True-False-torch.float16]": 0.1630231839999965,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-1023-True-True-torch.float16]": 0.14730254100001616,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-1024-False-False-torch.float16]": 0.13546284099999184,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-1024-False-True-torch.float16]": 0.14486511499998755,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-1024-True-False-torch.float16]": 0.15235340399996744,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-1024-True-True-torch.float16]": 0.13348256199998332,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-511-False-False-torch.float16]": 0.15991857100001994,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-511-False-True-torch.float16]": 0.1672904419999668,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-511-True-False-torch.float16]": 1.9616818229999922,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-511-True-True-torch.float16]": 2.681044533000005,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-512-False-False-torch.float16]": 0.16134574800003065,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-512-False-True-torch.float16]": 0.1694752360000109,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-512-True-False-torch.float16]": 0.1517858400000307,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-512-True-True-torch.float16]": 0.15261863700001754,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-1023-False-False-torch.float16]": 0.12782135099999437,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-1023-False-True-torch.float16]": 0.12864554300000464,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-1023-True-False-torch.float16]": 0.1325558130000104,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-1023-True-True-torch.float16]": 0.14826950899998792,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-1024-False-False-torch.float16]": 0.1324696390000213,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-1024-False-True-torch.float16]": 0.1133549259999711,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-1024-True-False-torch.float16]": 0.11709143700002755,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-1024-True-True-torch.float16]": 0.11898302299999841,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-511-False-False-torch.float16]": 0.11976096900002631,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-511-False-True-torch.float16]": 0.14592404699999406,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-511-True-False-torch.float16]": 0.14255564399999798,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-511-True-True-torch.float16]": 0.13948803100001328,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-512-False-False-torch.float16]": 0.12100858799999514,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-512-False-True-torch.float16]": 0.14399551100001418,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-512-True-False-torch.float16]": 0.13392968199997313,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-512-True-True-torch.float16]": 0.13400094299998955,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-1023-False-False-torch.float16]": 0.18065573999999174,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-1023-False-True-torch.float16]": 0.1730078210000272,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-1023-True-False-torch.float16]": 0.16104598300000816,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-1023-True-True-torch.float16]": 0.17457540399999516,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-1024-False-False-torch.float16]": 0.1655270940000264,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-1024-False-True-torch.float16]": 0.17906287099998508,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-1024-True-False-torch.float16]": 0.17501338899998586,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-1024-True-True-torch.float16]": 0.16476445200001422,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-511-False-False-torch.float16]": 0.14867826099998638,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-511-False-True-torch.float16]": 0.17248826200000167,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-511-True-False-torch.float16]": 0.16420958200001223,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-511-True-True-torch.float16]": 0.1709937589999697,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-512-False-False-torch.float16]": 0.17257754799996405,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-512-False-True-torch.float16]": 0.16326321600001847,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-512-True-False-torch.float16]": 0.1711246040000276,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-512-True-True-torch.float16]": 0.16878054200006432,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-1023-False-False-torch.float16]": 0.13456352700001162,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-1023-False-True-torch.float16]": 0.14066604400002802,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-1023-True-False-torch.float16]": 0.1509927489999825,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-1023-True-True-torch.float16]": 0.1505847729999914,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-1024-False-False-torch.float16]": 0.1479353690000096,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-1024-False-True-torch.float16]": 0.14757696700002043,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-1024-True-False-torch.float16]": 0.14725429399996415,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-1024-True-True-torch.float16]": 0.14421212699997454,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-511-False-False-torch.float16]": 0.14135365399999955,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-511-False-True-torch.float16]": 0.16014445999999793,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-511-True-False-torch.float16]": 0.1704987180000046,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-511-True-True-torch.float16]": 0.17711753999998336,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-512-False-False-torch.float16]": 0.15203431100002263,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-512-False-True-torch.float16]": 0.14885940199999936,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-512-True-False-torch.float16]": 0.14530539900002282,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-512-True-True-torch.float16]": 0.14836723100000881,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-1023-False-False-torch.float16]": 0.13685048299998925,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-1023-False-True-torch.float16]": 0.1467517849999922,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-1023-True-False-torch.float16]": 0.13602184899997383,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-1023-True-True-torch.float16]": 0.1365213800000049,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-1024-False-False-torch.float16]": 0.13905003699997565,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-1024-False-True-torch.float16]": 0.14752112799999395,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-1024-True-False-torch.float16]": 0.13336045899998794,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-1024-True-True-torch.float16]": 0.13716151200000581,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-511-False-False-torch.float16]": 0.1619932510000126,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-511-False-True-torch.float16]": 0.17062984099999312,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-511-True-False-torch.float16]": 0.16582041300000583,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-511-True-True-torch.float16]": 0.17587376800000243,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-512-False-False-torch.float16]": 0.1341358029999924,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-512-False-True-torch.float16]": 0.13693249799999307,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-512-True-False-torch.float16]": 0.13552402999999913,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-512-True-True-torch.float16]": 0.1400996459999817,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-1023-False-False-torch.float16]": 0.18105592400004866,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-1023-False-True-torch.float16]": 0.17412063000000444,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-1023-True-False-torch.float16]": 0.16566298999998708,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-1023-True-True-torch.float16]": 0.17890475900000524,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-1024-False-False-torch.float16]": 0.16695071000003736,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-1024-False-True-torch.float16]": 0.17026537200004555,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-1024-True-False-torch.float16]": 0.16167803300001538,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-1024-True-True-torch.float16]": 0.16755982399999425,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-511-False-False-torch.float16]": 0.16934171700000888,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-511-False-True-torch.float16]": 0.1721667449999984,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-511-True-False-torch.float16]": 0.14257341100002918,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-511-True-True-torch.float16]": 0.13982191400000943,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-512-False-False-torch.float16]": 0.16223500300000637,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-512-False-True-torch.float16]": 0.1468269710000243,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-512-True-False-torch.float16]": 0.16446783399996434,
-"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-512-True-True-torch.float16]": 0.16342565400000808,
-"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill[2048-128-torch.bfloat16]": 0.002354761999981747,
-"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill[2048-128-torch.float32]": 0.003990474000005406,
-"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill[2048-16-torch.bfloat16]": 0.14578403400003026,
-"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill[2048-16-torch.float32]": 0.18440348100003234,
-"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill[256-7-torch.bfloat16]": 0.14863136200000326,
-"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill[256-7-torch.float32]": 0.16455353599997125,
-"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill[300-33-torch.bfloat16]": 0.14238810100002297,
-"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill[300-33-torch.float32]": 0.15454987600000436,
-"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill_matches_update_stepwise": 0.16005512300000646,
-"tests/ops/test_causal_conv1d_lfm2.py::test_op_update[2048-torch.bfloat16]": 0.16433251200004406,
-"tests/ops/test_causal_conv1d_lfm2.py::test_op_update[2048-torch.float32]": 0.17109616300001562,
-"tests/ops/test_causal_conv1d_lfm2.py::test_op_update[256-torch.bfloat16]": 0.002471803000020145,
-"tests/ops/test_causal_conv1d_lfm2.py::test_op_update[256-torch.float32]": 0.0036539409999818417,
-"tests/ops/test_causal_conv1d_lfm2.py::test_op_update[300-torch.bfloat16]": 0.1860871119999956,
-"tests/ops/test_causal_conv1d_lfm2.py::test_op_update[300-torch.float32]": 0.16042178299994703,
-"tests/ops/test_causal_conv1d_lfm2.py::test_op_update_2d_legacy_call[torch.bfloat16]": 0.002684479999999212,
-"tests/ops/test_causal_conv1d_lfm2.py::test_op_update_2d_legacy_call[torch.float32]": 0.003018806999989465,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-K256_V128-bf16]": 34.14555530200002,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-K256_V128-fp32]": 0.0015812930000151937,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-K64_V128-bf16]": 10.833571021000012,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-K64_V128-fp32]": 0.0014093029999742157,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T100_unaligned-bf16]": 6.455326649999989,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T100_unaligned-fp32]": 0.0014287489999844638,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T16-bf16]": 0.24538836800005015,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T16-fp32]": 0.0015656799999703708,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T2048_long_seq-bf16]": 12.668593650000048,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T2048_long_seq-fp32]": 0.0013805789999423723,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T32_final_state-bf16]": 24.062110378,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T32_final_state-fp32]": 0.0016605539999829944,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T37_prime-bf16]": 0.1929992739999875,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T37_prime-fp32]": 0.0012747449999608307,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T512_long_seq-bf16]": 12.691462074000015,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T512_long_seq-fp32]": 0.0016076889999681043,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T64-bf16]": 0.14346918899997263,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T64-fp32]": 0.0016348099999845545,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T65_off_by_1-bf16]": 23.662088204000014,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T65_off_by_1-fp32]": 0.0011860600000090926,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-chunk32-bf16]": 3.965768204999989,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-chunk32-fp32]": 0.001447255000016412,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-decode-bf16]": 9.062846747000009,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-decode-fp32]": 0.0013245240000401282,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-init_final_state-bf16]": 9.925796228000024,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-init_final_state-fp32]": 0.0013046820000113257,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-l2norm-bf16]": 8.349837508999997,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-l2norm-fp32]": 0.003143040000026076,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-minimal-bf16]": 9.266539973999954,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-minimal-fp32]": 0.0023720459999765353,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-qwen3_5_T10782-bf16]": 2.985642958000028,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-qwen3_5_T10782-fp32]": 0.0012171040000339417,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-K256_V128-bf16]": 155.910470203,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-K256_V128-fp32]": 0.002704043999926853,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-K64_V128-bf16]": 12.9090155990001,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-K64_V128-fp32]": 0.002129796000076567,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T100_unaligned-bf16]": 0.21871101699991868,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T100_unaligned-fp32]": 0.0021877420000464554,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T16-bf16]": 0.15864285800000744,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T16-fp32]": 0.0013660679999816239,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T2048_long_seq-bf16]": 0.0011873299999365372,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T2048_long_seq-fp32]": 0.001046541000050638,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T32_final_state-bf16]": 10.403869537000048,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T32_final_state-fp32]": 0.0011562650000200847,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T37_prime-bf16]": 14.449041041999976,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T37_prime-fp32]": 0.0016036270000086006,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T512_long_seq-bf16]": 0.001057639000009658,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T512_long_seq-fp32]": 0.0012594840000019758,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T64-bf16]": 0.17224458300006518,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T64-fp32]": 0.0012783729999910065,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T65_off_by_1-bf16]": 50.995643884999936,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T65_off_by_1-fp32]": 0.0012408399999799258,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-chunk32-bf16]": 4.544282572000043,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-chunk32-fp32]": 0.0013325310000027457,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-decode-bf16]": 0.15052829699999393,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-decode-fp32]": 0.0014337850000742947,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-init_final_state-bf16]": 10.182990999000026,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-init_final_state-fp32]": 0.0012344669999606595,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-l2norm-bf16]": 0.001284571000098822,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-l2norm-fp32]": 0.001855837000050542,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-minimal-bf16]": 0.17450942299996086,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-minimal-fp32]": 0.0012786060000280486,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-qwen3_5_T10782-bf16]": 0.0012025940000057744,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-qwen3_5_T10782-fp32]": 0.0010826699999597622,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op_correlated_keys_stable_triangular_solve[cutile-in_kernel_l2]": 3.136417646000041,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op_correlated_keys_stable_triangular_solve[cutile-prenormalized]": 23.267026869000006,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op_correlated_keys_stable_triangular_solve[tilecpp-in_kernel_l2]": 0.0016193160000170792,
-"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op_correlated_keys_stable_triangular_solve[tilecpp-prenormalized]": 11.794935247000069,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-1024-0.2-False-False-torch.float32-1e-05]": 0.10600968199997851,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-1024-0.2-False-True-torch.float32-1e-05]": 0.10312861499994597,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-1024-0.2-True-False-torch.float32-1e-05]": 0.12304907599991566,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-1024-0.2-True-True-torch.float32-1e-05]": 0.284475719999989,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-4096-0.4-False-False-torch.float32-1e-05]": 0.13478110599999127,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-4096-0.4-False-True-torch.float32-1e-05]": 0.10648974299999736,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-4096-0.4-True-False-torch.float32-1e-05]": 0.16423257499997135,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-4096-0.4-True-True-torch.float32-1e-05]": 0.2570092930000669,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-65536-0.8-False-False-torch.float16-1e-05]": 0.16255204199995887,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-65536-0.8-False-False-torch.float32-1e-05]": 0.12891033799999718,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-65536-0.8-False-True-torch.float16-1e-05]": 0.13866054499999336,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-65536-0.8-False-True-torch.float32-1e-05]": 0.1456104319999838,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-65536-0.8-True-False-torch.float16-1e-05]": 0.14991940299998419,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-65536-0.8-True-False-torch.float32-1e-05]": 0.17718028699999877,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-65536-0.8-True-True-torch.float16-1e-05]": 0.24158105699999055,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-65536-0.8-True-True-torch.float32-1e-05]": 0.33441941999998903,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-1024-0.2-False-False-torch.float32-1e-05]": 0.14920504100001608,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-1024-0.2-False-True-torch.float32-1e-05]": 0.1304247440000097,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-1024-0.2-True-False-torch.float32-1e-05]": 0.1726272129999984,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-1024-0.2-True-True-torch.float32-1e-05]": 0.6701856219999911,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-4096-0.4-False-False-torch.float32-1e-05]": 0.15579314400000044,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-4096-0.4-False-True-torch.float32-1e-05]": 0.15893375199999582,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-4096-0.4-True-False-torch.float32-1e-05]": 0.1483170709999797,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-4096-0.4-True-True-torch.float32-1e-05]": 0.6782191570000009,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-65536-0.8-False-False-torch.float16-1e-05]": 0.14386986900001375,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-65536-0.8-False-False-torch.float32-1e-05]": 0.12917614299999514,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-65536-0.8-False-True-torch.float16-1e-05]": 0.14294584700002133,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-65536-0.8-False-True-torch.float32-1e-05]": 0.15003565300003174,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-65536-0.8-True-False-torch.float16-1e-05]": 0.1805788039999925,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-65536-0.8-True-False-torch.float32-1e-05]": 0.17246527699998637,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-65536-0.8-True-True-torch.float16-1e-05]": 0.6834634180000023,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-65536-0.8-True-True-torch.float32-1e-05]": 0.7500488979999886,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-1024-0.2-False-False-torch.float32-1e-05]": 0.11815156399998727,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-1024-0.2-False-True-torch.float32-1e-05]": 0.12823863099995947,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-1024-0.2-True-False-torch.float32-1e-05]": 0.17775300700003527,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-1024-0.2-True-True-torch.float32-1e-05]": 9.246091336999996,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-4096-0.4-False-False-torch.float32-1e-05]": 0.1250450219999948,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-4096-0.4-False-True-torch.float32-1e-05]": 0.14260297799998511,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-4096-0.4-True-False-torch.float32-1e-05]": 0.1654580550000162,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-4096-0.4-True-True-torch.float32-1e-05]": 2.084688195000041,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-65536-0.8-False-False-torch.float16-1e-05]": 0.15082200499998066,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-65536-0.8-False-False-torch.float32-1e-05]": 0.14133237400000098,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-65536-0.8-False-True-torch.float16-1e-05]": 0.13677736700000764,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-65536-0.8-False-True-torch.float32-1e-05]": 0.13135657199998718,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-65536-0.8-True-False-torch.float16-1e-05]": 0.17651256300004547,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-65536-0.8-True-False-torch.float32-1e-05]": 0.142647654000001,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-65536-0.8-True-True-torch.float16-1e-05]": 7.504311319999971,
-"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-65536-0.8-True-True-torch.float32-1e-05]": 2.0771611730000075,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-1-119]": 0.32350471100002665,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-1-256]": 0.7000436390000004,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-1-8192]": 1.2820395969999936,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-1-9]": 0.6826970619999599,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-4-119]": 0.1473952520000239,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-4-256]": 0.47067152000002466,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-4-8192]": 1.391928584000027,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-4-9]": 0.5241932470000279,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-8-119]": 0.1598726480000039,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-8-256]": 0.4782751750000216,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-8-8192]": 0.9217782249999971,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-8-9]": 0.5442961869999863,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-1-119]": 1.0515643879999743,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-1-256]": 1.0175993309999853,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-1-8192]": 1.6270505610000043,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-1-9]": 1.544888716000031,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-4-119]": 1.0455101689999537,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-4-256]": 1.067287704999984,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-4-8192]": 1.735697559000016,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-4-9]": 1.0380582600000139,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-8-119]": 1.0784166140000195,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-8-256]": 1.0838220349999688,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-8-8192]": 1.7482013030000019,
-"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-8-9]": 1.0194060790000208,
-"tests/ops/test_gemma_attention.py::TestGemmaAttention::test_op[cutile-1-12-4-512-64-256-30.0-True-torch.bfloat16]": 0.0011191919999760103,
-"tests/ops/test_gemma_attention.py::TestGemmaAttention::test_op[cutile-1-8-8-128-64-0-None-False-torch.float16]": 0.0010386120000021037,
-"tests/ops/test_gemma_attention.py::TestGemmaAttention::test_op[cutile-1-8-8-128-64-64-10.0-False-torch.float32]": 0.005305107999987513,
-"tests/ops/test_gemma_attention.py::TestGemmaAttention::test_op[cutile-1-8-8-128-64-64-10.0-True-torch.float16]": 0.0013315279999801533,
-"tests/ops/test_gemma_attention.py::TestGemmaAttention::test_op[tilecpp-1-12-4-512-64-256-30.0-True-torch.bfloat16]": 0.0011517960000162475,
-"tests/ops/test_gemma_attention.py::TestGemmaAttention::test_op[tilecpp-1-8-8-128-64-0-None-False-torch.float16]": 0.0011293390000162162,
-"tests/ops/test_gemma_attention.py::TestGemmaAttention::test_op[tilecpp-1-8-8-128-64-64-10.0-False-torch.float32]": 0.0012595069999576936,
-"tests/ops/test_gemma_attention.py::TestGemmaAttention::test_op[tilecpp-1-8-8-128-64-64-10.0-True-torch.float16]": 0.0011259899999913614,
-"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[cutile-1-8-4-256-64-64-10.0-torch.float16]": 0.0011946310000041649,
-"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[cutile-1-8-8-10019-64-256-50.0-torch.bfloat16]": 0.0011288890000002993,
-"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[cutile-1-8-8-10024-64-0-None-torch.float16]": 0.0010946310000008452,
-"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[cutile-1-8-8-128-64-0-10.0-torch.float16]": 0.0012052550000021256,
-"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[cutile-1-8-8-128-64-0-None-torch.float16]": 0.001140444999975898,
-"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[cutile-1-8-8-256-64-64-10.0-torch.float16]": 0.0010846949999745448,
-"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[tilecpp-1-8-4-256-64-64-10.0-torch.float16]": 0.0010664670000153365,
-"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[tilecpp-1-8-8-10019-64-256-50.0-torch.bfloat16]": 0.0010763829999689278,
-"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[tilecpp-1-8-8-10024-64-0-None-torch.float16]": 0.00102991499997529,
-"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[tilecpp-1-8-8-128-64-0-10.0-torch.float16]": 0.001129451999986486,
-"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[tilecpp-1-8-8-128-64-0-None-torch.float16]": 0.001078420999959917,
-"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[tilecpp-1-8-8-256-64-64-10.0-torch.float16]": 0.001116603999975041,
-"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[1024, 512, 256, 128]-[1024, 512, 256, 128]-[1024, 512, 256, 128]-False-torch.float16]": 4.823688075000035,
-"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[1024, 512, 256, 128]-[1024, 512, 256, 128]-[1024, 512, 256, 128]-True-torch.float16]": 4.650415747000011,
-"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[1024, 512, 256, 128]-[1024, 512, 256, 128]-[128, 128, 128, 128]-False-torch.float16]": 0.24808855499998117,
-"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[1024, 512, 256, 128]-[1024, 512, 256, 128]-[128, 128, 128, 128]-True-torch.float16]": 0.2588787589999697,
-"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[1024, 512, 256, 128]-[128, 128, 128, 128]-[1024, 512, 256, 128]-False-torch.float16]": 0.25241741300004605,
-"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[1024, 512, 256, 128]-[128, 128, 128, 128]-[1024, 512, 256, 128]-True-torch.float16]": 0.23944977399997924,
-"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[1024, 512, 256, 128]-[128, 128, 128, 128]-[128, 128, 128, 128]-False-torch.float16]": 0.23326034099994786,
-"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[1024, 512, 256, 128]-[128, 128, 128, 128]-[128, 128, 128, 128]-True-torch.float16]": 0.2673460400000067,
-"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[256, 256, 256, 256]-[1024, 512, 256, 128]-[1024, 512, 256, 128]-False-torch.float16]": 0.25860433299999386,
-"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[256, 256, 256, 256]-[1024, 512, 256, 128]-[1024, 512, 256, 128]-True-torch.float16]": 0.23226373900001818,
-"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[256, 256, 256, 256]-[1024, 512, 256, 128]-[128, 128, 128, 128]-False-torch.float16]": 0.27472651700003325,
-"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[256, 256, 256, 256]-[1024, 512, 256, 128]-[128, 128, 128, 128]-True-torch.float16]": 0.2816166019999855,
-"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[256, 256, 256, 256]-[128, 128, 128, 128]-[1024, 512, 256, 128]-False-torch.float16]": 0.2635969389999673,
-"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[256, 256, 256, 256]-[128, 128, 128, 128]-[1024, 512, 256, 128]-True-torch.float16]": 0.3072191029999658,
-"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[256, 256, 256, 256]-[128, 128, 128, 128]-[128, 128, 128, 128]-False-torch.float16]": 0.28163740799993775,
-"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[256, 256, 256, 256]-[128, 128, 128, 128]-[128, 128, 128, 128]-True-torch.float16]": 0.26203032399996573,
-"tests/ops/test_layer_norm.py::Test_LayerNorm2D::test_op[cutile-1004-2456-1.0-torch.float16]": 0.5494645040000137,
-"tests/ops/test_layer_norm.py::Test_LayerNorm2D::test_op[cutile-256-256-0.0-torch.float16]": 0.3738701210000386,
-"tests/ops/test_layer_norm.py::Test_LayerNorm2D::test_op[cutile-256-256-1.0-torch.float16]": 0.328458339000008,
-"tests/ops/test_layer_norm.py::Test_LayerNorm2D::test_op[cutile-512-256-0.0-torch.float32]": 0.3181496260000358,
-"tests/ops/test_layer_norm.py::Test_LayerNorm2D::test_op[cutile-9-9-0.0-torch.float16]": 0.3062416309999776,
-"tests/ops/test_layer_norm.py::Test_LayerNormNCHW::test_op[cutile-2-512-256-0.0-torch.float16]": 3.5908276599999454,
-"tests/ops/test_layer_norm.py::Test_LayerNormNCHW::test_op[cutile-256-256-128-1.0-torch.float16]": 1.017027594999945,
-"tests/ops/test_layer_norm.py::Test_LayerNormNCHW::test_op[cutile-512-256-128-0.0-torch.float32]": 1.2306447869999602,
-"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[cutile-256-256-0.0-torch.float16]": 0.3008881070000484,
-"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[cutile-256-256-0.0-torch.float32]": 0.35648271999997405,
-"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[cutile-256-256-1.0-torch.float16]": 0.33240054300000565,
-"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[cutile-9-9-0.0-torch.float16]": 0.34555510600000616,
-"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[tilecpp-256-256-0.0-torch.float16]": 0.6820813040000075,
-"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[tilecpp-256-256-0.0-torch.float32]": 0.7703972080000199,
-"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[tilecpp-256-256-1.0-torch.float16]": 0.17498283499998024,
-"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[tilecpp-9-9-0.0-torch.float16]": 0.6577410070000269,
-"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[triton-256-256-0.0-torch.float16]": 11.253448165999998,
-"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[triton-256-256-0.0-torch.float32]": 11.101862406999999,
-"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[triton-256-256-1.0-torch.float16]": 0.16244572299996207,
-"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[triton-9-9-0.0-torch.float16]": 12.96100288499997,
-"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[cutile-1024-1024-torch.bfloat16]": 5.294113060000029,
-"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[cutile-256-256-torch.bfloat16]": 3.4596467300000313,
-"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[cutile-30000-1024-torch.bfloat16]": 5.48426538800004,
-"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[cutile-4096-512-torch.bfloat16]": 4.682026176000079,
-"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[tilecpp-1024-1024-torch.bfloat16]": 0.0011270310000099926,
-"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[tilecpp-256-256-torch.bfloat16]": 3.7539826459999404,
-"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[tilecpp-30000-1024-torch.bfloat16]": 0.001112988000045334,
-"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[tilecpp-4096-512-torch.bfloat16]": 4.5028097939999725,
-"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[triton-1024-1024-torch.bfloat16]": 15.339369932000011,
-"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[triton-256-256-torch.bfloat16]": 9.014357998000037,
-"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[triton-30000-1024-torch.bfloat16]": 2.3042934439999954,
-"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[triton-4096-512-torch.bfloat16]": 13.353810894000048,
-"tests/ops/test_linear_gluact_linear.py::Test_LinearGluactLinear::test_op[cutile-1024-1024-1023-silu-torch.float32]": 0.0010109479999869109,
-"tests/ops/test_linear_gluact_linear.py::Test_LinearGluactLinear::test_op[cutile-1024-1024-1024-relu-torch.float32]": 0.0012960020000605255,
-"tests/ops/test_linear_gluact_linear.py::Test_LinearGluactLinear::test_op[cutile-1024-1024-1024-silu-torch.float32]": 0.0009952429999771084,
-"tests/ops/test_linear_gluact_linear.py::Test_LinearGluactLinear::test_op[cutile-1024-1024-512-gelu_tanh-torch.float32]": 0.001016684000035184,
-"tests/ops/test_linear_gluact_linear.py::Test_LinearGluactLinear::test_op[cutile-512-512-256-silu-torch.float16]": 0.0010190999999508676,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-T192-bf16]": 10.804827638999996,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-batch2-bf16]": 161.24556553899993,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-chunk128-ungrouped-bf16]": 877.0460355159998,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-groups2-bf16]": 181.31275430799997,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-large-bf16]": 313.236688532,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-minimal-bf16]": 179.96462089800002,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-single-chunk-bf16]": 224.72569259199997,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-standard-bf16]": 391.21901491000017,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-three-chunks-bf16]": 479.93455014300014,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-T192-bf16]": 1.721914646000073,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-batch2-bf16]": 1.8996541239999942,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-chunk128-ungrouped-bf16]": 239.46483165799998,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-groups2-bf16]": 33.04401553600002,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-large-bf16]": 115.46924479899997,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-minimal-bf16]": 36.718328801999974,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-single-chunk-bf16]": 35.89636320500006,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-standard-bf16]": 69.79899464399995,
-"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-three-chunks-bf16]": 4.973912800999869,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=False-1024-1024-1023-0-0-False-False-torch.bfloat16]": 0.001698712000006708,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=False-1024-1024-1024-0-0-False-False-torch.bfloat16]": 1.0313245700001517,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=False-3072-6144-2720-0-0-False-False-torch.bfloat16]": 0.00171145499996328,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=False-8-8-8-0-0-False-False-torch.bfloat16]": 1.1512047339999754,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=False-8192-8192-8192-0-0-False-False-torch.bfloat16]": 0.0013581960000692561,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=True-1024-1024-1023-0-0-False-False-torch.bfloat16]": 0.0013425549996100017,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=True-1024-1024-1024-0-0-False-False-torch.bfloat16]": 10.627573303999952,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=True-3072-6144-2720-0-0-False-False-torch.bfloat16]": 0.0016117840002607409,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=True-8-8-8-0-0-False-False-torch.bfloat16]": 10.36533936800015,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=True-8192-8192-8192-0-0-False-False-torch.bfloat16]": 0.003381121999836978,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=False-1024-1024-1023-0-0-False-False-torch.bfloat16]": 0.0013912989999198544,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=False-1024-1024-1024-0-0-False-False-torch.bfloat16]": 0.15971269700003177,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=False-3072-6144-2720-0-0-False-False-torch.bfloat16]": 0.0012407720000737754,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=False-8-8-8-0-0-False-False-torch.bfloat16]": 0.16769766000015807,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=False-8192-8192-8192-0-0-False-False-torch.bfloat16]": 0.0011982819997911065,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=True-1024-1024-1023-0-0-False-False-torch.bfloat16]": 0.0015493289998858017,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=True-1024-1024-1024-0-0-False-False-torch.bfloat16]": 0.20144192200007183,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=True-3072-6144-2720-0-0-False-False-torch.bfloat16]": 0.0017590199998949174,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=True-8-8-8-0-0-False-False-torch.bfloat16]": 0.1831429469999648,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=True-8192-8192-8192-0-0-False-False-torch.bfloat16]": 0.001498214999855918,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=False-1024-1024-1023-0-0-False-False-torch.bfloat16]": 0.0014956019999772252,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=False-1024-1024-1024-0-0-False-False-torch.bfloat16]": 1.9926009859999567,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=False-3072-6144-2720-0-0-False-False-torch.bfloat16]": 0.0054366379999919445,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=False-8-8-8-0-0-False-False-torch.bfloat16]": 1.6953474650000544,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=False-8192-8192-8192-0-0-False-False-torch.bfloat16]": 0.003446745000019291,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=True-1024-1024-1023-0-0-False-False-torch.bfloat16]": 0.0014867109999840977,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=True-1024-1024-1024-0-0-False-False-torch.bfloat16]": 84.14775664899992,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=True-3072-6144-2720-0-0-False-False-torch.bfloat16]": 0.0012711430000535984,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=True-8-8-8-0-0-False-False-torch.bfloat16]": 7.202428507000036,
-"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=True-8192-8192-8192-0-0-False-False-torch.bfloat16]": 0.0012883059999921898,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(2112, 7168)-256]": 32.49781374899976,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(2112, 7168)-32]": 13.538739293000162,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(24576, 1536)-256]": 29.007181547000073,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(24576, 1536)-32]": 13.426156044000209,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(24576, 7168)-256]": 32.78087770600018,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(24576, 7168)-32]": 13.454600078000112,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(3072, 1536)-256]": 29.490369185999953,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(3072, 1536)-32]": 13.47194484900001,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(32768, 512)-256]": 29.38135245500007,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(32768, 512)-32]": 12.218179340999995,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(36864, 7168)-256]": 32.93394729400006,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(36864, 7168)-32]": 13.581952347999732,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(4096, 7168)-256]": 32.560232521000216,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(4096, 7168)-32]": 13.414814254999783,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(576, 7168)-256]": 32.528902315000096,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(576, 7168)-32]": 13.73043925800016,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(7168, 16384)-256]": 40.83039530800011,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(7168, 16384)-32]": 16.118425183,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(7168, 18432)-256]": 43.54673422000019,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(7168, 18432)-32]": 16.91739924699982,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(7168, 2048)-256]": 31.793723662000275,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(7168, 2048)-32]": 15.399739515999954,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(7168, 256)-256]": 23.483363388000043,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(7168, 256)-32]": 8.968021157000067,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(2112, 7168)-256]": 28.983329069000092,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(2112, 7168)-32]": 11.794444148000025,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(24576, 1536)-256]": 20.429700571999888,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(24576, 1536)-32]": 9.14575155099999,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(24576, 7168)-256]": 29.463368593000155,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(24576, 7168)-32]": 11.783363923000024,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(3072, 1536)-256]": 20.645027264000078,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(3072, 1536)-32]": 8.958560648999992,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(32768, 512)-256]": 23.11345385700008,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(32768, 512)-32]": 10.167612499000143,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(36864, 7168)-256]": 29.790265470000122,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(36864, 7168)-32]": 11.407197665999888,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(4096, 7168)-256]": 28.928131410999868,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(4096, 7168)-32]": 11.467979963999937,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(576, 7168)-256]": 29.25809813800015,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(576, 7168)-32]": 11.33160093899994,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(7168, 16384)-256]": 42.86505574899991,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(7168, 16384)-32]": 15.579708902999982,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(7168, 18432)-256]": 45.94461435100004,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(7168, 18432)-32]": 15.96079606800015,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(7168, 2048)-256]": 21.447045315999958,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(7168, 2048)-32]": 9.517716336000035,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(7168, 256)-256]": 18.38246466800001,
-"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(7168, 256)-32]": 8.032651425999802,
-"tests/ops/test_mla.py::Test_MLA::test_op[cutile-1-64-64-torch.bfloat16-False]": 0.0020341000001735665,
-"tests/ops/test_mla.py::Test_MLA::test_op[cutile-1-64-64-torch.bfloat16-True]": 0.0019466869998723269,
-"tests/ops/test_mla.py::Test_MLA::test_op[cutile-4-64-64-torch.bfloat16-False]": 0.0018147729999782314,
-"tests/ops/test_mla.py::Test_MLA::test_op[cutile-4-64-64-torch.bfloat16-True]": 0.0019567249998999614,
-"tests/ops/test_mla.py::Test_MLA::test_op[tilecpp-1-64-64-torch.bfloat16-False]": 0.0014123959997505153,
-"tests/ops/test_mla.py::Test_MLA::test_op[tilecpp-1-64-64-torch.bfloat16-True]": 0.003017515999999887,
-"tests/ops/test_mla.py::Test_MLA::test_op[tilecpp-4-64-64-torch.bfloat16-False]": 0.0019464560000415077,
-"tests/ops/test_mla.py::Test_MLA::test_op[tilecpp-4-64-64-torch.bfloat16-True]": 0.0012725459998819133,
-"tests/ops/test_mla_decoding.py::Test_MLADecoding::test_op[cutile-512-64-torch.float16-128-False]": 0.0012992409997423238,
-"tests/ops/test_mla_decoding.py::Test_MLADecoding::test_op[cutile-512-64-torch.float16-16-True]": 1.2001498469999206,
-"tests/ops/test_mla_decoding.py::Test_MLADecoding::test_op[cutile-512-64-torch.float16-32-True]": 0.1901744800002234,
-"tests/ops/test_mla_decoding.py::Test_MLADecoding::test_op[cutile-512-64-torch.float16-64-False]": 0.0013748809999469813,
-"tests/ops/test_mla_decoding.py::Test_MLADecoding::test_op[tilecpp-512-64-torch.float16-128-False]": 0.184531490000154,
-"tests/ops/test_mla_decoding.py::Test_MLADecoding::test_op[tilecpp-512-64-torch.float16-16-True]": 1.4238348310002493,
-"tests/ops/test_mla_decoding.py::Test_MLADecoding::test_op[tilecpp-512-64-torch.float16-32-True]": 0.2481551829998807,
-"tests/ops/test_mla_decoding.py::Test_MLADecoding::test_op[tilecpp-512-64-torch.float16-64-False]": 0.1949527140002374,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-1024-16]": 5.111274784999978,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-1024-24]": 0.3534702060001109,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-1024-32]": 0.41331570999977885,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-1024-8]": 5.141728450999835,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-11049-16]": 4.298023443999909,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-11049-24]": 0.5759332840000297,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-11049-32]": 0.592321634999962,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-11049-8]": 4.281424561999984,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-129-16]": 4.106049649999932,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-129-24]": 0.41871581000032165,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-129-32]": 0.37227775299970745,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-129-8]": 4.008528389000048,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-8192-16]": 5.21658285500007,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-8192-24]": 0.5466707989999122,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-8192-32]": 0.5083783029995175,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-8192-8]": 5.209214628999916,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-1024-16]": 5.402807623000172,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-1024-24]": 0.3764936580000722,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-1024-32]": 0.3417966939998678,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-1024-8]": 5.39418216099989,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-11049-16]": 4.444225918000029,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-11049-24]": 0.4716659189998609,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-11049-32]": 0.43988633000003574,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-11049-8]": 4.401701700000103,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-129-16]": 4.372601151999788,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-129-24]": 4.1178118229997835,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-129-32]": 4.163751423999884,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-129-8]": 4.1696929400000045,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-8192-16]": 5.34822609999992,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-8192-24]": 0.44554770700005975,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-8192-32]": 0.4111638379999931,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-8192-8]": 5.410659229999965,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-1024-16]": 9.229566682000012,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-1024-24]": 9.135035447000007,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-1024-32]": 8.626214437999977,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-1024-8]": 9.186324478000017,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-11049-16]": 11.025601180000024,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-11049-24]": 10.919905671999999,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-11049-32]": 11.102441306000003,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-11049-8]": 11.017078188,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-129-16]": 11.656021877000057,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-129-24]": 11.34404866999995,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-129-32]": 10.773276414000009,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-129-8]": 11.841301876999978,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-8192-16]": 9.205838904000018,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-8192-24]": 9.293325793000008,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-8192-32]": 9.240028266999985,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-8192-8]": 9.310271812999957,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-1024-16]": 8.510449563000009,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-1024-24]": 8.997751401000016,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-1024-32]": 8.573211183000012,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-1024-8]": 9.13697672099994,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-11049-16]": 10.756421154000009,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-11049-24]": 10.78809905299994,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-11049-32]": 10.749918768999976,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-11049-8]": 10.68771011299998,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-129-16]": 10.24273523900007,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-129-24]": 10.770225883000023,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-129-32]": 10.237921552999978,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-129-8]": 10.833573743000045,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-8192-16]": 8.656009664999942,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-8192-24]": 9.112559836000003,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-8192-32]": 8.466501827999991,
-"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-8192-8]": 9.07810792700002,
-"tests/ops/test_moe.py::Test_MOE::test_op[cutile-torch.bfloat16-16-512-256-8-2]": 3.3310526930000037,
-"tests/ops/test_moe.py::Test_MOE::test_op[cutile-torch.bfloat16-32-1024-512-16-4]": 3.0324626290000083,
-"tests/ops/test_moe.py::Test_MOE::test_op[cutile-torch.bfloat16-64-1024-512-20-8]": 2.3785154160000843,
-"tests/ops/test_moe.py::Test_MOE::test_op[cutile-torch.float8_e4m3fn-16-512-256-8-2]": 0.001043299000002662,
-"tests/ops/test_moe.py::Test_MOE::test_op[cutile-torch.float8_e4m3fn-32-1024-512-16-4]": 0.004343569000070602,
-"tests/ops/test_moe.py::Test_MOE::test_op[cutile-torch.float8_e4m3fn-64-1024-512-20-8]": 0.000974563999932343,
-"tests/ops/test_moe.py::Test_MOE::test_op[tilecpp-torch.bfloat16-16-512-256-8-2]": 4.471029261000012,
-"tests/ops/test_moe.py::Test_MOE::test_op[tilecpp-torch.bfloat16-32-1024-512-16-4]": 4.344379652999976,
-"tests/ops/test_moe.py::Test_MOE::test_op[tilecpp-torch.bfloat16-64-1024-512-20-8]": 4.000724627000011,
-"tests/ops/test_moe.py::Test_MOE::test_op[tilecpp-torch.float8_e4m3fn-16-512-256-8-2]": 3.1618766749999736,
-"tests/ops/test_moe.py::Test_MOE::test_op[tilecpp-torch.float8_e4m3fn-32-1024-512-16-4]": 2.6574041160000093,
-"tests/ops/test_moe.py::Test_MOE::test_op[tilecpp-torch.float8_e4m3fn-64-1024-512-20-8]": 2.616043191000017,
-"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-2048-4096-1024-64-4-swiglu]": 0.25567506199996615,
-"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-2048-1408-60-4-swiglu]": 3.6229077520000033,
-"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-1-swiglu]": 0.24448472300002777,
-"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-2-swiglu]": 0.24423519100008662,
-"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-4-geglu]": 4.193745909000029,
-"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-4-gelu]": 3.8971774409999966,
-"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-4-reglu]": 3.2698521379999193,
-"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-4-relu]": 7.562132537000082,
-"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-4-relu_sq]": 2.8426037949999454,
-"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-4-silu]": 3.114225627000053,
-"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-4-swiglu]": 3.7158022960000494,
-"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-512-4096-1024-64-4-swiglu]": 0.27245957100001306,
-"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-512-4096-14336-8-2-swiglu]": 3.4785575760000143,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-K64_V128-bf16]": 0.0944921919999615,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-K64_V128-fp32]": 0.11989236199997322,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-T16-bf16]": 0.7590508840000325,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-T16-fp32]": 0.7022301830000401,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-T32_final_state-bf16]": 0.5697278220000044,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-T32_final_state-fp32]": 0.5258810839999342,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-T64-bf16]": 0.8684368080000695,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-T64-fp32]": 0.7964047049999863,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-decode-bf16]": 0.3758204299999761,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-decode-fp32]": 0.44204002100002526,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-init_final_state-bf16]": 0.8966458699999293,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-init_final_state-fp32]": 0.8668302309999376,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-l2norm-bf16]": 5.270891509999956,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-l2norm-fp32]": 4.935244673,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-minimal-bf16]": 0.3588850020000791,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-minimal-fp32]": 0.35174359699999513,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-qwen35_cached_decode-bf16]": 0.4550857750000432,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-qwen35_cached_decode-fp32]": 0.4618742570000336,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-K64_V128-bf16]": 0.08981612699994912,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-K64_V128-fp32]": 0.11608774400019684,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-T16-bf16]": 0.111798244000056,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-T16-fp32]": 0.13481082200007677,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-T32_final_state-bf16]": 331.2584225370001,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-T32_final_state-fp32]": 344.613440128,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-T64-bf16]": 0.11896196199995757,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-T64-fp32]": 0.11737765499992747,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-decode-bf16]": 10.170637780999982,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-decode-fp32]": 10.317554508999933,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-init_final_state-bf16]": 1.1770028390001244,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-init_final_state-fp32]": 1.0601740890001565,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-l2norm-bf16]": 170.04219274799993,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-l2norm-fp32]": 173.5755498860001,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-minimal-bf16]": 0.08769486300002427,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-minimal-fp32]": 0.1150568209998255,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-qwen35_cached_decode-bf16]": 1.99039623200008,
-"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-qwen35_cached_decode-fp32]": 2.083069488000092,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-None-256-18432-torch.float16]": 0.0012145189998591377,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-None-256-256-torch.float16]": 0.1696859180001411,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-None-256-256-torch.float32]": 0.2359799460000147,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-None-256-768-torch.float16]": 0.12880516200038983,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-None-31072-4096-torch.bfloat16]": 1.4639302070000042,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-None-4096-256-torch.bfloat16]": 0.1132105149997642,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_cached-256-18432-torch.float16]": 0.4872019570000248,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_cached-256-256-torch.float16]": 0.21862747899999135,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_cached-256-256-torch.float32]": 0.24752594900002123,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_cached-256-768-torch.float16]": 0.24735752400002298,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_cached-31072-4096-torch.bfloat16]": 0.3011879190000002,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_cached-4096-256-torch.bfloat16]": 0.21710307500001136,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_reload-256-18432-torch.float16]": 0.29643030900001577,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_reload-256-256-torch.float16]": 0.24406993800002397,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_reload-256-256-torch.float32]": 0.13890164800002935,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_reload-256-768-torch.float16]": 0.23419750399997952,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_reload-31072-4096-torch.bfloat16]": 0.323552253999992,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_reload-4096-256-torch.bfloat16]": 0.2240291719999732,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-static_persistent-256-18432-torch.float16]": 0.0011949980000451887,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-static_persistent-256-256-torch.float16]": 0.37197888000000034,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-static_persistent-256-256-torch.float32]": 1.0194483619999914,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-static_persistent-256-768-torch.float16]": 0.47905888299999333,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-static_persistent-31072-4096-torch.bfloat16]": 0.19357722199998761,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-static_persistent-4096-256-torch.bfloat16]": 0.32574264699999844,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-None-256-18432-torch.float16]": 0.002929114999972171,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-None-256-256-torch.float16]": 0.5408393470000021,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-None-256-256-torch.float32]": 0.6084590489999755,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-None-256-768-torch.float16]": 0.5846969469999976,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-None-31072-4096-torch.bfloat16]": 0.8514233310000066,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-None-4096-256-torch.bfloat16]": 0.6863244209999948,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_cached-256-18432-torch.float16]": 1.3319542580000245,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_cached-256-256-torch.float16]": 0.6493585289999828,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_cached-256-256-torch.float32]": 0.6087032510000085,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_cached-256-768-torch.float16]": 0.6259870350000085,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_cached-31072-4096-torch.bfloat16]": 0.7394776460000401,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_cached-4096-256-torch.bfloat16]": 0.6123646499999893,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_reload-256-18432-torch.float16]": 0.6224892939999904,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_reload-256-256-torch.float16]": 0.12179033900002878,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_reload-256-256-torch.float32]": 0.13575751100000844,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_reload-256-768-torch.float16]": 0.10379648599999314,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_reload-31072-4096-torch.bfloat16]": 0.6483690700000295,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_reload-4096-256-torch.bfloat16]": 0.6563800810000089,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-static_persistent-256-18432-torch.float16]": 0.0008712070000171934,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-static_persistent-256-256-torch.float16]": 0.001057717999998431,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-static_persistent-256-256-torch.float32]": 0.0008653339999966647,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-static_persistent-256-768-torch.float16]": 0.0008966770000142787,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-static_persistent-31072-4096-torch.bfloat16]": 0.0008580689999746482,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-static_persistent-4096-256-torch.bfloat16]": 0.0010249529999839524,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-None-256-18432-torch.float16]": 0.0011245509999753267,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-None-256-256-torch.float16]": 0.8841533060000017,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-None-256-256-torch.float32]": 0.23945403300001544,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-None-256-768-torch.float16]": 0.21268305399999576,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-None-31072-4096-torch.bfloat16]": 30.827323769999964,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-None-4096-256-torch.bfloat16]": 13.928867408000002,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_cached-256-18432-torch.float16]": 0.0013475350000078379,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_cached-256-256-torch.float16]": 0.0014474250000091615,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_cached-256-256-torch.float32]": 0.002251696000001857,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_cached-256-768-torch.float16]": 0.0014555979999784086,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_cached-31072-4096-torch.bfloat16]": 0.001312339000008933,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_cached-4096-256-torch.bfloat16]": 0.0013641810000137866,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_reload-256-18432-torch.float16]": 0.26333220500001175,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_reload-256-256-torch.float16]": 0.12676082199999428,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_reload-256-256-torch.float32]": 0.18360238200000367,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_reload-256-768-torch.float16]": 0.10419208799999069,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_reload-31072-4096-torch.bfloat16]": 0.28105024699999603,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_reload-4096-256-torch.bfloat16]": 0.217673321999996,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-static_persistent-256-18432-torch.float16]": 0.0011155560000304376,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-static_persistent-256-256-torch.float16]": 14.766655014000008,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-static_persistent-256-256-torch.float32]": 19.013794836999978,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-static_persistent-256-768-torch.float16]": 19.435050625000002,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-static_persistent-31072-4096-torch.bfloat16]": 0.1415303659999836,
-"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-static_persistent-4096-256-torch.bfloat16]": 0.13187533399997164,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-1-128-32-32-128-0.5]": 0.0008799309999858451,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-1-128-32-32-64-1.0]": 0.000987414999997327,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-1-128-32-8-256-0.25]": 0.0009195720000150231,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-1-128-32-8-64-1.0]": 0.0010028190000355153,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-2-128-32-32-64-1.0]": 0.0017341329999851496,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-2-128-32-8-128-0.5]": 0.0010293119999857936,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-2-128-32-8-256-0.25]": 0.0008811890000117728,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-2-128-32-8-64-1.0]": 0.0008969820000004347,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-3-423-73-155-92-1.0]": 0.00030910299997799484,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-3-423-73-213-92-1.0]": 0.00033604499998318715,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-1-128-32-32-128-0.5]": 0.25279559800000584,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-1-128-32-32-64-1.0]": 0.42817829400001983,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-1-128-32-8-256-0.25]": 0.3109690789999604,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-1-128-32-8-64-1.0]": 0.30127940399998465,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-2-128-32-32-64-1.0]": 0.959187959000019,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-2-128-32-8-128-0.5]": 0.2660131149999927,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-2-128-32-8-256-0.25]": 0.3156247139999664,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-2-128-32-8-64-1.0]": 0.31483996899999056,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-3-423-73-155-92-1.0]": 0.00032911200000285135,
-"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-3-423-73-213-92-1.0]": 0.00037700800001516654,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-1-128-32-32-128-0.5]": 0.001500221999947371,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-1-128-32-32-64-1.0]": 0.0015757480000218038,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-1-128-32-8-256-0.25]": 0.001345897999982526,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-1-128-32-8-64-1.0]": 0.002476546000025337,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-2-128-32-32-64-1.0]": 0.0014751840000144512,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-2-128-32-8-128-0.5]": 0.0014223690000108036,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-2-128-32-8-256-0.25]": 0.0013373179999973672,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-2-128-32-8-64-1.0]": 0.0014512409999838383,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-3-423-73-155-92-1.0]": 0.0005923900000084359,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-3-423-73-213-92-1.0]": 0.0005578509999963899,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-1-128-32-32-128-0.5]": 0.7062448290000134,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-1-128-32-32-64-1.0]": 0.6725436190000096,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-1-128-32-8-256-0.25]": 0.6990158290000181,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-1-128-32-8-64-1.0]": 0.6935691899999767,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-2-128-32-32-64-1.0]": 0.656579258999983,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-2-128-32-8-128-0.5]": 0.6846208380000007,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-2-128-32-8-256-0.25]": 0.6811236159999794,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-2-128-32-8-64-1.0]": 0.689695592000021,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-3-423-73-155-92-1.0]": 0.00046724999998559724,
-"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-3-423-73-213-92-1.0]": 0.00047745500000928587,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-1-128-32-32-128-0.5]": 0.000981827000003932,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-1-128-32-32-64-1.0]": 0.005015060999994603,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-1-128-32-8-256-0.25]": 0.0009932870000000094,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-1-128-32-8-64-1.0]": 0.002346565999999939,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-2-128-32-32-64-1.0]": 0.0033231320000197684,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-2-128-32-8-128-0.5]": 0.0008913529999858838,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-2-128-32-8-256-0.25]": 0.0010305999999786764,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-2-128-32-8-64-1.0]": 0.0009540780000349969,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-3-423-73-155-92-1.0]": 0.00035471199998937664,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-3-423-73-213-92-1.0]": 0.0003317740000170488,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-1-128-32-32-128-0.5]": 0.2582441460000098,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-1-128-32-32-64-1.0]": 0.1286133979999704,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-1-128-32-8-256-0.25]": 0.10134567199997946,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-1-128-32-8-64-1.0]": 0.39458845200005044,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-2-128-32-32-64-1.0]": 0.4515477520000104,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-2-128-32-8-128-0.5]": 0.266415453999997,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-2-128-32-8-256-0.25]": 1.5066377969999962,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-2-128-32-8-64-1.0]": 0.12740337399998225,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-3-423-73-155-92-1.0]": 0.0003252800000836942,
-"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-3-423-73-213-92-1.0]": 0.00036691700000801575,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float16-0.02-0.02-1-128-32-32-64-1.0]": 0.8323518839999906,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float16-0.02-0.02-1-128-32-8-256-0.25]": 0.9050862530000359,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float16-0.02-0.02-1-512-16-16-128-1.0]": 0.8270857569999919,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float16-0.02-0.02-2-128-32-8-128-0.5]": 0.3083214069999656,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float16-0.02-0.02-2-128-32-8-64-1.0]": 0.8938801050000222,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float32-1e_05-1e_05-1-128-32-32-64-1.0]": 1.1036504730000445,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float32-1e_05-1e_05-1-128-32-8-256-0.25]": 0.9360818880000181,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float32-1e_05-1e_05-1-512-16-16-128-1.0]": 0.9324006000000225,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float32-1e_05-1e_05-2-128-32-8-128-0.5]": 0.29601448300002176,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float32-1e_05-1e_05-2-128-32-8-64-1.0]": 0.8779124350000131,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float16-0.02-0.02-1-128-32-32-64-1.0]": 1.1606720139999993,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float16-0.02-0.02-1-128-32-8-256-0.25]": 1.2263153229999944,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float16-0.02-0.02-1-512-16-16-128-1.0]": 1.26907827200003,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float16-0.02-0.02-2-128-32-8-128-0.5]": 1.2675892170000225,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float16-0.02-0.02-2-128-32-8-64-1.0]": 1.2484200560000431,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float32-1e_05-1e_05-1-128-32-32-64-1.0]": 0.7095992449999926,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float32-1e_05-1e_05-1-128-32-8-256-0.25]": 0.7165506629999925,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float32-1e_05-1e_05-1-512-16-16-128-1.0]": 1.269967409000003,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float32-1e_05-1e_05-2-128-32-8-128-0.5]": 0.7372869099999946,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float32-1e_05-1e_05-2-128-32-8-64-1.0]": 0.7157132399999853,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float16-0.02-0.02-1-128-32-32-64-1.0]": 0.5770980860000066,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float16-0.02-0.02-1-128-32-8-256-0.25]": 0.3799968790000037,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float16-0.02-0.02-1-512-16-16-128-1.0]": 0.5858723300000008,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float16-0.02-0.02-2-128-32-8-128-0.5]": 0.3574194999999918,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float16-0.02-0.02-2-128-32-8-64-1.0]": 0.52481597299996,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float32-1e_05-1e_05-1-128-32-32-64-1.0]": 0.7385028840000132,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float32-1e_05-1e_05-1-128-32-8-256-0.25]": 0.371418704000007,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float32-1e_05-1e_05-1-512-16-16-128-1.0]": 0.5834604079999792,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float32-1e_05-1e_05-2-128-32-8-128-0.5]": 0.2592858649999812,
-"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float32-1e_05-1e_05-2-128-32-8-64-1.0]": 0.5284705159999703,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-16-1024-1024-torch.float16]": 0.1544476770000074,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-16-1024-2048-torch.float32]": 0.16053250199999525,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-16-1024-512-torch.float32]": 0.564700963000007,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-16-512-1024-torch.float32]": 0.11111388100002273,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-16-512-512-torch.float16]": 0.1280173600000012,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-32-1024-1024-torch.float32]": 0.1500453539999853,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-32-1024-4096-torch.float32]": 3.263329679000009,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-32-1024-512-torch.float16]": 0.14129447100000903,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-32-512-1024-torch.float16]": 0.1277376870000353,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-32-512-2048-torch.float32]": 0.14573377900003948,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-32-512-512-torch.float32]": 0.12669500999999173,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-8-1024-1024-torch.float32]": 0.5015671289999943,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-8-1024-2048-torch.float32]": 0.14174923500002024,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-8-1024-512-torch.float16]": 0.13078429300000494,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-8-512-1024-torch.float16]": 1.8565761000000123,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-8-512-2048-torch.float32]": 2.5198691180000026,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-8-512-512-torch.float16]": 2.03977047799998,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-16-1024-1024-torch.float16]": 0.1282323669999812,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-16-1024-2048-torch.float32]": 0.13880156500005114,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-16-1024-512-torch.float32]": 0.11508210900001359,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-16-512-1024-torch.float32]": 0.11153949400002716,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-16-512-512-torch.float16]": 0.11030644299998471,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-32-1024-1024-torch.float32]": 1.644652678,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-32-1024-4096-torch.float32]": 1.4344117000000551,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-32-1024-512-torch.float16]": 0.12830670600004623,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-32-512-1024-torch.float16]": 0.1335615450000205,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-32-512-2048-torch.float32]": 0.1378525810000042,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-32-512-512-torch.float32]": 0.1736636619999956,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-8-1024-1024-torch.float32]": 0.10914896800002793,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-8-1024-2048-torch.float32]": 0.12363685400001145,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-8-1024-512-torch.float16]": 0.12543401799999288,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-8-512-1024-torch.float16]": 1.1201716480000243,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-8-512-2048-torch.float32]": 0.6410045419999904,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-8-512-512-torch.float16]": 1.1399722439999778,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-13-256-1024-torch.float32]": 2.245818958000001,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-13-333-1500-torch.float32]": 0.5977622119999921,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-7-100-1000-torch.float32]": 0.5399959369999863,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-7-512-512-torch.float32]": 1.7718134330000055,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-8-100-512-torch.float32]": 0.1335519270000134,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-8-256-1000-torch.float32]": 2.8252676399999928,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-8-256-1500-torch.float32]": 3.073038362999995,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-8-256-3000-torch.float32]": 5.681967658000019,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-8-333-1024-torch.float32]": 0.43053188099997897,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-13-256-1024-torch.float32]": 0.1970527709999601,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-13-333-1500-torch.float32]": 0.09705926099999829,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-7-100-1000-torch.float32]": 0.12013645000001816,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-7-512-512-torch.float32]": 1.099693579999979,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-8-100-512-torch.float32]": 0.10874092399998858,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-8-256-1000-torch.float32]": 0.5753061970000033,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-8-256-1500-torch.float32]": 1.1179656559999955,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-8-256-3000-torch.float32]": 0.6512017119999882,
-"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-8-333-1024-torch.float32]": 0.09848443600000678,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-cutile-256-1009-torch.float16]": 0.7865783649999969,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-cutile-256-2048-torch.float16]": 0.8055962150000084,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-cutile-256-2048-torch.float32]": 0.7697850529999926,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-cutile-256-256-torch.float16]": 0.6564196019999997,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-cutile-256-256-torch.float32]": 0.6142878560000042,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-cutile-256-32768-torch.float32]": 1.909789910000029,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-cutile-256-9-torch.float32]": 0.704796588999983,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-tilecpp-256-1009-torch.float16]": 0.73680643900002,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-tilecpp-256-2048-torch.float16]": 0.717423843000006,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-tilecpp-256-2048-torch.float32]": 0.8053659029999949,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-tilecpp-256-256-torch.float16]": 0.7344794539999668,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-tilecpp-256-256-torch.float32]": 0.7040884609999694,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-tilecpp-256-32768-torch.float32]": 0.9588610870000309,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-tilecpp-256-9-torch.float32]": 0.7082036709999784,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-cutile-256-1009-torch.float16]": 0.4267691040000159,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-cutile-256-2048-torch.float16]": 0.4465061530000298,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-cutile-256-2048-torch.float32]": 0.4518230700000174,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-cutile-256-256-torch.float16]": 0.3366717210000445,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-cutile-256-256-torch.float32]": 0.3323490939999658,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-cutile-256-32768-torch.float32]": 0.7024051559999975,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-cutile-256-9-torch.float32]": 0.3770636549999722,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-tilecpp-256-1009-torch.float16]": 0.19114871499999708,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-tilecpp-256-2048-torch.float16]": 0.18129043499996556,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-tilecpp-256-2048-torch.float32]": 0.1883488899999861,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-tilecpp-256-256-torch.float16]": 0.2131124280000165,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-tilecpp-256-256-torch.float32]": 0.2133267369999885,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-tilecpp-256-32768-torch.float32]": 0.1905538239999771,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-tilecpp-256-9-torch.float32]": 0.2008843539999816,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-cutile-256-1009-torch.float16]": 0.36497849499997415,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-cutile-256-2048-torch.float16]": 0.3508156849999864,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-cutile-256-2048-torch.float32]": 0.3451664089999724,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-cutile-256-256-torch.float16]": 0.30490926800001716,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-cutile-256-256-torch.float32]": 0.2837287790000289,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-cutile-256-32768-torch.float32]": 0.7266430720000017,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-cutile-256-9-torch.float32]": 0.3083011279999539,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-tilecpp-256-1009-torch.float16]": 0.19135102900000334,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-tilecpp-256-2048-torch.float16]": 0.17928392899997903,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-tilecpp-256-2048-torch.float32]": 0.20950861100001816,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-tilecpp-256-256-torch.float16]": 0.18902183199998035,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-tilecpp-256-256-torch.float32]": 0.239912027000031,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-tilecpp-256-32768-torch.float32]": 0.18289522500001,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-tilecpp-256-9-torch.float32]": 0.192622077999971,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-cutile-256-1009-torch.float16]": 1.2599594309999986,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-cutile-256-2048-torch.float16]": 0.9904358259999526,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-cutile-256-2048-torch.float32]": 1.225880716000006,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-cutile-256-256-torch.float16]": 1.3301162889999887,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-cutile-256-256-torch.float32]": 1.4903591430000063,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-cutile-256-32768-torch.float32]": 4.096931289000025,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-cutile-256-9-torch.float32]": 1.326434321000022,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-tilecpp-256-1009-torch.float16]": 0.19688652099998194,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-tilecpp-256-2048-torch.float16]": 0.215477672999981,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-tilecpp-256-2048-torch.float32]": 0.17666121999999973,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-tilecpp-256-256-torch.float16]": 0.20071452100000897,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-tilecpp-256-256-torch.float32]": 0.2178864099999771,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-tilecpp-256-32768-torch.float32]": 0.18961246799997866,
-"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-tilecpp-256-9-torch.float32]": 0.19521378400003186,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-1-512-16-1]": 0.2972706339999718,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-1-512-16-2]": 0.2520950650000202,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-1-512-32-1]": 0.2624262370000281,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-1-512-32-2]": 0.3031928949999667,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-16-512-16-1]": 0.326696792000007,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-16-512-16-2]": 0.3535489360000099,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-16-512-32-1]": 0.31368875200001867,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-16-512-32-2]": 0.3221279269999968,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-2-512-16-1]": 0.2844566240000006,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-2-512-16-2]": 0.2873857630000032,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-2-512-32-1]": 0.29554731000001766,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-2-512-32-2]": 0.2891566600000033,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-8-512-16-1]": 0.29792004300000485,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-8-512-16-2]": 0.29647107099998493,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-8-512-32-1]": 0.3202111960000025,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-8-512-32-2]": 0.28815192399997613,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-1-512-16-1]": 0.26790931499999715,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-1-512-16-2]": 0.2576018459999716,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-1-512-32-1]": 0.2647199670000191,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-1-512-32-2]": 0.25549753600000713,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-16-512-16-1]": 0.3057405440000025,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-16-512-16-2]": 0.30637157699999307,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-16-512-32-1]": 0.30706461799996987,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-16-512-32-2]": 0.3207051430000263,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-2-512-16-1]": 0.2687486679999722,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-2-512-16-2]": 0.2629530460000069,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-2-512-32-1]": 0.29119628700001954,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-2-512-32-2]": 0.30879603299999303,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-8-512-16-1]": 0.28191644599999677,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-8-512-16-2]": 0.2864982540000085,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-8-512-32-1]": 0.2887532849999843,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-8-512-32-2]": 0.3107739020000224,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-1-512-16-1]": 0.6375652769999647,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-1-512-16-2]": 0.6257106299999577,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-1-512-32-1]": 0.630615382000002,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-1-512-32-2]": 0.6104170220000071,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-16-512-16-1]": 0.6669463939999787,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-16-512-16-2]": 0.6789410129999851,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-16-512-32-1]": 0.6544041580000339,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-16-512-32-2]": 0.6471595610000236,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-2-512-16-1]": 0.6598536760000115,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-2-512-16-2]": 0.6184611810000149,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-2-512-32-1]": 0.6679657209999732,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-2-512-32-2]": 0.6205526040000109,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-8-512-16-1]": 0.7280024300000036,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-8-512-16-2]": 0.6730777820000071,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-8-512-32-1]": 0.6673054739999884,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-8-512-32-2]": 0.6564612469999531,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-1-512-16-1]": 0.6803322500000206,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-1-512-16-2]": 0.6549353159999782,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-1-512-32-1]": 0.6856938950000142,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-1-512-32-2]": 0.6103896909999946,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-16-512-16-1]": 0.7281476089999899,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-16-512-16-2]": 0.7167569510000078,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-16-512-32-1]": 0.6636094979999996,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-16-512-32-2]": 0.6485378279999736,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-2-512-16-1]": 0.6276924359999896,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-2-512-16-2]": 0.6885734179999758,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-2-512-32-1]": 0.13172192200002542,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-2-512-32-2]": 0.674771021999959,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-8-512-16-1]": 0.12321420500001068,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-8-512-16-2]": 0.64477991199999,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-8-512-32-1]": 0.6553695179999863,
-"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-8-512-32-2]": 0.6604196450000188,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[cutile-2-1024-512-torch.float32]": 0.17153075199999535,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[cutile-4-512-1024-torch.float32]": 0.15966354599999022,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[cutile-8-256-2048-torch.float32]": 0.31712401200005047,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[cutile-8-512-1024-torch.float16]": 0.3497424289999742,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[cutile-8-512-512-torch.float16]": 0.36172342799997637,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[tilecpp-2-1024-512-torch.float32]": 1.1561888260000046,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[tilecpp-4-512-1024-torch.float32]": 1.0702655439999944,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[tilecpp-8-256-2048-torch.float32]": 1.1292605870000045,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[tilecpp-8-512-1024-torch.float16]": 1.151045785000008,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[tilecpp-8-512-512-torch.float16]": 1.1291592199999911,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-13-128-1024]": 0.3959966740000027,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-13-333-1500]": 0.5395875979999971,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-4-128-3000]": 0.5263445630000092,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-4-256-1000]": 0.4877196300000435,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-4-256-1500]": 0.5358027300000288,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-7-100-1000]": 0.44990856900000153,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-7-256-512]": 0.34473915300003455,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-8-100-512]": 0.12277813099996138,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-8-333-256]": 0.3099081320000039,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-13-128-1024]": 0.0986762920000217,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-13-333-1500]": 0.13156983899997954,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-4-128-3000]": 0.7251883950000035,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-4-256-1000]": 0.13374192399999174,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-4-256-1500]": 0.1187394050000421,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-7-100-1000]": 0.14279693299999963,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-7-256-512]": 0.12666253799997662,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-8-100-512]": 0.09650348200003123,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-8-333-256]": 1.0223252310000248,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward[cutile-2-512-1024-4096]": 0.14923376999999505,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward[cutile-4-1024-2048-8192]": 0.3025218830000256,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward[cutile-8-2048-4096-14336]": 0.2598767389999921,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward[tilecpp-2-512-1024-4096]": 0.12338330599999381,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward[tilecpp-4-1024-2048-8192]": 0.1530189529999859,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward[tilecpp-8-2048-4096-14336]": 0.7157162530000107,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-13-256-2048]": 0.114780278000012,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-13-333-1500]": 0.13553140099998018,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-7-100-1000]": 0.12360883100004116,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-7-512-1024]": 0.10678720600003544,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-8-100-1024]": 0.09709039900002381,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-8-256-1000]": 0.16881123199999593,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-8-256-1500]": 0.13848644999998783,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-8-256-3000]": 0.12726111200001355,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-8-333-512]": 0.19224987699999474,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-13-256-2048]": 0.11101977500001681,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-13-333-1500]": 0.12030891200001292,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-7-100-1000]": 0.11027593299996852,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-7-512-1024]": 0.11808077700001718,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-8-100-1024]": 0.10182934200000204,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-8-256-1000]": 0.09600050799997462,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-8-256-1500]": 0.0927385239999694,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-8-256-3000]": 0.09665702800000986,
-"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-8-333-512]": 0.0961453140000117,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-1024-1024-torch.float16]": 0.24652781500000742,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-1024-1024-torch.float32]": 0.2571273700000347,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-128-128-torch.float16]": 0.25852475700000355,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-128-128-torch.float32]": 0.28552432500001146,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-16-16-torch.float16]": 0.25384684699997706,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-16-16-torch.float32]": 0.26231107399999587,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-2-2-torch.float16]": 0.29460441800003423,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-2-2-torch.float32]": 0.32361879499998736,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-2048-2048-torch.float16]": 0.24727814199997056,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-2048-2048-torch.float32]": 0.2723342679999803,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-256-256-torch.float16]": 0.2369928799999741,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-256-256-torch.float32]": 0.2505765769999755,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-32-32-torch.float16]": 0.23435373000003779,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-32-32-torch.float32]": 0.24976834300002793,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-4-4-torch.float16]": 0.3053088140000568,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-4-4-torch.float32]": 0.2772383270000205,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-4096-4096-torch.float16]": 0.3369350719999886,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-4096-4096-torch.float32]": 0.2822110650000127,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-512-512-torch.float16]": 0.24468288599999255,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-512-512-torch.float32]": 0.28505478299999254,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-64-64-torch.float16]": 0.24737510100001714,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-64-64-torch.float32]": 0.2547445590000166,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-8-8-torch.float16]": 0.2896205109999812,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-8-8-torch.float32]": 0.2774279220000153,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-8192-8192-torch.float16]": 0.3507800850000251,
-"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-8192-8192-torch.float32]": 0.27921153099995877,
-"tests/suites/flashinfer/test_flashinfer_attention.py::Test_FlashInfer_DecodePaged::test_op[cutile-4-1024-128-float16]": 5.500744498000017,
-"tests/suites/flashinfer/test_flashinfer_attention.py::Test_FlashInfer_MLADecodePaged::test_op[cutile-32-4-1024-128-float16]": 0.001077948999977707,
-"tests/suites/flashinfer/test_flashinfer_attention.py::Test_FlashInfer_PrefillPaged::test_op[cutile-4-1024-128-128-float16]": 16.14699423799999,
-"tests/suites/flashinfer/test_flashinfer_attention.py::Test_FlashInfer_PrefillPaged::test_op[cutile-4-1024-192-128-float16]": 21.279657041999968,
-"tests/suites/flashinfer/test_flashinfer_attention.py::Test_FlashInfer_PrefillRagged::test_op[cutile-4-1024-128-float16]": 41.94995222699998,
-"tests/suites/flashinfer/test_flashinfer_attention.py::Test_FlashInfer_PrefillRagged::test_op[cutile-4-1024-192-float16]": 61.387417835000065,
-"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.0-0.0-False-True-torch.bfloat16-torch.bfloat16-4096-4096-4096-cutile]": 38.81290488800005,
-"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.0-0.0-False-True-torch.bfloat16-torch.bfloat16-8192-8192-8192-cutile]": 3.375743405000037,
-"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.0-0.0-False-True-torch.float16-torch.float16-4096-4096-4096-cutile]": 35.93972394899998,
-"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.0-0.0-False-True-torch.float16-torch.float16-8192-8192-8192-cutile]": 5.007637571000032,
-"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.0-0.0-False-True-torch.float8_e4m3fn-torch.bfloat16-4096-4096-4096-cutile]": 29.550462876999973,
-"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.0-0.0-False-True-torch.float8_e4m3fn-torch.bfloat16-8192-8192-8192-cutile]": 3.3172102300000574,
-"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.5-2.0-False-True-torch.bfloat16-torch.bfloat16-4096-4096-4096-cutile]": 0.037088714000049094,
-"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.5-2.0-False-True-torch.bfloat16-torch.bfloat16-8192-8192-8192-cutile]": 0.06635365500000034,
-"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.5-2.0-False-True-torch.float16-torch.float16-4096-4096-4096-cutile]": 0.03807624400002396,
-"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.5-2.0-False-True-torch.float16-torch.float16-8192-8192-8192-cutile]": 0.06601343299996643,
-"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.5-2.0-False-True-torch.float8_e4m3fn-torch.bfloat16-4096-4096-4096-cutile]": 0.03702211700004909,
-"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.5-2.0-False-True-torch.float8_e4m3fn-torch.bfloat16-8192-8192-8192-cutile]": 0.06543321800000967,
-"tests/suites/flashinfer/test_flashinfer_masked_bmm.py::Test_FlashInfer_MaskedBMM::test_op[cutile-False-False-torch.float16-6-4096-512-4096-7168]": 40.27446953399999,
-"tests/suites/flashinfer/test_flashinfer_masked_bmm.py::Test_FlashInfer_MaskedBMM::test_op[cutile-False-False-torch.float16-6-4096-512-7168-2048]": 1.5945390379998798,
-"tests/suites/flashinfer/test_flashinfer_masked_bmm.py::Test_FlashInfer_MaskedBMM::test_op[cutile-False-True-torch.float16-6-4096-512-4096-7168]": 40.15766551999991,
-"tests/suites/flashinfer/test_flashinfer_masked_bmm.py::Test_FlashInfer_MaskedBMM::test_op[cutile-False-True-torch.float16-6-4096-512-7168-2048]": 1.501094309999985,
-"tests/suites/flashinfer/test_flashinfer_masked_bmm.py::Test_FlashInfer_MaskedBMM::test_op[cutile-True-False-torch.float16-6-4096-512-4096-7168]": 43.485673185999985,
-"tests/suites/flashinfer/test_flashinfer_masked_bmm.py::Test_FlashInfer_MaskedBMM::test_op[cutile-True-False-torch.float16-6-4096-512-7168-2048]": 1.6346164129998897,
-"tests/suites/flashinfer/test_flashinfer_masked_bmm.py::Test_FlashInfer_MaskedBMM::test_op[cutile-True-True-torch.float16-6-4096-512-4096-7168]": 42.89487190299997,
-"tests/suites/flashinfer/test_flashinfer_masked_bmm.py::Test_FlashInfer_MaskedBMM::test_op[cutile-True-True-torch.float16-6-4096-512-7168-2048]": 1.5200476140000774,
-"tests/suites/flashinfer/test_flashinfer_ragged_block_scaled_bmm.py::Test_FlashInfer_RaggedBlockScaledBMM::test_op[False-True-torch.float8_e4m3fn-torch.bfloat16-2048-2048-128-16-cutile]": 4.834458009999935,
-"tests/suites/flashinfer/test_flashinfer_ragged_block_scaled_bmm.py::Test_FlashInfer_RaggedBlockScaledBMM::test_op[False-True-torch.float8_e4m3fn-torch.bfloat16-2048-2048-128-4-cutile]": 4.92752260899988,
-"tests/suites/flashinfer/test_flashinfer_ragged_block_scaled_bmm.py::Test_FlashInfer_RaggedBlockScaledBMM::test_op[False-True-torch.float8_e4m3fn-torch.bfloat16-2048-2048-512-16-cutile]": 0.4666226719999713,
-"tests/suites/flashinfer/test_flashinfer_ragged_block_scaled_bmm.py::Test_FlashInfer_RaggedBlockScaledBMM::test_op[False-True-torch.float8_e4m3fn-torch.bfloat16-2048-2048-512-4-cutile]": 0.47340543500001786,
-"tests/suites/flashinfer/test_flashinfer_ragged_block_scaled_bmm.py::Test_FlashInfer_RaggedBlockScaledBMM::test_op[False-True-torch.float8_e4m3fn-torch.bfloat16-4096-4096-128-16-cutile]": 0.4871669260000999,
-"tests/suites/flashinfer/test_flashinfer_ragged_block_scaled_bmm.py::Test_FlashInfer_RaggedBlockScaledBMM::test_op[False-True-torch.float8_e4m3fn-torch.bfloat16-4096-4096-128-4-cutile]": 0.46534517800000685,
-"tests/suites/flashinfer/test_flashinfer_ragged_block_scaled_bmm.py::Test_FlashInfer_RaggedBlockScaledBMM::test_op[False-True-torch.float8_e4m3fn-torch.bfloat16-4096-4096-512-16-cutile]": 0.5882562199999484,
-"tests/suites/flashinfer/test_flashinfer_ragged_block_scaled_bmm.py::Test_FlashInfer_RaggedBlockScaledBMM::test_op[False-True-torch.float8_e4m3fn-torch.bfloat16-4096-4096-512-4-cutile]": 0.4852029379999294,
-"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_dtypes[torch.bfloat16-4-256-256-256-cutile]": 0.00378045100012514,
-"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_dtypes[torch.float8_e4m3fn-4-256-256-256-cutile]": 14.222478249000005,
-"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_num_groups[1-256-256-256-torch.bfloat16-cutile]": 0.5854292450000003,
-"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_num_groups[4-256-256-256-torch.bfloat16-cutile]": 0.003694905000088511,
-"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_num_groups[8-256-256-256-torch.bfloat16-cutile]": 0.5585592399999086,
-"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_shapes[2-128-128-128-torch.bfloat16-False-False-cutile]": 0.0009979169999496662,
-"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_shapes[2-128-128-128-torch.bfloat16-True-False-cutile]": 0.5495889290000378,
-"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_shapes[4-256-256-256-torch.bfloat16-False-False-cutile]": 0.0009532440000157294,
-"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_shapes[4-256-256-256-torch.bfloat16-True-False-cutile]": 29.95001086800005,
-"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_shapes[4-512-512-512-torch.bfloat16-False-False-cutile]": 0.0009893319999036976,
-"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_shapes[4-512-512-512-torch.bfloat16-True-False-cutile]": 0.7069594399999914,
-"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_transpose[False-False-4-256-256-256-torch.bfloat16-cutile]": 0.0009819629999583412,
-"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_transpose[False-True-4-256-256-256-torch.bfloat16-cutile]": 0.0009049330000152622,
-"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_transpose[True-False-4-256-256-256-torch.bfloat16-cutile]": 0.0037330509999264905,
-"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_transpose[True-True-4-256-256-256-torch.bfloat16-cutile]": 0.0009019120000175462,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.float8_e4m3fn-128-2048-512]": 0.11031353199996374,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.float8_e4m3fn-128-2048-513]": 0.0023972129999947356,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.float8_e4m3fn-16-2048-512]": 0.12062581900011082,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.float8_e4m3fn-16-2048-513]": 0.003612463999843385,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.float8_e4m3fn-32-2048-512]": 0.11971032799999648,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.float8_e4m3fn-32-2048-513]": 0.0032853899999736313,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.float8_e4m3fn-64-2048-512]": 0.11281975700012481,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.float8_e4m3fn-64-2048-513]": 0.002241497999989406,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.int8-128-2048-512]": 0.11159406200010835,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.int8-128-2048-513]": 0.002332376999902408,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.int8-16-2048-512]": 0.11714265900002374,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.int8-16-2048-513]": 0.004179811999847516,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.int8-32-2048-512]": 0.09913440700006504,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.int8-32-2048-513]": 0.002265540000053079,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.int8-64-2048-512]": 0.12125481200007471,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.int8-64-2048-513]": 0.0023294789999681598,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.float8_e4m3fn-128-2048-512]": 0.02026439999985996,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.float8_e4m3fn-128-2048-513]": 0.0019206650000569425,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.float8_e4m3fn-16-2048-512]": 0.02416383999991467,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.float8_e4m3fn-16-2048-513]": 0.002192497999999432,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.float8_e4m3fn-32-2048-512]": 0.02184076800017465,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.float8_e4m3fn-32-2048-513]": 0.0023533969999789406,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.float8_e4m3fn-64-2048-512]": 0.02302743800009921,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.float8_e4m3fn-64-2048-513]": 0.0019278119999626142,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.int8-128-2048-512]": 0.02162144699991586,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.int8-128-2048-513]": 0.0018733729999667048,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.int8-16-2048-512]": 0.020984378000093784,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.int8-16-2048-513]": 0.002553160000047683,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.int8-32-2048-512]": 0.019848379000109162,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.int8-32-2048-513]": 0.0019840429998794207,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.int8-64-2048-512]": 0.021308007000016005,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.int8-64-2048-513]": 0.0019757859998890126,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.float8_e4m3fn-128-2048-512]": 0.0010872310000422658,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.float8_e4m3fn-128-2048-513]": 0.0010577760000103353,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.float8_e4m3fn-16-2048-512]": 0.0010631189999230628,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.float8_e4m3fn-16-2048-513]": 0.0027575510000019676,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.float8_e4m3fn-32-2048-512]": 0.002071383999975751,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.float8_e4m3fn-32-2048-513]": 0.0020125770000163357,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.float8_e4m3fn-64-2048-512]": 0.00205754300000649,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.float8_e4m3fn-64-2048-513]": 0.002016586999957326,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.int8-128-2048-512]": 0.001042043000012427,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.int8-128-2048-513]": 0.0010563449999949626,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.int8-16-2048-512]": 0.0010870129999887013,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.int8-16-2048-513]": 0.0013006930000187822,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.int8-32-2048-512]": 0.0034817000000373355,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.int8-32-2048-513]": 0.0011082999999985077,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.int8-64-2048-512]": 0.0011805799999535793,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.int8-64-2048-513]": 0.0010683029999825067,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.float8_e4m3fn-128-2048-512]": 0.1298981079999919,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.float8_e4m3fn-128-2048-513]": 0.004401974999979075,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.float8_e4m3fn-16-2048-512]": 0.11796498200004635,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.float8_e4m3fn-16-2048-513]": 0.0023944670000162205,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.float8_e4m3fn-32-2048-512]": 0.12101731199999222,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.float8_e4m3fn-32-2048-513]": 0.0025874160000114443,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.float8_e4m3fn-64-2048-512]": 0.12139068400000497,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.float8_e4m3fn-64-2048-513]": 0.0023390740000195365,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.int8-128-2048-512]": 0.12379594399999405,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.int8-128-2048-513]": 0.00319128599997498,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.int8-16-2048-512]": 0.1089112140000168,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.int8-16-2048-513]": 0.002963977999996814,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.int8-32-2048-512]": 0.11734048999997526,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.int8-32-2048-513]": 0.0027460629999893627,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.int8-64-2048-512]": 0.11968220299999643,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.int8-64-2048-513]": 0.0027529810000146426,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.float8_e4m3fn-128-2048-512]": 0.0012701199999867185,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.float8_e4m3fn-128-2048-513]": 0.001271915000017998,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.float8_e4m3fn-16-2048-512]": 0.0016126639999924919,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.float8_e4m3fn-16-2048-513]": 0.0013735240000016802,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.float8_e4m3fn-32-2048-512]": 0.0013160500000140019,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.float8_e4m3fn-32-2048-513]": 0.0013018880000004174,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.float8_e4m3fn-64-2048-512]": 0.0012834429999770691,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.float8_e4m3fn-64-2048-513]": 0.001297222999994574,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.int8-128-2048-512]": 0.0013615310000147929,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.int8-128-2048-513]": 0.00105126599996197,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.int8-16-2048-512]": 0.0012839030000293405,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.int8-16-2048-513]": 0.0012751699999853372,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.int8-32-2048-512]": 0.0013698189999900023,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.int8-32-2048-513]": 0.0013735349999990376,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.int8-64-2048-512]": 0.0013836549999837189,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.int8-64-2048-513]": 0.0013854649999984758,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.float8_e4m3fn-128-2048-512]": 0.0025042460000292976,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.float8_e4m3fn-128-2048-513]": 0.0026968729999907737,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.float8_e4m3fn-16-2048-512]": 0.00888015000001019,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.float8_e4m3fn-16-2048-513]": 0.002734170999985963,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.float8_e4m3fn-32-2048-512]": 0.0025781469999799356,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.float8_e4m3fn-32-2048-513]": 0.0029527800000153093,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.float8_e4m3fn-64-2048-512]": 0.002943780999999035,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.float8_e4m3fn-64-2048-513]": 0.00249038700002302,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.int8-128-2048-512]": 0.0023854279999966366,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.int8-128-2048-513]": 0.002409089999986236,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.int8-16-2048-512]": 0.002632204000008187,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.int8-16-2048-513]": 0.0023943790000089393,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.int8-32-2048-512]": 0.002419587000019874,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.int8-32-2048-513]": 0.0030364639999902465,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.int8-64-2048-512]": 0.0024048010000115028,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.int8-64-2048-513]": 0.002394336999969937,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.float8_e4m3fn-128-2048-512]": 0.0013937790000113637,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.float8_e4m3fn-128-2048-513]": 0.0010519719999990684,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.float8_e4m3fn-16-2048-512]": 0.0012701400000025842,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.float8_e4m3fn-16-2048-513]": 0.0012761620000105722,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.float8_e4m3fn-32-2048-512]": 0.0013099839999881624,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.float8_e4m3fn-32-2048-513]": 0.00129781400002571,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.float8_e4m3fn-64-2048-512]": 0.0012997200000199882,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.float8_e4m3fn-64-2048-513]": 0.0010013710000009723,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.int8-128-2048-512]": 0.0009465770000076645,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.int8-128-2048-513]": 0.0009526620000031016,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.int8-16-2048-512]": 0.0010229710000260184,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.int8-16-2048-513]": 0.0009920119999833332,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.int8-32-2048-512]": 0.0009955710000042473,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.int8-32-2048-513]": 0.0009588309999912781,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.int8-64-2048-512]": 0.0009489710000423202,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.int8-64-2048-513]": 0.0009327989999974307,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.float8_e4m3fn-128-2048-512]": 0.002599164999992354,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.float8_e4m3fn-128-2048-513]": 0.00270208099999536,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.float8_e4m3fn-16-2048-512]": 0.0019744709999827137,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.float8_e4m3fn-16-2048-513]": 0.002059626000004755,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.float8_e4m3fn-32-2048-512]": 0.002003030999986777,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.float8_e4m3fn-32-2048-513]": 0.0019698589999563865,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.float8_e4m3fn-64-2048-512]": 0.0029643350000014834,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.float8_e4m3fn-64-2048-513]": 0.0026759349999849746,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.int8-128-2048-512]": 0.0034896889999629366,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.int8-128-2048-513]": 0.00343791499997792,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.int8-16-2048-512]": 0.0028009099999906084,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.int8-16-2048-513]": 0.002516138000032697,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.int8-32-2048-512]": 0.0025233710000236442,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.int8-32-2048-513]": 0.006905336999977862,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.int8-64-2048-512]": 0.00845112999999742,
-"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.int8-64-2048-513]": 0.0019849740000097427,
-"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e4m3fn-torch.bfloat16-False-10000-4096-64-576-128-128]": 4.422321829999987,
-"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e4m3fn-torch.bfloat16-False-10000-4096-64-576-128-19]": 4.429476450999999,
-"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e4m3fn-torch.bfloat16-False-10000-4096-64-576-128-1]": 6.093136446999978,
-"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e4m3fn-torch.float16-False-10000-4096-64-576-128-128]": 4.46835468300003,
-"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e4m3fn-torch.float16-False-10000-4096-64-576-128-19]": 4.321749486000016,
-"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e4m3fn-torch.float16-False-10000-4096-64-576-128-1]": 6.773969932,
-"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e5m2-torch.bfloat16-False-10000-4096-64-576-128-128]": 4.402505599999984,
-"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e5m2-torch.bfloat16-False-10000-4096-64-576-128-19]": 4.293014835999941,
-"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e5m2-torch.bfloat16-False-10000-4096-64-576-128-1]": 6.036609229999954,
-"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e5m2-torch.float16-False-10000-4096-64-576-128-128]": 4.296337498000014,
-"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e5m2-torch.float16-False-10000-4096-64-576-128-19]": 4.286224460999989,
-"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e5m2-torch.float16-False-10000-4096-64-576-128-1]": 6.497528299999999,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.0-4-128-torch.bfloat16]": 0.379818145999991,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.0-4-128-torch.float16]": 0.3648856639999849,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.0-6-300-torch.float32]": 0.40317631700003176,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.0-63-41-torch.float32]": 0.3609889690000614,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.0-8-512-torch.float32]": 3.3540624800000387,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.1-4-128-torch.bfloat16]": 0.4015830079999887,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.1-4-128-torch.float16]": 0.409809078999956,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.1-6-300-torch.float32]": 0.4099561350000158,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.1-63-41-torch.float32]": 0.39212041899997985,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.1-8-512-torch.float32]": 0.6954730529999438,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-mean-1269-32000-torch.float32]": 3.160911125000041,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-mean-4-128-torch.bfloat16]": 2.570868039000061,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-mean-4-128-torch.float16]": 2.7095592389999865,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-mean-4-256-torch.float32]": 3.181209410000065,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-mean-6-300-torch.float32]": 1.721493301999942,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-mean-63-41-torch.float32]": 3.5110982759999843,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-mean-8-1024-torch.float32]": 2.194189193999989,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-sum-1269-32000-torch.float32]": 0.5789274429999978,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-sum-4-128-torch.bfloat16]": 0.30681820900002776,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-sum-4-128-torch.float16]": 0.35131119499999386,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-sum-4-256-torch.float32]": 0.3075265970000487,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-sum-6-300-torch.float32]": 0.30609461599999577,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-sum-63-41-torch.float32]": 0.3133414809999522,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-sum-8-1024-torch.float32]": 0.45131322200001023,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_ignore_index[cutile-4-256-torch.float32]": 0.39616800800001784,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_ignore_index[cutile-8-512-torch.float32]": 0.13955819800003155,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_predicted_tokens[cutile-4-128-torch.float32]": 0.3807122290000393,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_predicted_tokens[cutile-8-256-torch.float32]": 0.32133944800000336,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_reduction_none[cutile-4-128-torch.bfloat16]": 0.11736035600000605,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_reduction_none[cutile-4-128-torch.float16]": 0.13533503799993696,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_reduction_none[cutile-8-512-torch.float32]": 0.4786525860000097,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_softcap[cutile-30.0-4-128-torch.float32]": 0.558842664999986,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_softcap[cutile-30.0-8-512-torch.float32]": 0.5754022849999387,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_softcap[cutile-50.0-4-128-torch.float32]": 0.12197706400002062,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_softcap[cutile-50.0-8-512-torch.float32]": 0.14610924699996986,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_token_accuracy[cutile-4-128-torch.float32]": 0.3175275949998877,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_token_accuracy[cutile-8-256-torch.float32]": 0.3510554390000493,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_weight[cutile-4-128-torch.bfloat16]": 0.342138293000005,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_weight[cutile-8-512-torch.float32]": 0.5401239070000088,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_z_loss[cutile-0.0001-4-128-torch.float32]": 0.28947746000005736,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_z_loss[cutile-0.0001-8-512-torch.float32]": 0.49618863500001,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_z_loss[cutile-0.01-4-128-torch.float32]": 0.29698428200003946,
-"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_z_loss[cutile-0.01-8-512-torch.float32]": 0.49596251399998437,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(1, 1, 1023)-torch.float32]": 0.2635048359999814,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(2, 4, 512)-torch.float32]": 0.21429491100008136,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(2, 8, 4096)-torch.float32]": 0.270570189999944,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(3, 7, 256)-torch.float32]": 0.12626309599994556,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(4, 16, 2048)-torch.float32]": 0.23967090900003996,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(4, 256)-torch.bfloat16]": 0.2197506559999738,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(4, 256)-torch.float16]": 0.2158598680000523,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(4, 256)-torch.float32]": 0.21288227700000562,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(4, 300)-torch.float32]": 0.26061775300001955,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(1, 1, 1023)-torch.float32]": 0.2861739650000459,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(2, 4, 512)-torch.float32]": 0.22894521400002077,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(2, 8, 4096)-torch.float32]": 0.26558614800001124,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(3, 7, 256)-torch.float32]": 0.2639187770000717,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(4, 16, 2048)-torch.float32]": 0.2630981199999951,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(4, 256)-torch.bfloat16]": 0.21531280500005323,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(4, 256)-torch.float16]": 0.14850756499998852,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(4, 256)-torch.float32]": 0.15399357699993743,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(4, 300)-torch.float32]": 0.27627422899996645,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-False-(1, 1, 1023)-torch.float32]": 0.32732264000003397,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-False-(2, 8, 4096)-torch.float32]": 0.3082617750000054,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-False-(3, 7, 256)-torch.float32]": 0.2963613359999613,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-False-(4, 16, 2048)-torch.float32]": 0.2942472840000505,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-False-(4, 256)-torch.bfloat16]": 0.26688676100008024,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-False-(4, 256)-torch.float16]": 0.2782328429999552,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-False-(4, 256)-torch.float32]": 0.26916963899998336,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-False-(8, 512)-torch.float32]": 0.2800937689999614,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-True-(1, 1, 1023)-torch.float32]": 0.36573328300005414,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-True-(2, 8, 4096)-torch.float32]": 0.31654005400002916,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-True-(3, 7, 256)-torch.float32]": 0.30947869999999966,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-True-(4, 16, 2048)-torch.float32]": 0.29787492699995255,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-True-(4, 256)-torch.bfloat16]": 0.2872045139999386,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-True-(4, 256)-torch.float16]": 0.37859116100008805,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-True-(4, 256)-torch.float32]": 0.26870452500008923,
-"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-True-(8, 512)-torch.float32]": 0.30378763700002764,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-0.0-(2, 128, 512)-torch.float32]": 0.2815592980000474,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-0.0-(4, 256)-torch.bfloat16]": 0.40811988000007204,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-0.0-(4, 256)-torch.float16]": 0.4282282649999729,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-0.0-(4, 256)-torch.float32]": 0.4210465610000256,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-0.0-(8, 512)-torch.float32]": 0.40686332500001754,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-0.0-(9, 7, 41)-torch.float32]": 0.3927937610000072,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-1.0-(2, 128, 512)-torch.float32]": 0.3209429869999667,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-1.0-(4, 256)-torch.bfloat16]": 0.30116308900005606,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-1.0-(4, 256)-torch.float16]": 0.2818374539999695,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-1.0-(4, 256)-torch.float32]": 0.2954302819999839,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-1.0-(8, 512)-torch.float32]": 0.29372613600003206,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-1.0-(9, 7, 41)-torch.float32]": 0.3010073989999569,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_casting_modes[cutile-gemma]": 1.0009266440000033,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_casting_modes[cutile-llama]": 0.12444230800002742,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_casting_modes[cutile-none]": 1.0769020800000249,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(16, 1024)-torch.float32]": 1.266628039000011,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(2, 128, 512)-torch.float32]": 0.2574286709999569,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(2, 4, 512)-torch.float32]": 0.13434257899996283,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(4, 256)-torch.bfloat16]": 1.0055197759999714,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(4, 256)-torch.float16]": 0.996433173000014,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(4, 256)-torch.float32]": 0.9890289969999913,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(4, 300)-torch.float32]": 1.2910279889999288,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(8, 512)-torch.float32]": 1.0407075889999646,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(9, 7, 41)-torch.float32]": 1.0268256179999753,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(16, 1024)-torch.float32]": 0.27732648500000323,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(2, 128, 512)-torch.float32]": 0.2819659050000496,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(2, 4, 512)-torch.float32]": 0.12178856400004179,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(4, 256)-torch.bfloat16]": 0.2459941600000093,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(4, 256)-torch.float16]": 0.3390708750000613,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(4, 256)-torch.float32]": 0.23497518000004902,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(4, 300)-torch.float32]": 0.25844820700007176,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(8, 512)-torch.float32]": 0.26157047799995325,
-"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(9, 7, 41)-torch.float32]": 0.2567524780000099,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_backward[cutile-4-64-128-torch.bfloat16]": 0.3749721660000205,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_backward[cutile-4-64-128-torch.float16]": 0.38645493100000294,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_backward[cutile-4-64-128-torch.float32]": 0.3599989219999884,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_backward[cutile-63-41-41-torch.float32]": 0.3509176039999602,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_backward[cutile-8-128-256-torch.float32]": 0.3935359469999753,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_features[cutile]": 0.8931195850000222,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-mean-1024-1024-4096-torch.float32]": 2.428276935000042,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-mean-4-64-128-torch.bfloat16]": 0.6381496719999404,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-mean-4-64-128-torch.float16]": 0.6213224749999426,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-mean-4-64-128-torch.float32]": 3.3926111909999577,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-mean-6-64-150-torch.float32]": 2.2455626359999883,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-mean-63-41-41-torch.float32]": 0.5428794389999325,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-mean-8-128-256-torch.float32]": 0.4926945290000617,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-sum-1024-1024-4096-torch.float32]": 0.4546600199999489,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-sum-4-64-128-torch.bfloat16]": 0.30030336400000124,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-sum-4-64-128-torch.float16]": 0.35093713000009075,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-sum-4-64-128-torch.float32]": 0.30250346699995134,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-sum-6-64-150-torch.float32]": 0.3194214470000247,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-sum-63-41-41-torch.float32]": 0.33141202400003067,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-sum-8-128-256-torch.float32]": 0.29352554700000155,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_with_bias[cutile-4-64-128-torch.float16]": 0.10442059199999676,
-"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_with_bias[cutile-4-64-128-torch.float32]": 0.12437336799996501,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_backward[cutile-1-64-128-torch.float32]": 0.13061373199997206,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_backward[cutile-188-31-123-torch.float32]": 0.10401794000000564,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_backward[cutile-4-64-128-torch.bfloat16]": 0.14585906199999954,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_backward[cutile-4-64-128-torch.float16]": 0.14507258699995873,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_backward[cutile-4-64-128-torch.float32]": 0.15756124499995394,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_backward[cutile-63-41-41-torch.float32]": 0.10084613899999795,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_backward[cutile-8-128-256-torch.float32]": 0.11557211900003495,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-1-64-128-torch.float32]": 0.1296558839999875,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-188-31-123-torch.float32]": 0.46081631099997367,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-3-7-11-torch.float32]": 0.3818288779999648,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-4-64-128-torch.bfloat16]": 0.11006614899997658,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-4-64-128-torch.float16]": 0.12957521699996732,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-4-64-128-torch.float32]": 0.2832122000000368,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-4-64-32768-torch.float32]": 0.6164528519999521,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-63-41-41-torch.float32]": 0.24700828499993577,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-8-128-256-torch.float32]": 0.2838281040000652,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-1-64-128-torch.float32]": 0.13623359199993956,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-188-31-123-torch.float32]": 0.40850846299997556,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-3-7-11-torch.float32]": 0.3921210450000103,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-4-64-128-torch.bfloat16]": 0.10292532300002222,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-4-64-128-torch.float16]": 0.12865193900000804,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-4-64-128-torch.float32]": 0.270669923000014,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-4-64-32768-torch.float32]": 0.6206561860000193,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-63-41-41-torch.float32]": 0.23142654800000173,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-8-128-256-torch.float32]": 0.2958168709999427,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-0.5-0.2-1024-1024-4096-torch.bfloat16]": 0.7059056290000285,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-0.5-0.2-1024-1024-4096-torch.float32]": 0.15015266200003907,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-0.5-0.2-141-31-123-torch.bfloat16]": 0.4912041709999926,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-0.5-0.2-141-31-123-torch.float32]": 0.13825898399994685,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-1.0-0.5-1024-1024-4096-torch.bfloat16]": 0.6781390420000548,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-1.0-0.5-1024-1024-4096-torch.float32]": 0.13063072199997805,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-1.0-0.5-141-31-123-torch.bfloat16]": 0.4177853599999821,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-1.0-0.5-141-31-123-torch.float32]": 0.12946166099999346,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-2.0-0.8-1024-1024-4096-torch.bfloat16]": 0.7515545449999763,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-2.0-0.8-1024-1024-4096-torch.float32]": 0.13275465000003805,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-2.0-0.8-141-31-123-torch.bfloat16]": 0.43759040399999094,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-2.0-0.8-141-31-123-torch.float32]": 0.13328161500004398,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-0.5-0.2-1024-1024-4096-torch.bfloat16]": 0.692591779000054,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-0.5-0.2-1024-1024-4096-torch.float32]": 0.1290601779999747,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-0.5-0.2-141-31-123-torch.bfloat16]": 0.4456057560000204,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-0.5-0.2-141-31-123-torch.float32]": 0.1390929039999378,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-1.0-0.5-1024-1024-4096-torch.bfloat16]": 0.7129217930000777,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-1.0-0.5-1024-1024-4096-torch.float32]": 0.13821379299997716,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-1.0-0.5-141-31-123-torch.bfloat16]": 0.4507779329999835,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-1.0-0.5-141-31-123-torch.float32]": 0.15111299500006226,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-2.0-0.8-1024-1024-4096-torch.bfloat16]": 0.7616332710000506,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-2.0-0.8-1024-1024-4096-torch.float32]": 0.14211394300002667,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-2.0-0.8-141-31-123-torch.bfloat16]": 0.48317339399989123,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-2.0-0.8-141-31-123-torch.float32]": 0.14454519100002017,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_with_labels[cutile-16-64-128-torch.float32]": 0.14098488500002304,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_with_labels[cutile-4-64-128-torch.float32]": 0.30629156500003774,
-"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_with_labels[cutile-8-128-256-torch.float32]": 0.34735127400000465,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.bfloat16-3-(1, 1, 16, 32)]": 9.234989677999977,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.bfloat16-3-(1, 2, 64, 32)]": 0.8407793979999951,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.bfloat16-3-(2, 4, 32, 64)]": 10.274809227000048,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.bfloat16-7-(1, 1, 16, 32)]": 0.6495614170000295,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.bfloat16-7-(1, 2, 64, 32)]": 0.5139435100000469,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.bfloat16-7-(2, 4, 32, 64)]": 0.8230611790000353,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float16-3-(1, 1, 16, 32)]": 9.264641014000006,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float16-3-(1, 2, 64, 32)]": 0.7598626760000116,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float16-3-(2, 4, 32, 64)]": 10.182950301999938,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float16-7-(1, 1, 16, 32)]": 0.7277264920000448,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float16-7-(1, 2, 64, 32)]": 0.7214517349999596,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float16-7-(2, 4, 32, 64)]": 0.566320103999999,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float32-3-(1, 1, 16, 32)]": 9.17226262999992,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float32-3-(1, 2, 64, 32)]": 14.821911311000008,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float32-3-(2, 4, 32, 64)]": 18.907695537999928,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float32-7-(1, 1, 16, 32)]": 14.567234485999961,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float32-7-(1, 2, 64, 32)]": 0.19789021700006515,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float32-7-(2, 4, 32, 64)]": 18.984154584000066,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_backward[cutile-(1, 1, 16, 32)]": 0.14162032800010138,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_backward[cutile-(1, 2, 64, 32)]": 0.12063179599999785,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_backward[cutile-(2, 4, 32, 64)]": 0.12426462999997057,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_dilation[cutile-1-(1, 1, 16, 32)]": 0.14030261400006339,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_dilation[cutile-1-(1, 2, 64, 32)]": 0.22174198399994793,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_dilation[cutile-1-(2, 4, 32, 64)]": 0.12861878600006094,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_dilation[cutile-2-(1, 1, 16, 32)]": 0.9729280630000403,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_dilation[cutile-2-(1, 2, 64, 32)]": 0.8851545279999868,
-"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_dilation[cutile-2-(2, 4, 32, 64)]": 1.3465234529999748,
-"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_backward[cutile-(4, 256)-torch.bfloat16]": 0.2815402699999936,
-"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_backward[cutile-(4, 256)-torch.float16]": 0.262497294999946,
-"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_backward[cutile-(4, 256)-torch.float32]": 0.2765871670000024,
-"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_backward[cutile-(8, 512)-torch.float32]": 0.26390888600002427,
-"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_backward[cutile-(9, 41, 341)-torch.float32]": 0.32741563899992343,
-"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_backward[cutile-(9, 7, 41)-torch.float32]": 0.24201767900001414,
-"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(16, 1024)-torch.float32]": 0.2312800560000028,
-"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(2, 2, 8)-torch.float32]": 0.2570046650000677,
-"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(2, 4, 512)-torch.float32]": 0.13619086200003494,
-"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(4, 256)-torch.bfloat16]": 0.22729205299998512,
-"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(4, 256)-torch.float16]": 0.2335968060000937,
-"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(4, 256)-torch.float32]": 0.2111349890000156,
-"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(4, 300)-torch.float32]": 0.26037221700005375,
-"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(8, 512)-torch.float32]": 0.21903659000003017,
-"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(9, 41, 341)-torch.float32]": 0.2978285430000369,
-"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(9, 7, 41)-torch.float32]": 0.23437085400001934,
-"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_backward[cutile-2-4-2-8-torch.bfloat16]": 0.333594249999976,
-"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_backward[cutile-2-4-2-8-torch.float16]": 0.3393604340000138,
-"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_backward[cutile-2-4-2-8-torch.float32]": 0.3151261669999599,
-"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_backward[cutile-2-63-21-2163-torch.float32]": 0.6788409000000115,
-"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_backward[cutile-4-8-4-16-torch.float32]": 0.33366673399996216,
-"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_forward[cutile-16-32-1-4096-torch.float32]": 0.40998079600001347,
-"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_forward[cutile-2-4-2-8-torch.bfloat16]": 0.25382374399998753,
-"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_forward[cutile-2-4-2-8-torch.float16]": 0.276295771999969,
-"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_forward[cutile-2-4-2-8-torch.float32]": 0.29700012899996864,
-"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_forward[cutile-2-6-3-10-torch.float32]": 0.2786880639999936,
-"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_forward[cutile-2-63-21-2163-torch.float32]": 0.5292069809999589,
-"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_forward[cutile-4-8-4-16-torch.float32]": 0.2895203039999501,
-"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op[cutile-cispo-1-4-128]": 0.8623190000000704,
-"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op[cutile-cispo-2-8-256]": 0.8093943019999301,
-"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op[cutile-cispo-4-16-512]": 0.8789487599999575,
-"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op[cutile-grpo-1-4-128]": 0.8796952930000543,
-"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op[cutile-grpo-2-8-256]": 0.950683755,
-"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op[cutile-grpo-4-16-512]": 0.9466196860000196,
-"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op_backward[cutile]": 1.0329972950000297,
-"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op_with_kl[cutile-2-8-256]": 0.3204350159999194,
-"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op_with_mask[cutile-2-8-256]": 0.9781392899999446,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op[cutile-16-1024-0.5-torch.float32]": 0.4577902450000124,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op[cutile-4-256-0.0-torch.float32]": 0.2874218209999526,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op[cutile-4-256-0.5-torch.bfloat16]": 0.34346896099998503,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op[cutile-4-256-0.5-torch.float16]": 0.264479209000001,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op[cutile-4-256-0.5-torch.float32]": 0.16111326700001882,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op[cutile-4-256-1.0-torch.float32]": 0.28913875599999983,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op[cutile-8-512-0.3-torch.float32]": 0.35919587799998,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op[cutile-8-512-0.5-torch.float32]": 0.2919787469999733,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_all_ignored[cutile-20-32-torch.bfloat16]": 0.11588889299997618,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_all_ignored[cutile-20-32-torch.float32]": 0.13874779599996145,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_nondefault_ignore_index[cutile-16-1024-0.3-2-torch.bfloat16]": 0.5623627310000074,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_nondefault_ignore_index[cutile-8-512-0.5-2-torch.float32]": 0.34937917299998844,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_nondefault_ignore_index[cutile-8-512-0.5-42-torch.float32]": 0.3415405800000144,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_not_last_layer[cutile-16441-1271-0.5-torch.float32]": 0.5754381569999509,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_not_last_layer[cutile-2048-3200-0.5-torch.bfloat16]": 0.7639935499999524,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_not_last_layer[cutile-2048-3200-0.5-torch.float32]": 0.6131995320000101,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_not_last_layer[cutile-8-512-0.5-torch.float16]": 0.328431691999981,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_symmetry[cutile-16441-1271-torch.float32]": 0.11948138999997582,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_symmetry[cutile-2048-3200-torch.bfloat16]": 0.12606795399995008,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_symmetry[cutile-2048-3200-torch.float32]": 0.16556394500008764,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_with_labels[cutile-16-1024-0.5-torch.float32]": 0.4864762819999555,
-"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_with_labels[cutile-8-512-0.5-torch.float32]": 0.340069925000023,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-batchmean-(16, 1024)-torch.float32]": 0.09366543000004413,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-batchmean-(4, 256)-torch.bfloat16]": 0.09267055399999435,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-batchmean-(4, 256)-torch.float16]": 0.09522952999998324,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-batchmean-(4, 256)-torch.float32]": 0.13124347999996644,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-batchmean-(4, 300)-torch.float32]": 0.09350290099996528,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-batchmean-(8, 512)-torch.float32]": 0.12788445000001047,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-mean-(16, 1024)-torch.float32]": 0.10568721100003131,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-mean-(4, 256)-torch.bfloat16]": 0.10084045600001446,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-mean-(4, 256)-torch.float16]": 0.09355430899995554,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-mean-(4, 256)-torch.float32]": 0.1469760750000546,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-mean-(4, 300)-torch.float32]": 0.09409216599999581,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-mean-(8, 512)-torch.float32]": 0.11240647199997511,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-none-(16, 1024)-torch.float32]": 0.24148958999995784,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-none-(4, 256)-torch.bfloat16]": 0.22407572200000914,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-none-(4, 256)-torch.float16]": 0.22926582399998097,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-none-(4, 256)-torch.float32]": 0.1986978930000305,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-none-(4, 300)-torch.float32]": 0.28432740699997794,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-none-(8, 512)-torch.float32]": 0.2141598110000018,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-sum-(16, 1024)-torch.float32]": 0.25848541700003125,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-sum-(4, 256)-torch.bfloat16]": 0.24864298600004986,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-sum-(4, 256)-torch.float16]": 0.25829549399998086,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-sum-(4, 256)-torch.float32]": 0.23771883900002422,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-sum-(4, 300)-torch.float32]": 0.30239559200003896,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-sum-(8, 512)-torch.float32]": 0.2451307100000122,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-batchmean-(16, 1024)-torch.float32]": 0.11781124699990642,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-batchmean-(4, 256)-torch.bfloat16]": 0.12373190299996395,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-batchmean-(4, 256)-torch.float16]": 0.11896896000001789,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-batchmean-(4, 256)-torch.float32]": 0.10430994799992277,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-batchmean-(4, 300)-torch.float32]": 0.12418865999995887,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-batchmean-(8, 512)-torch.float32]": 0.11099903599995287,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-mean-(16, 1024)-torch.float32]": 0.11453725699999495,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-mean-(4, 256)-torch.bfloat16]": 0.10107970000001387,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-mean-(4, 256)-torch.float16]": 0.11367509400002973,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-mean-(4, 256)-torch.float32]": 0.1328057319999516,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-mean-(4, 300)-torch.float32]": 0.10942242500004795,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-mean-(8, 512)-torch.float32]": 0.10272073100003354,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-none-(16, 1024)-torch.float32]": 0.23624789800004464,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-none-(4, 256)-torch.bfloat16]": 0.22786811200001011,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-none-(4, 256)-torch.float16]": 0.23235735199995133,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-none-(4, 256)-torch.float32]": 0.1943922030000067,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-none-(4, 300)-torch.float32]": 0.2687828560000298,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-none-(8, 512)-torch.float32]": 0.23797188699995786,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-sum-(16, 1024)-torch.float32]": 0.2406713130000071,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-sum-(4, 256)-torch.bfloat16]": 0.23851392799997484,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-sum-(4, 256)-torch.float16]": 0.23556907300002194,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-sum-(4, 256)-torch.float32]": 0.2462521910001101,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-sum-(4, 300)-torch.float32]": 0.2988604889999351,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-sum-(8, 512)-torch.float32]": 0.21677717500000426,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-batchmean-(16, 1024)-torch.float32]": 0.09900344499993707,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-batchmean-(4, 256)-torch.bfloat16]": 0.10168561000000409,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-batchmean-(4, 256)-torch.float16]": 0.1045315050000113,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-batchmean-(4, 256)-torch.float32]": 0.10024696599992922,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-batchmean-(4, 300)-torch.float32]": 0.09614502399995217,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-batchmean-(8, 512)-torch.float32]": 0.09774484099995107,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-mean-(16, 1024)-torch.float32]": 0.11533936299997549,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-mean-(4, 256)-torch.bfloat16]": 0.10963956800003416,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-mean-(4, 256)-torch.float16]": 0.1135712159999116,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-mean-(4, 256)-torch.float32]": 0.1307901820000552,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-mean-(4, 300)-torch.float32]": 0.09942755700006956,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-mean-(8, 512)-torch.float32]": 0.10863941900004193,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-sum-(16, 1024)-torch.float32]": 0.24483066599998438,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-sum-(4, 256)-torch.bfloat16]": 0.23643329200001517,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-sum-(4, 256)-torch.float16]": 0.23504991599997993,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-sum-(4, 256)-torch.float32]": 0.23081800099998873,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-sum-(4, 300)-torch.float32]": 0.3490863769999919,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-sum-(8, 512)-torch.float32]": 0.2327264829999649,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-batchmean-(16, 1024)-torch.float32]": 0.11076296299995647,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-batchmean-(4, 256)-torch.bfloat16]": 0.10806345500003545,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-batchmean-(4, 256)-torch.float16]": 0.10738246599999002,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-batchmean-(4, 256)-torch.float32]": 0.10495024200002945,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-batchmean-(4, 300)-torch.float32]": 0.10436307300005865,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-batchmean-(8, 512)-torch.float32]": 0.10762750199990023,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-mean-(16, 1024)-torch.float32]": 0.10861098899999888,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-mean-(4, 256)-torch.bfloat16]": 0.10686001599998463,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-mean-(4, 256)-torch.float16]": 0.10783704400000715,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-mean-(4, 256)-torch.float32]": 0.1357639859999722,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-mean-(4, 300)-torch.float32]": 0.11287432199998193,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-mean-(8, 512)-torch.float32]": 0.11523748199999773,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-sum-(16, 1024)-torch.float32]": 0.23557057300001816,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-sum-(4, 256)-torch.bfloat16]": 0.23181284799989044,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-sum-(4, 256)-torch.float16]": 0.23382750700005772,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-sum-(4, 256)-torch.float32]": 0.19270603600006098,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-sum-(4, 300)-torch.float32]": 0.2726386250000701,
-"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-sum-(8, 512)-torch.float32]": 0.23609806399997524,
-"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op[cutile-(16, 1024)-torch.float32]": 0.2856756709999786,
-"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op[cutile-(2, 4, 512)-torch.float32]": 0.1720510899999681,
-"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op[cutile-(4, 256)-torch.bfloat16]": 0.32817168200000424,
-"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op[cutile-(4, 256)-torch.float16]": 0.26339214299997593,
-"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op[cutile-(4, 256)-torch.float32]": 0.2370314209999833,
-"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op[cutile-(4, 300)-torch.float32]": 0.28623422499998696,
-"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op[cutile-(8, 512)-torch.float32]": 0.2576301239999452,
-"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op_backward[cutile-(16, 1024)-torch.float32]": 0.32725877099994705,
-"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op_backward[cutile-(2, 4, 512)-torch.float32]": 0.13838748700004544,
-"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op_backward[cutile-(4, 256)-torch.bfloat16]": 0.29161081399996647,
-"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op_backward[cutile-(4, 256)-torch.float16]": 0.30402560299995685,
-"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op_backward[cutile-(4, 256)-torch.float32]": 0.30969124299997475,
-"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op_backward[cutile-(4, 300)-torch.float32]": 0.308890942000005,
-"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op_backward[cutile-(8, 512)-torch.float32]": 0.32381828800004087,
-"tests/suites/liger/test_llama4_rope.py::Test_Liger_Llama4Rope::test_op[cutile-(1, 16, 8, 128)-torch.float32]": 0.2600568160000307,
-"tests/suites/liger/test_llama4_rope.py::Test_Liger_Llama4Rope::test_op[cutile-(1, 4, 2, 48)-torch.float32]": 0.2577741900000774,
-"tests/suites/liger/test_llama4_rope.py::Test_Liger_Llama4Rope::test_op[cutile-(2, 8, 4, 64)-torch.bfloat16]": 0.2747740789999398,
-"tests/suites/liger/test_llama4_rope.py::Test_Liger_Llama4Rope::test_op[cutile-(2, 8, 4, 64)-torch.float16]": 0.26404408400003376,
-"tests/suites/liger/test_llama4_rope.py::Test_Liger_Llama4Rope::test_op[cutile-(2, 8, 4, 64)-torch.float32]": 0.25635744300001306,
-"tests/suites/liger/test_llama4_rope.py::Test_Liger_Llama4Rope::test_op_backward[cutile-(2, 8, 4, 64)-torch.bfloat16]": 0.2721970910000664,
-"tests/suites/liger/test_llama4_rope.py::Test_Liger_Llama4Rope::test_op_backward[cutile-(2, 8, 4, 64)-torch.float16]": 0.2497211870000342,
-"tests/suites/liger/test_llama4_rope.py::Test_Liger_Llama4Rope::test_op_backward[cutile-(2, 8, 4, 64)-torch.float32]": 0.26610899799993604,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.bfloat16-(64, 512)-per_token]": 0.26926677099999097,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.bfloat16-(64, 512)-shared]": 0.26079465000003665,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.bfloat16-(8, 128)-per_token]": 0.2840044769999963,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.bfloat16-(8, 128)-shared]": 0.2836450189999482,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.float16-(64, 512)-per_token]": 0.2610631729999682,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.float16-(64, 512)-shared]": 0.26109641800002237,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.float16-(8, 128)-per_token]": 0.2433949489999918,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.float16-(8, 128)-shared]": 0.2636863710000057,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.float32-(64, 512)-per_token]": 0.26737971900001867,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.float32-(64, 512)-shared]": 0.25291841199998544,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.float32-(8, 128)-per_token]": 0.24825860900000407,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.float32-(8, 128)-shared]": 0.2699248999999213,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op_backward[cutile-torch.float32-(64, 512)-per_token]": 0.2959113580000121,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op_backward[cutile-torch.float32-(64, 512)-shared]": 0.3234623710000051,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op_backward[cutile-torch.float32-(8, 128)-per_token]": 0.3141607620000286,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op_backward[cutile-torch.float32-(8, 128)-shared]": 0.32091786099994124,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op_casting_modes[cutile-gemma]": 0.23784995599999093,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op_casting_modes[cutile-llama]": 0.1370767439999554,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op_casting_modes[cutile-none]": 0.25702521300007675,
-"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op_no_shift[cutile]": 0.45188440800001217,
-"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op[cutile-(1, 2, 16)-torch.float32]": 0.3811860650000085,
-"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op[cutile-(2, 1, 8)-torch.bfloat16]": 0.4543372869999871,
-"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op[cutile-(2, 1, 8)-torch.float16]": 0.13876650499992138,
-"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op[cutile-(2, 1, 8)-torch.float32]": 0.4059867529999792,
-"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_backward[cutile-(1, 1, 16)-torch.float32]": 0.36533959399997684,
-"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_backward[cutile-(2, 1, 8)-torch.float32]": 0.3528440989999808,
-"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_backward_conv_params[cutile-2-1-8-1-torch.float32]": 0.16146606099999872,
-"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_backward_conv_params[cutile-2-2-8-1-torch.float32]": 0.1343358189999435,
-"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_backward_conv_params[cutile-2-2-8-2-torch.bfloat16]": 0.3299980309999455,
-"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_backward_conv_params[cutile-2-2-8-2-torch.float32]": 0.11154870999996547,
-"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_with_bias[cutile-(2, 1, 8)-torch.bfloat16]": 0.1325867670000207,
-"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_with_bias[cutile-(2, 1, 8)-torch.float16]": 0.13974886299990885,
-"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_with_bias[cutile-(2, 1, 8)-torch.float32]": 0.1480583369999522,
-"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op[cutile-(16, 1024)-torch.float32]": 0.4557513190000009,
-"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op[cutile-(2, 4, 512)-torch.float32]": 0.1390834839999684,
-"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op[cutile-(4, 256)-torch.bfloat16]": 0.47883119499999793,
-"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op[cutile-(4, 256)-torch.float16]": 0.4497351559999174,
-"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op[cutile-(4, 256)-torch.float32]": 0.48260862799997994,
-"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op[cutile-(4, 300)-torch.float32]": 0.4554377479999516,
-"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op[cutile-(8, 512)-torch.float32]": 0.47858826199995974,
-"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op_backward[cutile-(4, 256)-torch.bfloat16]": 0.32603045000001885,
-"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op_backward[cutile-(4, 256)-torch.float16]": 0.33815738399999873,
-"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op_backward[cutile-(4, 256)-torch.float32]": 0.3196604269999739,
-"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op_backward[cutile-(8, 512)-torch.float32]": 0.33115211299997327,
-"tests/suites/liger/test_qwen2vl_mrope.py::Test_Liger_Qwen2VLMRope::test_op[cutile-(1, 2, 4, 48)-torch.float32]": 0.3026307860000088,
-"tests/suites/liger/test_qwen2vl_mrope.py::Test_Liger_Qwen2VLMRope::test_op[cutile-(1, 4, 8, 128)-torch.float32]": 0.3482843400000206,
-"tests/suites/liger/test_qwen2vl_mrope.py::Test_Liger_Qwen2VLMRope::test_op[cutile-(2, 8, 16, 64)-torch.bfloat16]": 0.33647474800000055,
-"tests/suites/liger/test_qwen2vl_mrope.py::Test_Liger_Qwen2VLMRope::test_op[cutile-(2, 8, 16, 64)-torch.float16]": 0.3281254400000648,
-"tests/suites/liger/test_qwen2vl_mrope.py::Test_Liger_Qwen2VLMRope::test_op[cutile-(2, 8, 16, 64)-torch.float32]": 0.3253525910000121,
-"tests/suites/liger/test_qwen2vl_mrope.py::Test_Liger_Qwen2VLMRope::test_op_backward[cutile-(2, 8, 16, 64)-torch.bfloat16]": 0.5042521320000901,
-"tests/suites/liger/test_qwen2vl_mrope.py::Test_Liger_Qwen2VLMRope::test_op_backward[cutile-(2, 8, 16, 64)-torch.float16]": 0.5147260079999683,
-"tests/suites/liger/test_qwen2vl_mrope.py::Test_Liger_Qwen2VLMRope::test_op_backward[cutile-(2, 8, 16, 64)-torch.float32]": 0.49184773299998596,
-"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(1, 1, 1023)-torch.float32]": 0.2480095269999083,
-"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(2, 4, 512)-torch.float32]": 0.21437805000005028,
-"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(2, 8, 4096)-torch.float32]": 0.23900213599995368,
-"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(3, 7, 256)-torch.float32]": 0.131413278000025,
-"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(4, 16, 2048)-torch.float32]": 0.22468961300000956,
-"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(4, 256)-torch.bfloat16]": 0.22017332300004,
-"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(4, 256)-torch.float16]": 0.22120678600003885,
-"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(4, 256)-torch.float32]": 0.2286510840000915,
-"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(4, 300)-torch.float32]": 0.25446997599999577,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op[cutile-(16, 1024)-torch.float32]": 0.24325194499999725,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op[cutile-(2, 4, 512)-torch.float32]": 0.12916967599994678,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op[cutile-(4, 256)-torch.bfloat16]": 0.25121591499998885,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op[cutile-(4, 256)-torch.float16]": 0.2469593600000053,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op[cutile-(4, 256)-torch.float32]": 0.19699383899995837,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op[cutile-(4, 300)-torch.float32]": 0.23816735800005517,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op[cutile-(8, 512)-torch.float32]": 0.23758295900000803,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_backward[cutile-(4, 256)-torch.bfloat16]": 0.28624768500003483,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_backward[cutile-(4, 256)-torch.float16]": 0.2800323059998959,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_backward[cutile-(4, 256)-torch.float32]": 0.22649855200000957,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_backward[cutile-(8, 512)-torch.float32]": 0.2651934550000874,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_casting_modes[cutile-(4, 256)-torch.bfloat16-gemma]": 0.23178914099997883,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_casting_modes[cutile-(4, 256)-torch.bfloat16-llama]": 0.13248091299999487,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_casting_modes[cutile-(4, 256)-torch.bfloat16-none]": 0.2521086699999273,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_casting_modes[cutile-(4, 256)-torch.float32-gemma]": 0.22395213099991906,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_casting_modes[cutile-(4, 256)-torch.float32-llama]": 0.1273590229999968,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_casting_modes[cutile-(4, 256)-torch.float32-none]": 0.2580410170000391,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_no_affine[cutile-(4, 256)-torch.bfloat16]": 0.24185257999999976,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_no_affine[cutile-(4, 256)-torch.float16]": 0.25961181200000283,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_no_affine[cutile-(4, 256)-torch.float32]": 0.24945223600002464,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_no_affine[cutile-(8, 512)-torch.float32]": 0.25092012400000385,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_with_offset[cutile-0.0-(4, 256)-torch.bfloat16]": 0.10760840499995084,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_with_offset[cutile-0.0-(4, 256)-torch.float16]": 0.12107590299996218,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_with_offset[cutile-0.0-(4, 256)-torch.float32]": 0.1366407179999669,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_with_offset[cutile-1.0-(4, 256)-torch.bfloat16]": 0.11495359199994937,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_with_offset[cutile-1.0-(4, 256)-torch.float16]": 0.1112785289999465,
-"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_with_offset[cutile-1.0-(4, 256)-torch.float32]": 0.10952054200004113,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-False-(1, 2, 4, 48)-torch.float32]": 0.13822962099999359,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-False-(1, 4, 8, 128)-torch.float32]": 0.13281797399992,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-False-(2, 8, 16, 64)-torch.bfloat16]": 0.30177599899997176,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-False-(2, 8, 16, 64)-torch.float16]": 0.27296848600002477,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-False-(2, 8, 16, 64)-torch.float32]": 0.29564305500002774,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-False-(3, 4, 7, 64)-torch.float32]": 0.27092564700001276,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-True-(1, 2, 4, 48)-torch.float32]": 0.30190486500004,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-True-(1, 4, 8, 128)-torch.float32]": 0.305673699999943,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-True-(2, 8, 16, 64)-torch.bfloat16]": 0.2883709899999758,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-True-(2, 8, 16, 64)-torch.float16]": 0.31881759699996337,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-True-(2, 8, 16, 64)-torch.float32]": 0.3102699090000556,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-True-(3, 4, 7, 64)-torch.float32]": 0.305253074999996,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op_backward[cutile-(1, 2, 4, 48)-torch.float32]": 0.5322552989999849,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op_backward[cutile-(1, 4, 8, 128)-torch.float32]": 0.49630790000003344,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op_backward[cutile-(2, 8, 16, 64)-torch.bfloat16]": 0.4772267950000355,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op_backward[cutile-(2, 8, 16, 64)-torch.float16]": 0.4826273689999425,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op_backward[cutile-(2, 8, 16, 64)-torch.float32]": 0.48085209400005624,
-"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op_backward[cutile-(3, 4, 7, 64)-torch.float32]": 0.45513294500000256,
-"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op[cutile-(1, 1023)-torch.float32]": 0.5971517039999981,
-"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op[cutile-(1, 2, 4096)-torch.float32]": 0.16104309099995362,
-"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op[cutile-(1, 4096)-torch.float32]": 0.6001372590000642,
-"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op[cutile-(2, 8)-torch.bfloat16]": 0.20739190799997687,
-"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op[cutile-(2, 8)-torch.float16]": 0.43246369399997775,
-"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op[cutile-(2, 8)-torch.float32]": 0.43433102599993845,
-"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op[cutile-(3, 7, 256)-torch.float32]": 0.529327582999997,
-"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op[cutile-(4, 16)-torch.float32]": 0.4329678330000206,
-"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op_backward[cutile-(2, 8)-torch.bfloat16]": 0.27342963499995676,
-"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op_backward[cutile-(2, 8)-torch.float16]": 0.2533724460000144,
-"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op_backward[cutile-(2, 8)-torch.float32]": 1.2574856509999677,
-"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op_backward[cutile-(4, 16)-torch.float32]": 1.348643109999955,
-"tests/suites/liger/test_sparsemax.py::Test_Liger_Sparsemax::test_op[cutile-1-(2, 128, 512)-torch.float32]": 0.9104693610000254,
-"tests/suites/liger/test_sparsemax.py::Test_Liger_Sparsemax::test_op[cutile-1-(4, 256)-torch.float32]": 0.13236234999999397,
-"tests/suites/liger/test_sparsemax.py::Test_Liger_Sparsemax::test_op[cutile-1-(5, 123, 123)-torch.float32]": 0.1460293540000066,
-"tests/suites/liger/test_sparsemax.py::Test_Liger_Sparsemax::test_op[cutile-_1-(2, 128, 512)-torch.float32]": 1.1528961059999574,
-"tests/suites/liger/test_sparsemax.py::Test_Liger_Sparsemax::test_op[cutile-_1-(4, 256)-torch.float32]": 1.2955208969999603,
-"tests/suites/liger/test_sparsemax.py::Test_Liger_Sparsemax::test_op[cutile-_1-(5, 123, 123)-torch.float32]": 1.3510876439999606,
-"tests/suites/liger/test_sparsemax.py::Test_Liger_Sparsemax::test_op_backward[cutile-(2, 128, 512)-torch.float32]": 0.9523264270000027,
-"tests/suites/liger/test_sparsemax.py::Test_Liger_Sparsemax::test_op_backward[cutile-(4, 256)-torch.float32]": 0.9873392350000358,
-"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_backward[cutile-(2, 256, 256)-torch.float32]": 0.24815551700004335,
-"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_backward[cutile-(2, 8, 8)-torch.float32]": 0.3007585649999669,
-"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_backward[cutile-(4, 256)-torch.bfloat16]": 0.2602785600000743,
-"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_backward[cutile-(4, 256)-torch.float16]": 0.27952288400001635,
-"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_backward[cutile-(4, 256)-torch.float32]": 0.25903023200004327,
-"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_backward[cutile-(6, 42, 123)-torch.float32]": 0.25737957999996297,
-"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_backward[cutile-(9, 7, 41)-torch.float32]": 0.32049090000003844,
-"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_forward[cutile-(2, 256, 256)-torch.float32]": 0.2903391980000265,
-"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_forward[cutile-(2, 8, 8)-torch.float32]": 0.2653622149999819,
-"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_forward[cutile-(4, 256)-torch.bfloat16]": 0.24032394099998555,
-"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_forward[cutile-(4, 256)-torch.float16]": 0.24565399799996612,
-"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_forward[cutile-(4, 256)-torch.float32]": 0.25164382999992085,
-"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_forward[cutile-(4, 300)-torch.float32]": 0.2578569479999828,
-"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_forward[cutile-(6, 42, 123)-torch.float32]": 0.25793733299997257,
-"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_forward[cutile-(9, 7, 41)-torch.float32]": 0.2609263740000074,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-2-1-1024-128]": 0.12153015400002687,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-2-2-512-512]": 0.12952137700006006,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-2-4-127-128]": 0.11663116400001172,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-4-1-1024-128]": 0.12991460800003551,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-4-2-512-512]": 0.11975346399998443,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-4-4-127-128]": 0.13274288600001682,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-None-1-1024-128]": 0.13293671299999232,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-None-2-512-512]": 0.11982793499998934,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-None-4-127-128]": 0.11945723700000599,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-2-1-1024-128]": 0.11220423299994309,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-2-1-1024-256]": 0.10262780600004362,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-2-2-1024-64]": 0.1101494700000103,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-2-2-512-512]": 0.104542145000039,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-2-4-127-128]": 0.10877230500000223,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-4-1-1024-128]": 0.10614156999997704,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-4-1-1024-256]": 0.11627781599997888,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-4-2-1024-64]": 0.1123424049999926,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-4-2-512-512]": 0.11945644900009711,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-4-4-127-128]": 0.11173109699996076,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-None-1-1024-128]": 0.14291954200001555,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-None-1-1024-256]": 0.12607550199999196,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-None-2-1024-64]": 0.13145273600002838,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-None-2-512-512]": 0.12394130900008804,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-None-4-127-128]": 0.14084131800001387,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-2-1-1024-128]": 0.11191818999998304,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-2-1-1024-256]": 0.17429282900008047,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-2-2-1024-64]": 0.10768001799999638,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-2-2-512-512]": 0.12009257199997592,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-2-4-127-128]": 0.12197674600002983,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-4-1-1024-128]": 0.14254702800002406,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-4-1-1024-256]": 0.1591937899999607,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-4-2-1024-64]": 0.13528095200001644,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-4-2-512-512]": 0.14746473600001764,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-4-4-127-128]": 0.1419192319999638,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-None-1-1024-128]": 0.14066432999993594,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-None-1-1024-256]": 0.11461302099996828,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-None-2-1024-64]": 0.12596390899994958,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-None-2-512-512]": 0.10648840199996812,
-"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-None-4-127-128]": 0.11800934299998289,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-batchmean-1-4096-32000]": 2.402667156000007,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-batchmean-3-423-32000]": 2.400004329000069,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-batchmean-41-401-1271]": 0.45598296999997956,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-mean-1-4096-32000]": 2.3978507800000557,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-mean-3-423-32000]": 2.437172847999989,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-mean-41-401-1271]": 0.42728186700003334,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-none-1-4096-32000]": 0.7751403330000244,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-none-3-423-32000]": 0.7968421780000199,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-none-41-401-1271]": 0.3543314849999888,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-sum-1-4096-32000]": 2.374297253000009,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-sum-3-423-32000]": 2.4322816179999336,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-sum-41-401-1271]": 0.43240130399999543,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-batchmean-1-4096-32000]": 3.1641787000000363,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-batchmean-3-423-32000]": 3.0919004289999634,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-batchmean-41-401-1271]": 0.3484173610000312,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-mean-1-4096-32000]": 3.191643508000027,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-mean-3-423-32000]": 3.228168586000038,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-mean-41-401-1271]": 0.3761592679999808,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-none-1-4096-32000]": 0.5098973579999324,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-none-3-423-32000]": 0.48869310400004906,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-none-41-401-1271]": 0.3703627309999433,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-sum-1-4096-32000]": 3.1247677499999895,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-sum-3-423-32000]": 3.16601150799994,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-sum-41-401-1271]": 0.36838142500005233,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-batchmean-1-4096-32000]": 0.17253410600000052,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-batchmean-3-423-32000]": 0.15101328899993405,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-batchmean-41-401-1271]": 0.17253634400003648,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-mean-1-4096-32000]": 0.1396459769999865,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-mean-3-423-32000]": 0.14453663500000857,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-mean-41-401-1271]": 0.1665612890000716,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-sum-1-4096-32000]": 0.16502732499998274,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-sum-3-423-32000]": 0.13408178800000314,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-sum-41-401-1271]": 0.16813406499994699,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-batchmean-1-4096-32000]": 1.8032774609999933,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-batchmean-3-423-32000]": 1.7746416079999676,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-batchmean-41-401-1271]": 0.365040449999924,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-mean-1-4096-32000]": 1.7550247890000037,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-mean-3-423-32000]": 1.7552137170000606,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-mean-41-401-1271]": 0.38454197899994824,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-none-1-4096-32000]": 0.6096591399999625,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-none-3-423-32000]": 0.6025380119999681,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-none-41-401-1271]": 0.38912262899998495,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-sum-1-4096-32000]": 1.7479629649999993,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-sum-3-423-32000]": 1.80043728499993,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-sum-41-401-1271]": 0.3795242430000485,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-batchmean-1-4096-32000]": 1.7642669580000074,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-batchmean-3-423-32000]": 1.7289347730000486,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-batchmean-41-401-1271]": 0.3834586249999461,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-mean-1-4096-32000]": 1.7751263670000412,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-mean-3-423-32000]": 1.8184812329999431,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-mean-41-401-1271]": 0.35036454200002254,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-none-1-4096-32000]": 0.6120711159999814,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-none-3-423-32000]": 0.5884223629999497,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-none-41-401-1271]": 0.378681805000042,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-sum-1-4096-32000]": 1.7553342290000273,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-sum-3-423-32000]": 1.759391277000077,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-sum-41-401-1271]": 0.37069877999994105,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-batchmean-1-4096-32000]": 1.758624382999983,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-batchmean-3-423-32000]": 1.7194179869999857,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-batchmean-41-401-1271]": 0.38763051600000153,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-mean-1-4096-32000]": 1.728356437000059,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-mean-3-423-32000]": 1.7344754900000225,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-mean-41-401-1271]": 0.3686805590000404,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-none-1-4096-32000]": 0.6786840590000338,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-none-3-423-32000]": 0.5945790430000102,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-none-41-401-1271]": 0.3937336240000491,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-sum-1-4096-32000]": 1.7359105830000203,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-sum-3-423-32000]": 1.8079542539999807,
-"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-sum-41-401-1271]": 0.36859491100005926,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_all_ignored[cutile]": 0.13273032599994394,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_backward[cutile-torch.bfloat16-1-64-32000]": 0.14227341500003376,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_backward[cutile-torch.bfloat16-2-32-1024]": 0.2526748709999538,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_backward[cutile-torch.float16-1-64-32000]": 0.45767964200001643,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_backward[cutile-torch.float16-2-32-1024]": 0.2367091030000097,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_backward_large_vocab[cutile-torch.bfloat16-1-16-100000]": 0.14077825099991514,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_backward_large_vocab[cutile-torch.bfloat16-2-8-128256]": 0.1312319000000457,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_backward_logit_scaling[cutile-torch.bfloat16-2-16-1024]": 0.12552602200003093,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_backward_softcapping[cutile-torch.bfloat16-2-16-1024]": 0.15496368700001995,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_basic[cutile-torch.bfloat16-1-128-50257]": 1.4744824429999994,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_basic[cutile-torch.bfloat16-2-64-1024]": 0.021098639999991065,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_basic[cutile-torch.bfloat16-4-32-32000]": 0.29374728599998434,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_basic[cutile-torch.float16-1-128-50257]": 1.3021989919999442,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_basic[cutile-torch.float16-2-64-1024]": 0.03567045000005464,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_basic[cutile-torch.float16-4-32-32000]": 0.021631551000041327,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_large_vocab[cutile-torch.bfloat16-1-16-100000]": 0.17227404799996293,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_large_vocab[cutile-torch.bfloat16-2-8-128256]": 0.19675905299999386,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_logit_scaling[cutile-torch.bfloat16-2-16-1024]": 0.11437288999997008,
-"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_softcapping[cutile-torch.bfloat16-2-16-1024]": 0.14249756100002742,
-"tests/suites/unsloth/test_fp8.py::Test_Unsloth_ActQuant::test_op[cutile-(1024, 4096)-128]": 0.002401671000029637,
-"tests/suites/unsloth/test_fp8.py::Test_Unsloth_ActQuant::test_op[cutile-(128, 512)-128]": 0.0031153049999943505,
-"tests/suites/unsloth/test_fp8.py::Test_Unsloth_ActQuant::test_op[cutile-(2048, 8192)-128]": 0.0028097349999711696,
-"tests/suites/unsloth/test_fp8.py::Test_Unsloth_ActQuant::test_op[cutile-(256, 256)-128]": 0.10317757700005359,
-"tests/suites/unsloth/test_fp8.py::Test_Unsloth_ActQuant::test_op[cutile-(64, 1024)-128]": 0.0031037869999295253,
-"tests/suites/unsloth/test_fp8.py::Test_Unsloth_ActQuant::test_op_non_square[cutile-(128, 384)-128]": 0.0018842910000103075,
-"tests/suites/unsloth/test_fp8.py::Test_Unsloth_ActQuant::test_op_non_square[cutile-(256, 128)-128]": 0.0018005739999580328,
-"tests/suites/unsloth/test_fp8.py::Test_Unsloth_W8A8_FP8_Matmul::test_op[cutile-1024-2048-1024-[128, 128]]": 0.7399935870000149,
-"tests/suites/unsloth/test_fp8.py::Test_Unsloth_W8A8_FP8_Matmul::test_op[cutile-128-128-128-[128, 128]]": 0.3964506749999259,
-"tests/suites/unsloth/test_fp8.py::Test_Unsloth_W8A8_FP8_Matmul::test_op[cutile-256-256-256-[128, 128]]": 0.569933366999976,
-"tests/suites/unsloth/test_fp8.py::Test_Unsloth_W8A8_FP8_Matmul::test_op[cutile-512-1024-512-[128, 128]]": 0.6617622630000142,
-"tests/suites/unsloth/test_fp8.py::Test_Unsloth_WeightDequant::test_op[cutile-2048-4096-128]": 0.30518428599998515,
-"tests/suites/unsloth/test_fp8.py::Test_Unsloth_WeightDequant::test_op[cutile-256-256-128]": 0.308578553000018,
-"tests/suites/unsloth/test_fp8.py::Test_Unsloth_WeightDequant::test_op[cutile-384-768-128]": 0.3001257370000303,
-"tests/suites/unsloth/test_fp8.py::Test_Unsloth_WeightDequant::test_op[cutile-4096-8192-128]": 0.35215963299992836,
-"tests/suites/unsloth/test_fp8.py::Test_Unsloth_WeightDequant::test_op[cutile-512-1024-128]": 0.311540736999973,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Backward::test_op[cutile-torch.bfloat16-1024-1024]": 0.12356254799999533,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Backward::test_op[cutile-torch.bfloat16-2048-5120]": 0.13649315100002468,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Backward::test_op[cutile-torch.bfloat16-256-512]": 0.11626140200002055,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Backward::test_op[cutile-torch.bfloat16-4096-8192]": 0.1414377140000056,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Backward::test_op[cutile-torch.float16-1024-1024]": 0.1038506329999791,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Backward::test_op[cutile-torch.float16-2048-5120]": 0.10944024099995886,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Backward::test_op[cutile-torch.float16-256-512]": 0.12044250499997133,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Backward::test_op[cutile-torch.float16-4096-8192]": 0.1496392629999832,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.bfloat16-(1, 1, 1024)]": 0.4204019070000413,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.bfloat16-(2, 1024, 8192)]": 0.4771459959999902,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.bfloat16-(2, 128, 256)]": 0.4655557380000346,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.bfloat16-(2, 256, 4096)]": 0.4033601510000153,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.bfloat16-(4, 512, 5120)]": 0.4297753989999933,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.bfloat16-(4, 64, 512)]": 0.44709165500000836,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.float16-(1, 1, 1024)]": 0.4471374179999543,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.float16-(2, 1024, 8192)]": 0.37741454600001134,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.float16-(2, 128, 256)]": 0.45697944500000176,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.float16-(2, 256, 4096)]": 0.4952614599999947,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.float16-(4, 512, 5120)]": 0.4105787860000305,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.float16-(4, 64, 512)]": 0.4454500300000177,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Backward::test_op[cutile-torch.bfloat16-1024-1024]": 0.11853776899999957,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Backward::test_op[cutile-torch.bfloat16-2048-5120]": 0.14103745700003856,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Backward::test_op[cutile-torch.bfloat16-256-512]": 0.13318017500000678,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Backward::test_op[cutile-torch.bfloat16-4096-8192]": 0.16117435199998909,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Backward::test_op[cutile-torch.float16-1024-1024]": 0.13652236699999776,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Backward::test_op[cutile-torch.float16-2048-5120]": 0.12238820300001407,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Backward::test_op[cutile-torch.float16-256-512]": 0.15773447899999837,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Backward::test_op[cutile-torch.float16-4096-8192]": 0.16042635599995947,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.bfloat16-(1, 1, 1024)]": 0.4833043049999901,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.bfloat16-(2, 1024, 8192)]": 0.5244869339999809,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.bfloat16-(2, 128, 256)]": 0.4885312659999954,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.bfloat16-(2, 256, 4096)]": 0.48563852600000246,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.bfloat16-(4, 512, 5120)]": 0.49121207700000014,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.bfloat16-(4, 64, 512)]": 0.48250601400002324,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.float16-(1, 1, 1024)]": 0.5068433389999996,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.float16-(2, 1024, 8192)]": 0.502084352000054,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.float16-(2, 128, 256)]": 0.4924958539999693,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.float16-(2, 256, 4096)]": 0.49456470500001615,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.float16-(4, 512, 5120)]": 0.4972137429999748,
-"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.float16-(4, 64, 512)]": 0.4734334210000952,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_W_grad[cutile-torch.bfloat16-4-16-64-64-1]": 0.0028234840000322947,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_W_grad[cutile-torch.bfloat16-4-32-128-128-1]": 0.002611618999935672,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_W_grad[cutile-torch.bfloat16-8-64-512-256-1]": 9.149405005000062,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_W_grad[cutile-torch.float16-4-16-64-64-1]": 0.009352100999876711,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_W_grad[cutile-torch.float16-4-32-128-128-1]": 0.00400365400003011,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_W_grad[cutile-torch.float16-8-64-512-256-1]": 0.0037439489999542275,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_W_grad_topk4[cutile-torch.bfloat16-4-16-128-128-4]": 0.0038545919999251055,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_W_grad_topk4[cutile-torch.bfloat16-8-8-256-128-4]": 0.0037878589999991163,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_X_grad[cutile-torch.bfloat16-4-16-64-64-1]": 7.166191269000024,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_X_grad[cutile-torch.bfloat16-4-32-128-128-1]": 6.082815801000038,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_X_grad[cutile-torch.bfloat16-8-64-512-256-1]": 0.3101235900001029,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_X_grad[cutile-torch.float16-4-16-64-64-1]": 7.5960813259999895,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_X_grad[cutile-torch.float16-4-32-128-128-1]": 6.396920233999992,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_X_grad[cutile-torch.float16-8-64-512-256-1]": 9.347685664999858,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_X_grad_topk4[cutile-torch.bfloat16-4-16-128-128-4]": 6.179076599000268,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_X_grad_topk4[cutile-torch.bfloat16-8-8-256-128-4]": 7.563866540000049,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_deterministic[cutile-8-32-256-128]": 1.7479724060000308,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_edge_cases[cutile-torch.bfloat16-[0, 0, 0, 64]-64-64]": 0.002172737000023517,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_edge_cases[cutile-torch.bfloat16-[128]-256-128]": 2.279568230000109,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_edge_cases[cutile-torch.bfloat16-[32, 16, 8, 8]-64-64]": 0.3689976939999724,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_edge_cases[cutile-torch.bfloat16-[64, 0, 32, 0]-128-128]": 1.6462430789999871,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.bfloat16-16-16-768-512-1]": 5.103825131000008,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.bfloat16-16-32-2048-1024-1]": 9.536944302000052,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.bfloat16-4-16-64-64-1]": 1.3144567200000665,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.bfloat16-4-32-128-128-1]": 1.6714290840000103,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.bfloat16-8-16-128-128-2]": 1.8089903390000472,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.bfloat16-8-64-512-256-1]": 3.265927354000098,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.bfloat16-8-8-64-128-1]": 1.578274126999986,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.float16-16-16-768-512-1]": 4.817008771000019,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.float16-16-32-2048-1024-1]": 9.776041471000042,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.float16-4-16-64-64-1]": 1.1316823419999196,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.float16-4-32-128-128-1]": 1.705107916999964,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.float16-8-16-128-128-2]": 1.852530809999962,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.float16-8-64-512-256-1]": 2.8786963709999327,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.float16-8-8-64-128-1]": 1.4088828019999369,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward_large[cutile-torch.bfloat16-128-8-768-512-1]": 6.506502144999956,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward_topk4[cutile-torch.bfloat16-4-16-128-128-4]": 1.6790749449999112,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward_topk4[cutile-torch.bfloat16-8-8-256-128-4]": 1.779323575000035,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward_upstream_model[cutile-torch.bfloat16-128-8-768-2048-8]": 4.471815114999913,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward_upstream_model[cutile-torch.bfloat16-16-64-8192-5120-1]": 3.0509513709999965,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_non_power_of_2[cutile-torch.bfloat16-4-20-300-192]": 2.731657468999913,
-"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_non_power_of_2[cutile-torch.bfloat16-8-13-100-96]": 2.1090372630000616,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.bfloat16-(16, 1024)]": 0.47331382800007304,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.bfloat16-(2, 4, 512)]": 0.002503036999996766,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.bfloat16-(21, 349, 1024)]": 0.5058167060000187,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.bfloat16-(32, 4096)]": 0.6127465890001531,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.bfloat16-(4, 256)]": 0.4309699970000338,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.bfloat16-(4, 300)]": 0.782303762999959,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.bfloat16-(8, 512)]": 0.4889395019999938,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.float16-(16, 1024)]": 0.498614110999938,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.float16-(2, 4, 512)]": 0.003152456000066195,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.float16-(21, 349, 1024)]": 0.5171284170000945,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.float16-(32, 4096)]": 0.6725543740000148,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.float16-(4, 256)]": 0.13955435200011834,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.float16-(4, 300)]": 0.7243690560000005,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.float16-(8, 512)]": 0.5126720020000448,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op_backward[cutile-(2, 4, 512)-torch.bfloat16]": 0.004608588999872154,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op_backward[cutile-(4, 2048)-torch.bfloat16]": 1.1440183979999574,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op_backward[cutile-(4, 256)-torch.float16]": 1.0254637639999373,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op_backward[cutile-(4, 256)-torch.float32]": 0.955539905999899,
-"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op_backward[cutile-(8, 512)-torch.bfloat16]": 0.5367806830000745,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward[cutile-(2, 4, 512)-torch.bfloat16]": 0.6038072480000665,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward[cutile-(4, 2048)-torch.bfloat16]": 1.1505372660000148,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward[cutile-(4, 256)-torch.float16]": 0.5025023550000469,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward[cutile-(4, 256)-torch.float32]": 1.0599668529999917,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward[cutile-(8, 512)-torch.bfloat16]": 0.10780232199999773,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward_gemma[cutile-(16, 1024)-torch.bfloat16]": 0.47811547200001314,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward_gemma[cutile-(2, 4, 512)-torch.bfloat16]": 0.4884928679998666,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward_gemma[cutile-(4, 256)-torch.float16]": 0.6205145599998332,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward_gemma[cutile-(8, 512)-torch.bfloat16]": 0.3423722170000474,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_gemma[cutile-torch.bfloat16-(16, 1024)]": 0.453243567999948,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_gemma[cutile-torch.bfloat16-(4, 256)]": 0.45447026800002277,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_gemma[cutile-torch.bfloat16-(8, 512)]": 0.5059729890000426,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_gemma[cutile-torch.float16-(16, 1024)]": 0.45336702700001297,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_gemma[cutile-torch.float16-(4, 256)]": 0.4871675019999202,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_gemma[cutile-torch.float16-(8, 512)]": 0.45881626100003814,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.bfloat16-(16, 1024)]": 0.439909784000065,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.bfloat16-(2, 4, 512)]": 0.0024520839999695454,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.bfloat16-(21, 349, 1024)]": 0.48320130599995537,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.bfloat16-(32, 4096)]": 0.5556934789999559,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.bfloat16-(4, 256)]": 0.45846769600007065,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.bfloat16-(4, 300)]": 0.6884496300000364,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.bfloat16-(8, 512)]": 0.42911033699999734,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.float16-(16, 1024)]": 0.46960611100007554,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.float16-(2, 4, 512)]": 0.0032867369999394214,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.float16-(21, 349, 1024)]": 0.4435703779998903,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.float16-(32, 4096)]": 0.7202101219999122,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.float16-(4, 256)]": 0.4466256279999925,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.float16-(4, 300)]": 0.6313630800000283,
-"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.float16-(8, 512)]": 0.41283722600007877,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op[cutile-torch.bfloat16-1-128-4-128]": 0.398573474999921,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op[cutile-torch.bfloat16-2-64-8-64]": 0.387200779000068,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op[cutile-torch.bfloat16-4-32-16-64]": 0.3937358389999872,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op[cutile-torch.float16-1-128-4-128]": 0.421525416999998,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op[cutile-torch.float16-2-64-8-64]": 0.41535505299998476,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op[cutile-torch.float16-4-32-16-64]": 0.4223318509999672,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op_backward[cutile-torch.bfloat16-1-128-4-128]": 0.22691756100016391,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op_backward[cutile-torch.bfloat16-2-64-8-64]": 0.24482456500015815,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op_backward[cutile-torch.float16-1-128-4-128]": 0.10363365399985014,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op_backward[cutile-torch.float16-2-64-8-64]": 0.0982356179999897,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op[cutile-torch.bfloat16-1-32-8-128-128]": 0.5303091130000439,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op[cutile-torch.bfloat16-2-8-8-64-64]": 0.4726201849998688,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op[cutile-torch.bfloat16-4-16-4-32-64]": 0.551470256000016,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op[cutile-torch.float16-1-32-8-128-128]": 0.459458126000186,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op[cutile-torch.float16-2-8-8-64-64]": 0.4842655579999473,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op[cutile-torch.float16-4-16-4-32-64]": 0.5686654249999492,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op_backward[cutile-torch.bfloat16-1-32-8-128-128]": 0.29786790200000723,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op_backward[cutile-torch.bfloat16-2-8-8-64-64]": 0.1251643809999905,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op_backward[cutile-torch.float16-1-32-8-128-128]": 0.27610774599997967,
-"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op_backward[cutile-torch.float16-2-8-8-64-64]": 0.28433568200011905,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.bfloat16-1024-1024]": 0.1095306109999683,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.bfloat16-2048-5120]": 0.10229543400009788,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.bfloat16-256-512]": 0.127607385000033,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.bfloat16-4096-8192]": 0.1593039609999778,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.bfloat16-512-4096]": 0.12156767900012255,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.float16-1024-1024]": 0.11182399100016482,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.float16-2048-5120]": 0.10933062899982815,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.float16-256-512]": 0.09966446999987966,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.float16-4096-8192]": 0.14333101899990197,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.float16-512-4096]": 0.10145540900009564,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.bfloat16-(1, 1, 1024)]": 0.45417870199992194,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.bfloat16-(2, 1024, 8192)]": 0.3948951569998371,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.bfloat16-(2, 128, 256)]": 0.4381393590000471,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.bfloat16-(2, 256, 4096)]": 0.46734088800008067,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.bfloat16-(4, 512, 5120)]": 0.43648219399995014,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.bfloat16-(4, 64, 512)]": 0.459770540999898,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.float16-(1, 1, 1024)]": 0.42055715499998314,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.float16-(2, 1024, 8192)]": 0.43389688800004933,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.float16-(2, 128, 256)]": 0.47711276300003647,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.float16-(2, 256, 4096)]": 0.4221766739999566,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.float16-(4, 512, 5120)]": 0.3791067550000662,
-"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.float16-(4, 64, 512)]": 0.4389662319999843
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-0-torch.bfloat16-none]": 0.6252736070000253,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-0-torch.bfloat16-tanh]": 0.5366465959999971,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-0-torch.float16-none]": 0.6586260119999565,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-0-torch.float16-tanh]": 0.5260472429999936,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-0-torch.float32-none]": 0.6681368269999552,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-0-torch.float32-tanh]": 0.5696530079999889,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-1-torch.bfloat16-none]": 0.27626397300002736,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-1-torch.bfloat16-tanh]": 0.24053664199999503,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-1-torch.float16-none]": 0.27805149299999243,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-1-torch.float16-tanh]": 0.2552596440000343,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-1-torch.float32-none]": 0.31431461699997953,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(16, 4)-1-torch.float32-tanh]": 0.2444632659999968,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-0-torch.bfloat16-none]": 0.6187644629999909,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-0-torch.bfloat16-tanh]": 0.5320870319999926,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-0-torch.float16-none]": 0.694602605999961,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-0-torch.float16-tanh]": 0.5813207159999649,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-0-torch.float32-none]": 0.6647503670000106,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-0-torch.float32-tanh]": 0.48146434899999235,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-1-torch.bfloat16-none]": 0.7012354450000089,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-1-torch.bfloat16-tanh]": 0.555743528999983,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-1-torch.float16-none]": 0.7668686639999862,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-1-torch.float16-tanh]": 0.6189072390000092,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-1-torch.float32-none]": 0.7683462239999983,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-1-torch.float32-tanh]": 0.6170357620000004,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-2-torch.bfloat16-none]": 0.671095825000009,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-2-torch.bfloat16-tanh]": 0.548967872999981,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-2-torch.float16-none]": 0.6662282890000313,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-2-torch.float16-tanh]": 0.5688067070000216,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-2-torch.float32-none]": 1.5601185940000164,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-2-torch.float32-tanh]": 0.599976737999981,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-3-torch.bfloat16-none]": 0.6771518389999756,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-3-torch.bfloat16-tanh]": 0.5595318930000133,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-3-torch.float16-none]": 0.7257824389999428,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-3-torch.float16-tanh]": 0.6013019189999511,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-3-torch.float32-none]": 0.7893354439999882,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(256, 64, 16, 4)-3-torch.float32-tanh]": 0.5864020950000111,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(4,)-0-torch.bfloat16-none]": 0.6475792809999916,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(4,)-0-torch.bfloat16-tanh]": 0.5015541029999611,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(4,)-0-torch.float16-none]": 0.7146322100000191,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(4,)-0-torch.float16-tanh]": 0.5369434959999637,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(4,)-0-torch.float32-none]": 1.3722410030000276,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(4,)-0-torch.float32-tanh]": 0.5981410819999553,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-0-torch.bfloat16-none]": 0.6272955609999542,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-0-torch.bfloat16-tanh]": 0.5612066080000204,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-0-torch.float16-none]": 0.7448494790000098,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-0-torch.float16-tanh]": 0.5839434609999614,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-0-torch.float32-none]": 0.6962103850000005,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-0-torch.float32-tanh]": 0.5268335090000278,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-1-torch.bfloat16-none]": 0.2807838649999894,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-1-torch.bfloat16-tanh]": 0.27180579699998475,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-1-torch.float16-none]": 0.2922866000000113,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-1-torch.float16-tanh]": 0.28802287999999976,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-1-torch.float32-none]": 0.28392996399998083,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-1-torch.float32-tanh]": 0.24618515899999238,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-2-torch.bfloat16-none]": 0.18619695700002126,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-2-torch.bfloat16-tanh]": 0.18979039200002035,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-2-torch.float16-none]": 0.18812568600000645,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-2-torch.float16-tanh]": 0.19171628799998075,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-2-torch.float32-none]": 0.1836512950000042,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[cutile-(64, 16, 4)-2-torch.float32-tanh]": 0.18813155899999856,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-0-torch.bfloat16-none]": 0.17089947099998426,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-0-torch.bfloat16-tanh]": 0.17269759799998496,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-0-torch.float16-none]": 0.17044648300003473,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-0-torch.float16-tanh]": 0.17488221399997883,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-0-torch.float32-none]": 0.17318249500002025,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-0-torch.float32-tanh]": 0.1723762059999956,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-1-torch.bfloat16-none]": 0.17448207799998272,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-1-torch.bfloat16-tanh]": 0.16823858100002553,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-1-torch.float16-none]": 0.17670486900001947,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-1-torch.float16-tanh]": 0.18313881299997092,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-1-torch.float32-none]": 0.17292352100002972,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(16, 4)-1-torch.float32-tanh]": 0.17658024199997158,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-0-torch.bfloat16-none]": 1.7244771050000054,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-0-torch.bfloat16-tanh]": 1.6796671830000207,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-0-torch.float16-none]": 1.7424329170000306,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-0-torch.float16-tanh]": 1.9586351280000258,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-0-torch.float32-none]": 2.3361259780000125,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-0-torch.float32-tanh]": 1.8608870400000228,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-1-torch.bfloat16-none]": 0.14342149899997025,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-1-torch.bfloat16-tanh]": 0.1431789540000068,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-1-torch.float16-none]": 0.1426891410000053,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-1-torch.float16-tanh]": 0.14196992700001942,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-1-torch.float32-none]": 0.1485111659999916,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-1-torch.float32-tanh]": 0.14271204000002058,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-2-torch.bfloat16-none]": 0.1460144840000055,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-2-torch.bfloat16-tanh]": 0.15082480799995324,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-2-torch.float16-none]": 0.14646101199997474,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-2-torch.float16-tanh]": 0.14655821600001673,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-2-torch.float32-none]": 0.14825087399998438,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-2-torch.float32-tanh]": 0.14653732899998317,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-3-torch.bfloat16-none]": 0.17275282299999617,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-3-torch.bfloat16-tanh]": 0.17117529000003628,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-3-torch.float16-none]": 0.17120266399999196,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-3-torch.float16-tanh]": 0.16951792500000806,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-3-torch.float32-none]": 0.1679981039999916,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(256, 64, 16, 4)-3-torch.float32-tanh]": 0.17079913399999214,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(4,)-0-torch.bfloat16-none]": 0.36250598800000944,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(4,)-0-torch.bfloat16-tanh]": 1.6799269350000259,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(4,)-0-torch.float16-none]": 0.17130128900001296,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(4,)-0-torch.float16-tanh]": 0.16792116300004523,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(4,)-0-torch.float32-none]": 0.171889186000044,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(4,)-0-torch.float32-tanh]": 0.17310887500002536,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-0-torch.bfloat16-none]": 0.16941371099997582,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-0-torch.bfloat16-tanh]": 0.16707567399996037,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-0-torch.float16-none]": 0.17281126199998198,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-0-torch.float16-tanh]": 0.17114473500001282,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-0-torch.float32-none]": 0.17147508899998343,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-0-torch.float32-tanh]": 0.17021270600000094,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-1-torch.bfloat16-none]": 0.14327915700002336,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-1-torch.bfloat16-tanh]": 0.14754422100000397,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-1-torch.float16-none]": 0.18984858900000745,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-1-torch.float16-tanh]": 0.1505900769999755,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-1-torch.float32-none]": 0.1701512439999533,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-1-torch.float32-tanh]": 0.16566800100000023,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-2-torch.bfloat16-none]": 0.15215651499997307,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-2-torch.bfloat16-tanh]": 0.1478557200000239,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-2-torch.float16-none]": 0.15232452699999044,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-2-torch.float16-tanh]": 0.15148400999999012,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-2-torch.float32-none]": 0.15614014499996642,
+"tests/ops/activation/test_geglu.py::Test_GEGLU::test_op[tilecpp-(64, 16, 4)-2-torch.float32-tanh]": 0.15815364300001988,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[cutile-256-2048-none-torch.bfloat16]": 0.5156313530000318,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[cutile-256-2048-none-torch.float16]": 0.5581859729999792,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[cutile-256-2048-none-torch.float32]": 0.6838103039999908,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[cutile-256-2048-tanh-torch.bfloat16]": 0.44549970800002825,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[cutile-256-2048-tanh-torch.float16]": 0.4320038360000069,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[cutile-256-2048-tanh-torch.float32]": 0.5101320480000027,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[tilecpp-256-2048-none-torch.bfloat16]": 2.0005139499999984,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[tilecpp-256-2048-none-torch.float16]": 2.1448217740000075,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[tilecpp-256-2048-none-torch.float32]": 1.940828191000037,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[tilecpp-256-2048-tanh-torch.bfloat16]": 1.7858262229999866,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[tilecpp-256-2048-tanh-torch.float16]": 1.8969022770000095,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op[tilecpp-256-2048-tanh-torch.float32]": 1.805668359000009,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op_unaligned[cutile-none-torch.float16]": 0.6977925399999947,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op_unaligned[cutile-none-torch.float32]": 0.6430743770000049,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op_unaligned[cutile-tanh-torch.float16]": 0.6807180039999992,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op_unaligned[cutile-tanh-torch.float32]": 0.5633762679999847,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op_unaligned[tilecpp-none-torch.float16]": 0.15145184399997902,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op_unaligned[tilecpp-none-torch.float32]": 0.16306560000001014,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op_unaligned[tilecpp-tanh-torch.float16]": 0.16577545299995222,
+"tests/ops/activation/test_gelu.py::Test_GeLU::test_op_unaligned[tilecpp-tanh-torch.float32]": 0.15825863300000265,
+"tests/ops/activation/test_relu.py::Test_ReLU::test_op[cutile-256-2048-torch.float32]": 0.4512697929999945,
+"tests/ops/activation/test_relu.py::Test_ReLU::test_op[cutile-256-256-torch.float16]": 0.43123341200001164,
+"tests/ops/activation/test_relu.py::Test_ReLU::test_op[cutile-256-32768-torch.float16]": 0.4408274639999945,
+"tests/ops/activation/test_relu.py::Test_ReLU::test_op[cutile-256-64-torch.float16]": 0.4152050220000092,
+"tests/ops/activation/test_relu.py::Test_ReLU::test_op[tilecpp-256-2048-torch.float32]": 1.0660012399999914,
+"tests/ops/activation/test_relu.py::Test_ReLU::test_op[tilecpp-256-256-torch.float16]": 0.3177552139999875,
+"tests/ops/activation/test_relu.py::Test_ReLU::test_op[tilecpp-256-32768-torch.float16]": 0.28071671100002504,
+"tests/ops/activation/test_relu.py::Test_ReLU::test_op[tilecpp-256-64-torch.float16]": 1.6863623780000125,
+"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_backward_corner_cases[cutile-1-1-64-64-True-torch.float16]": 22.292557976000012,
+"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_backward_corner_cases[cutile-1-32-128-64-True-torch.float16]": 20.35623326199999,
+"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_backward_irregular_shapes[cutile-1-4-127-64-True-torch.float16]": 20.72181408000003,
+"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_backward_irregular_shapes[cutile-2-8-2047-128-True-torch.bfloat16]": 19.225585593999995,
+"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_backward_regular_shapes[cutile-2-8-256-64-True-torch.float16]": 22.221181735000016,
+"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_backward_regular_shapes[cutile-2-8-512-128-False-torch.bfloat16]": 17.165542563000002,
+"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_forward_with_lse_matches_reference[cutile-1-8-256-128-True-torch.float16]": 1.804993152999998,
+"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_forward_with_lse_matches_reference[cutile-2-8-512-64-False-torch.bfloat16]": 1.0849780659999624,
+"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_functional_api_inference_mode[cutile]": 2.2683870100000263,
+"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward::test_op_functional_api_training_mode[cutile]": 20.976387465000073,
+"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward_GQA::test_op_backward_gqa[cutile-1-8-1-128-64-True-torch.float16]": 23.80138439199999,
+"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward_GQA::test_op_backward_gqa[cutile-1-8-4-128-64-True-torch.float16]": 20.238588308000033,
+"tests/ops/experimental/test_attention_backward.py::Test_FMHA_Backward_Numerical::test_op_numerical_gradient_check[cutile]": 0.2040473250000332,
+"tests/ops/experimental/test_fused_linear_cross_entropy.py::TestFusedLinearCrossEntropy::test_op_chunk_size_consistency[cutile]": 0.8741912499999671,
+"tests/ops/experimental/test_fused_linear_cross_entropy.py::TestFusedLinearCrossEntropy::test_op_forward_matches_pytorch[cutile-1-256-384-4096-torch.float16-sum]": 0.40579074800001536,
+"tests/ops/experimental/test_fused_linear_cross_entropy.py::TestFusedLinearCrossEntropy::test_op_forward_matches_pytorch[cutile-2-128-256-2048-torch.float16-mean]": 0.7984253390000049,
+"tests/ops/experimental/test_fused_linear_cross_entropy.py::TestFusedLinearCrossEntropy::test_op_forward_matches_pytorch[cutile-2-64-256-2048-torch.bfloat16-mean]": 0.6397966540000652,
+"tests/ops/experimental/test_fused_linear_cross_entropy.py::TestFusedLinearCrossEntropy::test_op_peak_memory_less_than_pytorch[cutile]": 1.7168269320000036,
+"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op[cutile-torch.float32-2-512-1024-4096]": 2.5507270850000054,
+"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op[cutile-torch.float32-2-64-256-512]": 2.271181330000047,
+"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op[cutile-torch.float32-4-128-512-1024]": 2.1244005650000872,
+"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op[cutile-torch.float32-8-256-512-2048]": 3.2640309640000282,
+"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-11-127-750-1500]": 0.6866470389999222,
+"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-13-128-512-1024]": 0.17915703800002802,
+"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-13-333-500-1000]": 0.5431909759999485,
+"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-4-100-256-512]": 0.17987008599993715,
+"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-4-128-300-600]": 2.686751162000064,
+"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-4-128-500-1000]": 2.7286812100000475,
+"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-4-128-750-1500]": 3.877476143000081,
+"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-4-333-512-1024]": 0.46266575800007104,
+"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-7-100-300-600]": 0.49896254500004034,
+"tests/ops/experimental/test_fused_swiglu_backward.py::Test_FusedSwiGLUBackward::test_op_irregular[cutile-torch.float32-7-64-256-512]": 0.17064024100000097,
+"tests/ops/experimental/test_mhc.py::Test_MHC::test_op_mhc_apply_residual[cutile-128-2-1024]": 0.3713822119999577,
+"tests/ops/experimental/test_mhc.py::Test_MHC::test_op_mhc_apply_residual[cutile-128-4-1024]": 0.40023054599998886,
+"tests/ops/experimental/test_mhc.py::Test_MHC::test_op_mhc_apply_residual[cutile-64-8-2048]": 0.545054748000041,
+"tests/ops/experimental/test_mhc.py::Test_MHC::test_op_mhc_gemm_rms_scale_bf16_precision[cutile-128-1000-8]": 28.15871181899996,
+"tests/ops/experimental/test_mhc.py::Test_MHC::test_op_mhc_gemm_rms_scale_bf16_precision[cutile-128-1024-4]": 30.146112172000016,
+"tests/ops/experimental/test_mhc.py::Test_MHC::test_op_mhc_gemm_rms_scale_bf16_precision[cutile-128-1024-8]": 30.627178904999994,
+"tests/ops/experimental/test_mhc.py::Test_MHC::test_op_mhc_sinkhorn[cutile-256-4-torch.float32]": 0.38189294499989046,
+"tests/ops/experimental/test_mhc.py::Test_MHC::test_op_mhc_sinkhorn[cutile-256-8-torch.float32]": 0.31523472499998206,
+"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_input_dtypes[cutile-torch.bfloat16]": 0.0004234189999863247,
+"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_input_dtypes[cutile-torch.float16]": 0.0004080870000109371,
+"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_input_dtypes[cutile-torch.float32]": 0.0003923379999832832,
+"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_invalid_inputs[cutile]": 0.00042589699998529795,
+"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-1-64]": 0.0004201710000018011,
+"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-128-64]": 0.0006882359999735854,
+"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-128-80]": 0.0006694010000387607,
+"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-160-80]": 0.0008795259999772043,
+"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-17-64]": 0.000444087000005311,
+"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-192-64]": 0.0004234499999711261,
+"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-256-48]": 0.0008766539999669476,
+"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-33-48]": 0.00042128199999069693,
+"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-512-256]": 0.0006528669999852355,
+"tests/ops/experimental/test_nvfp4_quantize.py::TestNVFP4Quantize::test_op_output_shapes[cutile-64-64]": 0.0004596009999886519,
+"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-None-256-256-torch.float16]": 0.3211140059999593,
+"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-None-256-256-torch.float32]": 0.3249114569999847,
+"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-None-4096-256-torch.bfloat16]": 0.511558253000004,
+"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-multi_wave_cached-256-256-torch.float16]": 0.289372123999982,
+"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-multi_wave_cached-256-256-torch.float32]": 0.2937340020000079,
+"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-multi_wave_cached-4096-256-torch.bfloat16]": 0.3022755929999903,
+"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-multi_wave_reload-256-256-torch.float16]": 0.16900348800001552,
+"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-multi_wave_reload-256-256-torch.float32]": 0.16160941600000456,
+"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-multi_wave_reload-4096-256-torch.bfloat16]": 0.3264167529999895,
+"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-static_persistent-256-256-torch.float16]": 0.5072222229999852,
+"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-static_persistent-256-256-torch.float32]": 1.700077738999994,
+"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormAutogradBackward::test_op[cutile-static_persistent-4096-256-torch.bfloat16]": 0.16868510500000866,
+"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormBackward::test_op[cutile-2003-2001-torch.float16]": 1.9460601870000005,
+"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormBackward::test_op[cutile-256-256-torch.float16]": 2.0134191850000036,
+"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormBackward::test_op[cutile-256-256-torch.float32]": 0.6046953809999707,
+"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormBackward::test_op[cutile-31072-4096-torch.bfloat16]": 13.623849871999994,
+"tests/ops/experimental/test_rmsnorm_backward.py::Test_RMSNormBackward::test_op[cutile-4096-256-torch.bfloat16]": 1.9930657619999863,
+"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_basic[cutile-torch.bfloat16-1-16-64-128-128-64-1-64]": 40.478604236999985,
+"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_basic[cutile-torch.bfloat16-1-64-128-256-128-64-1-256]": 90.23874729900001,
+"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_basic[cutile-torch.bfloat16-2-32-128-256-128-64-1-128]": 77.95847832099997,
+"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_gqa[cutile-torch.bfloat16-1-16-64-128-128-64-4-64]": 20.17174170300001,
+"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_gqa[cutile-torch.bfloat16-1-32-64-128-128-64-8-64]": 21.79622415199998,
+"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_invalid_kernel_configs[cutile]": 0.1812811510001211,
+"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_irregular_shapes[cutile-torch.bfloat16-1-16-1-128-128-64-1-64]": 43.02938131299993,
+"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_irregular_shapes[cutile-torch.bfloat16-1-16-33-77-128-64-1-64]": 39.75952677599997,
+"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_tile_h_configs[cutile-torch.bfloat16-16-1-16-32-TILE_H with small TILE_N]": 0.16745616500008964,
+"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_tile_h_configs[cutile-torch.bfloat16-16-1-4-64-small TILE_H]": 0.18799661499997455,
+"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_tile_h_configs[cutile-torch.bfloat16-16-16-1-64-1:1 mapping fallback]": 1.197208326000009,
+"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_tile_h_configs[cutile-torch.bfloat16-16-4-4-64-GQA group_size=4]": 0.25032408799995665,
+"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_tile_h_configs[cutile-torch.bfloat16-18-6-1-64-odd prime group_size=3 fallback]": 1.218409630999986,
+"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_tile_h_configs[cutile-torch.bfloat16-24-4-2-64-non_pow2 group_size=6, TILE_H=2]": 1.3397402149998925,
+"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_tile_h_configs[cutile-torch.bfloat16-24-8-1-64-non_pow2 group_size=3 fallback]": 1.114458710000008,
+"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_tile_h_configs[cutile-torch.bfloat16-64-1-16-64-large TILE_H]": 0.25679259900027773,
+"tests/ops/experimental/test_sparse_mla.py::Test_SparseMLA::test_op_topk_equals_skv[cutile-torch.bfloat16-1-16-64-128-64-1]": 0.17360443900008704,
+"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_first_tokens[cutile]": 0.15809969700001147,
+"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_gqa[cutile]": 0.16348675400001866,
+"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_long_context_mistral[cutile]": 0.15155569199998808,
+"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_multi_head[cutile]": 0.1660481430000118,
+"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_seq_not_divisible_by_tile[cutile]": 0.16325673700009702,
+"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_small_window[cutile]": 0.17103788399992936,
+"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_various_configs[cutile-1024-512]": 0.1548568929999874,
+"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_various_configs[cutile-2048-1024]": 0.1547801699999809,
+"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_various_configs[cutile-4096-2048]": 0.14675246799993147,
+"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_various_configs[cutile-4096-4096]": 0.14512579699987782,
+"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_various_configs[cutile-512-256]": 0.15543424000009054,
+"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_very_small_seq[cutile]": 0.16847861400015063,
+"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_window_equals_seq[cutile]": 1.3499497870000141,
+"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_window_equals_tile[cutile]": 0.15631962300005853,
+"tests/ops/experimental/test_swa_attention.py::TestSWAAttention::test_op_window_of_one[cutile]": 0.1697301280000829,
+"tests/ops/test_attention.py::Test_FMHA::test_op[cutile-1-1-9-128-False-torch.bfloat16]": 0.001681533999999374,
+"tests/ops/test_attention.py::Test_FMHA::test_op[cutile-1-32-2047-128-True-torch.float16]": 0.001577050999912899,
+"tests/ops/test_attention.py::Test_FMHA::test_op[cutile-2-32-4095-128-True-torch.bfloat16]": 0.001568184999996447,
+"tests/ops/test_attention.py::Test_FMHA::test_op[cutile-2-32-4095-128-True-torch.float8_e5m2]": 0.0017434369998454713,
+"tests/ops/test_attention.py::Test_FMHA::test_op[tilecpp-1-1-9-128-False-torch.bfloat16]": 0.0015903529999832244,
+"tests/ops/test_attention.py::Test_FMHA::test_op[tilecpp-1-32-2047-128-True-torch.float16]": 0.0016021320000163541,
+"tests/ops/test_attention.py::Test_FMHA::test_op[tilecpp-2-32-4095-128-True-torch.bfloat16]": 0.0015858869999192393,
+"tests/ops/test_attention.py::Test_FMHA::test_op[tilecpp-2-32-4095-128-True-torch.float8_e5m2]": 0.0017568759999448957,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-128-0.125-64-8-8-128-1-1]": 37.59067348799999,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-128-0.125-64-8-8-128-1-2]": 36.420128318000025,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-128-0.125-64-8-8-128-128-1]": 29.82492418000004,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-128-0.125-64-8-8-128-128-2]": 1.046193518999928,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-128-0.125-64-8-8-32-1-1]": 35.16873793500008,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-128-0.125-64-8-8-32-1-2]": 34.25014749699994,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-128-0.125-64-8-8-32-128-1]": 0.0012103069999511717,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-128-0.125-64-8-8-32-128-2]": 0.0011389329999929032,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-None-0.125-64-8-8-128-1-1]": 1.5381478489999836,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-None-0.125-64-8-8-128-1-2]": 1.5253856919999862,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-None-0.125-64-8-8-128-128-1]": 1.7305600620001087,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-None-0.125-64-8-8-128-128-2]": 1.7053485010001168,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-None-0.125-64-8-8-32-1-1]": 1.4262797960001308,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-None-0.125-64-8-8-32-1-2]": 35.46572530200001,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-None-0.125-64-8-8-32-128-1]": 0.0024891709999508294,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-0-None-0.125-64-8-8-32-128-2]": 0.0022651120000034553,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-128-0.125-64-8-8-128-1-1]": 0.14492825000013454,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-128-0.125-64-8-8-128-1-2]": 0.13646925399996235,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-128-0.125-64-8-8-128-128-1]": 0.13789151100002073,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-128-0.125-64-8-8-128-128-2]": 0.13273738900011267,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-128-0.125-64-8-8-32-1-1]": 0.1355514970000513,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-128-0.125-64-8-8-32-1-2]": 0.14167446000010386,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-128-0.125-64-8-8-32-128-1]": 0.001205584999979692,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-128-0.125-64-8-8-32-128-2]": 0.0010413839999046104,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-None-0.125-64-8-8-128-1-1]": 33.76884441799996,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-None-0.125-64-8-8-128-1-2]": 34.000542729000074,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-None-0.125-64-8-8-128-128-1]": 27.04095373399997,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-None-0.125-64-8-8-128-128-2]": 1.1519550399999616,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-None-0.125-64-8-8-32-1-1]": 30.92734718099996,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-None-0.125-64-8-8-32-1-2]": 0.16494574599994394,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-None-0.125-64-8-8-32-128-1]": 0.0013743129999284065,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[cutile-5-None-0.125-64-8-8-32-128-2]": 0.0013006359999963024,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-128-0.125-64-8-8-128-1-1]": 4.7122802850001335,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-128-0.125-64-8-8-128-1-2]": 0.1554573519998712,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-128-0.125-64-8-8-128-128-1]": 0.1528295389999812,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-128-0.125-64-8-8-128-128-2]": 0.15234006000002864,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-128-0.125-64-8-8-32-1-1]": 0.16690589800009548,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-128-0.125-64-8-8-32-1-2]": 0.17571100700001807,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-128-0.125-64-8-8-32-128-1]": 0.001183490000130405,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-128-0.125-64-8-8-32-128-2]": 0.0011042170000337137,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-None-0.125-64-8-8-128-1-1]": 5.162757644000067,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-None-0.125-64-8-8-128-1-2]": 0.17277970799989362,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-None-0.125-64-8-8-128-128-1]": 0.1565598439999576,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-None-0.125-64-8-8-128-128-2]": 0.16205757300008372,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-None-0.125-64-8-8-32-1-1]": 0.1578552960000934,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-None-0.125-64-8-8-32-1-2]": 0.16684363999991092,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-None-0.125-64-8-8-32-128-1]": 0.002008330000080605,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-0-None-0.125-64-8-8-32-128-2]": 0.0018347350001022278,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-128-0.125-64-8-8-128-1-1]": 0.14848934799999824,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-128-0.125-64-8-8-128-1-2]": 0.14855583900009606,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-128-0.125-64-8-8-128-128-1]": 0.15178161200003615,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-128-0.125-64-8-8-128-128-2]": 0.15233574000001227,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-128-0.125-64-8-8-32-1-1]": 0.14500215699990804,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-128-0.125-64-8-8-32-1-2]": 0.13962041499996758,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-128-0.125-64-8-8-32-128-1]": 0.0011742460000050414,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-128-0.125-64-8-8-32-128-2]": 0.0024215420000928134,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-None-0.125-64-8-8-128-1-1]": 0.14274271399995087,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-None-0.125-64-8-8-128-1-2]": 0.12423686499994346,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-None-0.125-64-8-8-128-128-1]": 0.13142085400011183,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-None-0.125-64-8-8-128-128-2]": 0.1334278979999226,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-None-0.125-64-8-8-32-1-1]": 0.1339199539999072,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-None-0.125-64-8-8-32-1-2]": 0.14244148200009477,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-None-0.125-64-8-8-32-128-1]": 0.0010990969999511435,
+"tests/ops/test_attention_sink.py::Test_AttentionSink::test_op[tilecpp-5-None-0.125-64-8-8-32-128-2]": 0.0010480550000693256,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0033111399999654623,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.003445837000015217,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.003378058000009787,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0033322069999996984,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.003272797000022365,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.003257751999967695,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0032800820000318254,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.003331474999981765,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.003294052000001102,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.003295795000013868,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0032025030000113475,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0033396440000217353,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.0020987810000008267,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0020558459999620027,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.002028514999977915,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.002026446999991549,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.002049733999996306,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0020694729999490846,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.002709654000000228,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0020728979999944386,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.002097547999994731,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.002095107000002372,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0020455660000209264,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-False-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0035435510000354498,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.002155022000010831,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0020823140000061358,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0021187520000296445,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0021068480000110412,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0021080090000111795,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.002086411999982829,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0021210109999856286,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0023914619999914066,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.0021564280000347935,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.002301778000003196,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.002110341999980392,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.002057349000040176,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.002075552000007974,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.002080084000027682,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0020869690000040464,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.002058727999980192,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.002039312999983167,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0022134820000019317,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0020745589999933145,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0020248079999589663,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.002060512000014114,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.002015518999996857,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0036230400000079044,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-128-True-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0031527250000067397,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0020091259999901467,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.002044472000022779,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.002073412000015651,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.00202067499998293,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.002037557000022616,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.002067136999983177,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0020020209999813687,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0022066029999905368,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.0020060479999983727,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0020140040000171666,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.002035118999970109,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0020649089999835724,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.001998543999974345,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.002018500000019685,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0020132969999906436,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.002040475000001152,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.002015841000002183,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0020120159999805765,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0020247100000005958,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0020225990000426464,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.002026152000013326,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0019971799999893847,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0020051619999890136,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-False-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0019850509999912447,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0020554760000095484,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0020242440000117767,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.002047147000013183,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0021798459999899933,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.002291025000033642,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0019979730000443396,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.003071691999991799,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.002075096999988091,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.002183148999961304,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.002062198999993825,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0020729089999633743,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.002015029999995477,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.0020822070000008353,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.002100625999986505,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0020476259999782087,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.002033247000014171,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.002030483999988064,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0020026140000197756,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.002028536000011627,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.002034033000029467,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.002047844999992776,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0020605750000015632,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.002029554000017697,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-2064-64-True-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.002081242999992128,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.002007620000000543,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.003059592999989036,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0020401470000024347,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.002015664999987621,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0022569799999985207,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.002030381000025727,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0020113570000148684,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.001978188000038017,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.0020135019999827364,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.001983581999979833,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0020364129999848046,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0020510630000103447,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.002027510000004895,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.002067702000005056,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.002004394000039156,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.001985573000013119,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0019991000000061376,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.001992733000008684,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0020370910000053755,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.002023081999993792,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.002025456000041004,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0020426170000291677,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0019970879999959834,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-False-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0020110919999751786,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0020255339999835087,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.002065194000010706,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.002051881999989291,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0020608069999923373,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0020234330000050704,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0020693090000065695,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.002025927999994792,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0022135459999788054,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.0020430330000067443,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0019963409999945725,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0021244859999853816,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0020057870000016464,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.002055396999992354,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0020155209999757062,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.002019487999973535,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0019960340000011456,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.002087406000015335,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0020167049999884057,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0025290809999773955,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0020292239999832873,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.002038181999949984,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.002023368999999775,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.001994535999983782,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-128-True-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0020063020000122833,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.002002328999992642,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.002038584999951354,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0020235470000216083,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0019897370000308,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.00203477899998461,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0022072910000190404,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.002172128000012208,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0020923010000046816,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.002020000999976901,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0020139100000164945,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0019843839999680313,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.002082663000010143,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.0020956890000149997,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0020338510000215138,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0019906610000077762,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.001969864999978199,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.001994362999965915,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0019974959999444764,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0020155789999876106,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0020170070000347096,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0020058029999745486,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.002029351999993878,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0019873199999835833,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-False-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.002258667999996078,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.00202678700000547,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0019883940000227085,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0020133499999985816,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0020098270000232787,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.002055388999991692,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.002003490000021202,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.002050883999999087,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.002042052999996713,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.0020312419999868325,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0020181419999687478,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0020141360000138775,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.001965745000035213,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.002068508000093061,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0021128619999899456,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0021431449999624874,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0022050370000101793,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0021995200000048953,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0020461129999773675,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0022922429999994165,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0020532339999590477,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0023429819999591928,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0022206479998203577,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.002039761000020235,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-128-496-64-True-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.002039187999969272,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0018352340000262757,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0019927399999914996,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.002104698000010785,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0019881349999764097,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0017347780000136481,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.001992847000025222,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0017606070000226737,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0017436709999856248,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.001840114000003723,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0019437139999638475,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0019941379999863784,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0020524079999972855,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.0020041650000166555,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0023279510000122627,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.002021519999999555,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0020027090000098724,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0019978590000562235,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0019113080000181526,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0019955609999726676,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0020216799999843715,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0020226549999904364,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0019915610000111883,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0021158300000081454,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-False-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.001821522999989611,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.001993018000007396,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0020134349999807455,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0019939569999962714,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0020198459999960505,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0019948510000062925,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0020018200000038178,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.002034922999968103,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0020101769999598673,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.0021637529999622984,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0020677699999964716,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0021419150000383524,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0019979190000185554,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.002042363000015257,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.002054952999969828,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.002016891000010901,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.001999179999984335,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0020195359999490847,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0020385990000022502,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0020593620000113333,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0020612509999864415,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0020486660000074153,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0020523729999695206,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0020756169999742724,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-128-True-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.003041631999991523,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.002648849999985714,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.002030364000006557,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0022351380000031895,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.002033734999997705,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0020205619999842384,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0020373579999954927,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.002023379000007708,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.002045298000012963,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.002034279000014294,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0020807160000231306,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.002017622999943569,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0028267979999725412,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.0020410159999926236,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0021049399999810703,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0021434740000074726,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0022555129999659584,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0020147980000047028,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0020789049999905274,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.002022958000026165,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0020205499999974563,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.002038165000016079,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.002070540000005394,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0021047460000147566,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-False-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.002068911999970169,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.002030851000029088,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0020490210000332354,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0020036119999815583,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0020487750000199867,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.002049203000012767,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0020476709999854847,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.002064925000013318,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0020227239999712765,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.001998593000053006,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0022750430000257893,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0020252700000185087,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.002246470000017098,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.002008227000004581,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0020862499999907413,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0020321509999519094,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.002059065999986842,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.002023154000028171,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.002012067000009665,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.002039106000012225,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.0020886350000068887,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0020646289999888268,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0021889239999950405,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0020237809999912315,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-2064-64-True-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0020367560000238427,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0020249600000283863,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0020409430000256634,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.002044167999997626,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0020204900000351245,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.002053212999982179,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0020057169999745383,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0020290700000202833,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0020298899999886544,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.002059701000007408,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0020434099999988575,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0020264890000021296,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0020709819999922274,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.0021873810000272442,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.002050887999985207,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0020612410000069303,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0020374629999935223,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0021649919999902068,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0020584439999993265,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0020061429999884695,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.002002727000018467,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.00205104899998787,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0020198160000006737,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.002072964000007005,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-False-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.002024974999983442,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.002039780000018254,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0020407560000137437,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0020435510000140766,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.0019932980000021416,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0020198200000152156,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0021972169999742164,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0020076169999754256,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.002006599000026199,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.002075790999981564,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0020555580000234386,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.002056241000019554,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0020679120000011153,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.002052554999977474,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.0021240119999959006,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0021010670000123355,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0021420989999967333,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.0021047529999975723,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0021132589999979245,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0021507250000070144,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.002131171000030463,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0021435669999618767,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.002081390999961741,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0021475640000119256,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-128-True-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0020860820000052627,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.002162961000010455,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.002102188000009164,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.0021478980000040337,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.002207290000001194,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0021190630000091915,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0021613540000089415,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.002175988000004736,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.002721696000008933,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.0021267609999426895,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.002124215000009144,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0022113710000155606,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0021061709999514733,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.002192079999986163,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.002226339999992888,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0021472399999993286,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0021439730000167856,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.002106931000014356,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0021145309999894835,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0021732989999918573,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.002266470999984449,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0021405500000071243,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.00216096799999832,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.002153974999998809,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-False-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.0021735660000103962,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-False]": 0.0021270740000147725,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-False-None-0.0-bnsd-True]": 0.0021332579999864265,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-False-None-0.0-nsbd-False]": 0.002183549000022822,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-False-None-0.1-bnsd-False]": 0.002111512999988463,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-False-alibi-0.0-bnsd-False]": 0.0022865079999974114,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-False-alibi-0.1-nsbd-True]": 0.0023218880000115405,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-False-matrix-0.0-bnsd-False]": 0.0021200080000198795,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-False-vector-0.0-bnsd-False]": 0.0020984110000199507,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-0-True-None-0.0-bnsd-False]": 0.002144528000030732,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-512-False-None-0.0-bnsd-False]": 0.0023370730000067397,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-512-True-None-0.0-nsbd-True]": 0.0021874619999948663,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-cutile-512-True-matrix-0.1-nsbd-True]": 0.0021003319999977066,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-False]": 0.003439524000015126,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-False-None-0.0-bnsd-True]": 0.002111827000021549,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-False-None-0.0-nsbd-False]": 0.0026962190000006103,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-False-None-0.1-bnsd-False]": 0.0020549009999797363,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-False-alibi-0.0-bnsd-False]": 0.002010432000020046,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-False-alibi-0.1-nsbd-True]": 0.0021596889999955238,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-False-matrix-0.0-bnsd-False]": 0.0023364360000073248,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-False-vector-0.0-bnsd-False]": 0.003387417999988429,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-0-True-None-0.0-bnsd-False]": 0.0034404779999590573,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-512-False-None-0.0-bnsd-False]": 0.0033720239999865953,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-512-True-None-0.0-nsbd-True]": 0.0021125370000163457,
+"tests/ops/test_attention_variant.py::Test_FMHA_variant::test_op[4-32-1-257-496-64-True-torch.bfloat16-triton-512-True-matrix-0.1-nsbd-True]": 0.002442146999982242,
+"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-128-2064-128-False-torch.bfloat16-False]": 0.0018007719999957317,
+"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-128-2064-128-False-torch.bfloat16-True]": 0.0017700949999834847,
+"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-128-2064-128-True-torch.bfloat16-False]": 0.00174520399997391,
+"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-128-2064-128-True-torch.bfloat16-True]": 0.0017745529999899645,
+"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-128-496-128-False-torch.bfloat16-False]": 0.0017873889999862058,
+"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-128-496-128-False-torch.bfloat16-True]": 0.0017492689999869526,
+"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-128-496-128-True-torch.bfloat16-False]": 0.0019646009999689795,
+"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-128-496-128-True-torch.bfloat16-True]": 0.0017558140000062394,
+"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-257-2064-128-False-torch.bfloat16-False]": 0.0017666520000148012,
+"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-257-2064-128-False-torch.bfloat16-True]": 0.0017474560000323436,
+"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-257-2064-128-True-torch.bfloat16-False]": 0.0017422369999735565,
+"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-257-2064-128-True-torch.bfloat16-True]": 0.0019524389999503455,
+"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-257-496-128-False-torch.bfloat16-False]": 0.0017388880000055451,
+"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-257-496-128-False-torch.bfloat16-True]": 0.0017686110000170174,
+"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-257-496-128-True-torch.bfloat16-False]": 0.0017292920000215872,
+"tests/ops/test_attention_varlen.py::Test_FMHA_varlen::test_op[cutile-4-32-1-257-496-128-True-torch.bfloat16-True]": 0.0017664150000200607,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-1023-False-False-torch.float16]": 0.1367885909999984,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-1023-False-True-torch.float16]": 0.13776529500000834,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-1023-True-False-torch.float16]": 0.14198892699999988,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-1023-True-True-torch.float16]": 0.1371275180000282,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-1024-False-False-torch.float16]": 0.13740175199998816,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-1024-False-True-torch.float16]": 0.1357802660000118,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-1024-True-False-torch.float16]": 0.1346945320000259,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-1024-True-True-torch.float16]": 0.13566309199995885,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-511-False-False-torch.float16]": 0.1398865609999973,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-511-False-True-torch.float16]": 0.13600638299996604,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-511-True-False-torch.float16]": 0.14049500199999443,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-511-True-True-torch.float16]": 0.1373220949999734,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-512-False-False-torch.float16]": 0.14117118100003268,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-512-False-True-torch.float16]": 0.14052478599998608,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-512-True-False-torch.float16]": 0.1394910829999958,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-256-512-True-True-torch.float16]": 0.13813763400000312,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-1023-False-False-torch.float16]": 0.1778659689999813,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-1023-False-True-torch.float16]": 0.17555543200001011,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-1023-True-False-torch.float16]": 0.1752707920000205,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-1023-True-True-torch.float16]": 0.17451101300002847,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-1024-False-False-torch.float16]": 0.1824153069999852,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-1024-False-True-torch.float16]": 0.17335845299999164,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-1024-True-False-torch.float16]": 0.168717934,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-1024-True-True-torch.float16]": 0.17543948399998044,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-511-False-False-torch.float16]": 2.186541876000007,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-511-False-True-torch.float16]": 0.14735276800001884,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-511-True-False-torch.float16]": 0.14272245099999736,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-511-True-True-torch.float16]": 0.1415517260000172,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-512-False-False-torch.float16]": 0.2084274260000143,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-512-False-True-torch.float16]": 0.17686461400001008,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-512-True-False-torch.float16]": 0.17451376100001426,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-1024-512-512-True-True-torch.float16]": 0.17753269800002158,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-1023-False-False-torch.float16]": 0.16106736800003318,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-1023-False-True-torch.float16]": 0.15554205599994475,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-1023-True-False-torch.float16]": 0.1490688249999721,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-1023-True-True-torch.float16]": 0.14628613399997903,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-1024-False-False-torch.float16]": 0.15355599599996594,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-1024-False-True-torch.float16]": 0.15145123700006025,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-1024-True-False-torch.float16]": 0.1460192380000649,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-1024-True-True-torch.float16]": 0.15270596600004183,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-511-False-False-torch.float16]": 0.17738789199989924,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-511-False-True-torch.float16]": 0.14701637799998934,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-511-True-False-torch.float16]": 0.15469931100000167,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-511-True-True-torch.float16]": 0.17150166600004013,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-512-False-False-torch.float16]": 0.1683891999999787,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-512-False-True-torch.float16]": 0.14067571200007478,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-512-True-False-torch.float16]": 0.14513682399996242,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-256-512-True-True-torch.float16]": 0.14881510299994716,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-1023-False-False-torch.float16]": 0.15484318800000096,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-1023-False-True-torch.float16]": 0.13916746999998963,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-1023-True-False-torch.float16]": 0.1367332529999885,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-1023-True-True-torch.float16]": 0.13938769599997158,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-1024-False-False-torch.float16]": 0.14542050800002926,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-1024-False-True-torch.float16]": 0.1364735479999979,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-1024-True-False-torch.float16]": 0.13608420900004603,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-1024-True-True-torch.float16]": 0.13969830499999603,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-511-False-False-torch.float16]": 0.1498679480000078,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-511-False-True-torch.float16]": 0.1525021049999964,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-511-True-False-torch.float16]": 0.1492159679999645,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-511-True-True-torch.float16]": 0.151932520999992,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-512-False-False-torch.float16]": 1.7911859200000038,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-512-False-True-torch.float16]": 0.15287007800003494,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-512-True-False-torch.float16]": 0.14512648399994532,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-False-4-128-512-512-True-True-torch.float16]": 0.1451350530000468,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-1023-False-False-torch.float16]": 72.28453170900002,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-1023-False-True-torch.float16]": 293.03343604500014,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-1023-True-False-torch.float16]": 19.642332332000024,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-1023-True-True-torch.float16]": 26.778804793000063,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-1024-False-False-torch.float16]": 16.036066987999902,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-1024-False-True-torch.float16]": 15.331060702000059,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-1024-True-False-torch.float16]": 19.46626863000006,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-1024-True-True-torch.float16]": 15.6132359799999,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-511-False-False-torch.float16]": 158.290498696,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-511-False-True-torch.float16]": 185.9608495980001,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-511-True-False-torch.float16]": 17.678049858999998,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-511-True-True-torch.float16]": 31.890701977999925,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-512-False-False-torch.float16]": 22.208092426999997,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-512-False-True-torch.float16]": 25.504170145000103,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-512-True-False-torch.float16]": 21.936462887999937,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-256-512-True-True-torch.float16]": 22.00874279900006,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-1023-False-False-torch.float16]": 68.85947150999999,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-1023-False-True-torch.float16]": 301.1968222090002,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-1023-True-False-torch.float16]": 14.533433569999943,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-1023-True-True-torch.float16]": 27.57549889599977,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-1024-False-False-torch.float16]": 19.843438633999938,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-1024-False-True-torch.float16]": 19.253786158999986,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-1024-True-False-torch.float16]": 15.05535700499999,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-1024-True-True-torch.float16]": 16.458007624999993,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-511-False-False-torch.float16]": 160.38439373200026,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-511-False-True-torch.float16]": 182.57038911800032,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-511-True-False-torch.float16]": 24.691955391000192,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-511-True-True-torch.float16]": 31.567412470000136,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-512-False-False-torch.float16]": 22.671363233000193,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-512-False-True-torch.float16]": 19.205410948000235,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-512-True-False-torch.float16]": 19.99135865500034,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-1024-512-512-True-True-torch.float16]": 21.548341671000117,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-1023-False-False-torch.float16]": 62.36220984099998,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-1023-False-True-torch.float16]": 300.003211932,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-1023-True-False-torch.float16]": 19.136103876999982,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-1023-True-True-torch.float16]": 32.640933915999994,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-1024-False-False-torch.float16]": 19.052600523000024,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-1024-False-True-torch.float16]": 18.683894199000065,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-1024-True-False-torch.float16]": 21.40511845399999,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-1024-True-True-torch.float16]": 15.417424671000049,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-511-False-False-torch.float16]": 159.50202545899998,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-511-False-True-torch.float16]": 189.91430599199998,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-511-True-False-torch.float16]": 27.43602206700001,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-511-True-True-torch.float16]": 39.46324411499998,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-512-False-False-torch.float16]": 25.12586579399988,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-512-False-True-torch.float16]": 21.492233179000095,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-512-True-False-torch.float16]": 19.37927721899996,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-256-512-True-True-torch.float16]": 22.18841402199996,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-1023-False-False-torch.float16]": 64.64629589799995,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-1023-False-True-torch.float16]": 298.1146448059999,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-1023-True-False-torch.float16]": 20.998536285,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-1023-True-True-torch.float16]": 36.851745059999985,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-1024-False-False-torch.float16]": 21.08459634199994,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-1024-False-True-torch.float16]": 18.532988854999985,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-1024-True-False-torch.float16]": 21.043876144000023,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-1024-True-True-torch.float16]": 20.96114934900004,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-511-False-False-torch.float16]": 157.19289479899976,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-511-False-True-torch.float16]": 200.78464333700003,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-511-True-False-torch.float16]": 27.559489696000128,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-511-True-True-torch.float16]": 32.45479717499995,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-512-False-False-torch.float16]": 25.236537768000062,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-512-False-True-torch.float16]": 20.350812722,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-512-True-False-torch.float16]": 23.464420715999836,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[cutile-True-4-128-512-512-True-True-torch.float16]": 21.000315984000054,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-1023-False-False-torch.float16]": 0.1640439140000467,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-1023-False-True-torch.float16]": 0.16821766899997215,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-1023-True-False-torch.float16]": 0.16785368099996845,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-1023-True-True-torch.float16]": 0.16926906599999825,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-1024-False-False-torch.float16]": 0.1507112280000058,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-1024-False-True-torch.float16]": 0.15516907899998955,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-1024-True-False-torch.float16]": 0.15487382400002048,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-1024-True-True-torch.float16]": 0.15343766500001266,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-511-False-False-torch.float16]": 0.16452163900001437,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-511-False-True-torch.float16]": 0.16840741099997558,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-511-True-False-torch.float16]": 0.16231652499999427,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-511-True-True-torch.float16]": 0.1637238719999914,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-512-False-False-torch.float16]": 0.16659715200003689,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-512-False-True-torch.float16]": 0.16889704200002598,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-512-True-False-torch.float16]": 0.16437081199998715,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-256-512-True-True-torch.float16]": 0.16626172700000552,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-1023-False-False-torch.float16]": 1.8095410950000144,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-1023-False-True-torch.float16]": 2.579071354000007,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-1023-True-False-torch.float16]": 1.4798257849999743,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-1023-True-True-torch.float16]": 0.14921134799999436,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-1024-False-False-torch.float16]": 0.17516094299995189,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-1024-False-True-torch.float16]": 0.17423673899997993,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-1024-True-False-torch.float16]": 0.17579161400001908,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-1024-True-True-torch.float16]": 0.18119660100001056,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-511-False-False-torch.float16]": 0.14196331499999815,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-511-False-True-torch.float16]": 0.1230539070000134,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-511-True-False-torch.float16]": 0.14370655300001545,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-511-True-True-torch.float16]": 0.15421638700004792,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-512-False-False-torch.float16]": 0.15583197300000506,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-512-False-True-torch.float16]": 0.16159665600002882,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-512-True-False-torch.float16]": 0.1535261659999776,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-1024-512-512-True-True-torch.float16]": 0.15348742699998752,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-1023-False-False-torch.float16]": 0.1637744699999928,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-1023-False-True-torch.float16]": 0.1656980420000309,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-1023-True-False-torch.float16]": 0.16371149799999785,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-1023-True-True-torch.float16]": 0.16326224599998795,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-1024-False-False-torch.float16]": 0.16263321700000688,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-1024-False-True-torch.float16]": 0.16588081499997998,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-1024-True-False-torch.float16]": 0.16171916100000772,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-1024-True-True-torch.float16]": 0.16162931099998445,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-511-False-False-torch.float16]": 1.557925076000032,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-511-False-True-torch.float16]": 0.1589915469999994,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-511-True-False-torch.float16]": 0.16939636600002927,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-511-True-True-torch.float16]": 2.984922915999988,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-512-False-False-torch.float16]": 0.16706404999999336,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-512-False-True-torch.float16]": 0.1706362889999582,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-512-True-False-torch.float16]": 0.16593739099999993,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-256-512-True-True-torch.float16]": 0.1712706910000179,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-1023-False-False-torch.float16]": 0.16989069499999232,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-1023-False-True-torch.float16]": 0.16792753900000434,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-1023-True-False-torch.float16]": 0.167781854000026,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-1023-True-True-torch.float16]": 0.17399706099999435,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-1024-False-False-torch.float16]": 0.16704937599996583,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-1024-False-True-torch.float16]": 0.17123698399998943,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-1024-True-False-torch.float16]": 0.16542418999998176,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-1024-True-True-torch.float16]": 0.16653658699999596,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-511-False-False-torch.float16]": 0.16634358100000668,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-511-False-True-torch.float16]": 0.16454152000000022,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-511-True-False-torch.float16]": 0.16309424199999967,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-511-True-True-torch.float16]": 0.1650026339999897,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-512-False-False-torch.float16]": 1.878442184999983,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-512-False-True-torch.float16]": 2.629644847999998,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-512-True-False-torch.float16]": 2.814757564999951,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-False-4-128-512-512-True-True-torch.float16]": 0.17103388800001085,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-1023-False-False-torch.float16]": 0.16634093400000438,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-1023-False-True-torch.float16]": 0.1666926270000033,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-1023-True-False-torch.float16]": 0.1703681719999679,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-1023-True-True-torch.float16]": 0.17084533499999566,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-1024-False-False-torch.float16]": 0.16510646899999415,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-1024-False-True-torch.float16]": 0.17090587800001344,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-1024-True-False-torch.float16]": 0.1679632479999782,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-1024-True-True-torch.float16]": 0.16741725700001098,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-511-False-False-torch.float16]": 0.16881491199998777,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-511-False-True-torch.float16]": 0.16840817400000674,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-511-True-False-torch.float16]": 0.16503399900000204,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-511-True-True-torch.float16]": 0.17108605800004284,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-512-False-False-torch.float16]": 0.16770613799997136,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-512-False-True-torch.float16]": 0.16557665400000587,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-512-True-False-torch.float16]": 0.16900695000001065,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-256-512-True-True-torch.float16]": 0.172259696000026,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-1023-False-False-torch.float16]": 0.15437842299999716,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-1023-False-True-torch.float16]": 0.15779088400000774,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-1023-True-False-torch.float16]": 0.16089461400002847,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-1023-True-True-torch.float16]": 0.1647540190000143,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-1024-False-False-torch.float16]": 0.15362475099999529,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-1024-False-True-torch.float16]": 0.15734883200002514,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-1024-True-False-torch.float16]": 0.14738469200003124,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-1024-True-True-torch.float16]": 0.15145667999999546,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-511-False-False-torch.float16]": 0.2014488119999669,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-511-False-True-torch.float16]": 0.16189061899999047,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-511-True-False-torch.float16]": 0.16523385499996834,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-511-True-True-torch.float16]": 0.15154169300004128,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-512-False-False-torch.float16]": 0.1605814829999872,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-512-False-True-torch.float16]": 0.1627329020000161,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-512-True-False-torch.float16]": 0.16153217199999403,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-1024-512-512-True-True-torch.float16]": 0.161152407000003,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-1023-False-False-torch.float16]": 0.17095211599996674,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-1023-False-True-torch.float16]": 0.16937471700001083,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-1023-True-False-torch.float16]": 0.1705921130000263,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-1023-True-True-torch.float16]": 0.1720550589999732,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-1024-False-False-torch.float16]": 0.17316532500001358,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-1024-False-True-torch.float16]": 0.17428626599999575,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-1024-True-False-torch.float16]": 0.1730934470000136,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-1024-True-True-torch.float16]": 0.17508152999999993,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-511-False-False-torch.float16]": 0.178083431999994,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-511-False-True-torch.float16]": 0.1729298559999961,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-511-True-False-torch.float16]": 0.17357469800001013,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-511-True-True-torch.float16]": 0.17317185799998924,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-512-False-False-torch.float16]": 0.18066334100001313,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-512-False-True-torch.float16]": 0.18120880899999747,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-512-True-False-torch.float16]": 0.17869864299998994,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-256-512-True-True-torch.float16]": 0.1763324209999837,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-1023-False-False-torch.float16]": 0.16605064599997377,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-1023-False-True-torch.float16]": 0.17000975799996354,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-1023-True-False-torch.float16]": 0.17330756100000144,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-1023-True-True-torch.float16]": 0.16938998699995977,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-1024-False-False-torch.float16]": 0.16958278399999926,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-1024-False-True-torch.float16]": 0.16748567999999864,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-1024-True-False-torch.float16]": 0.16604001599998242,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-1024-True-True-torch.float16]": 0.16722875800002157,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-511-False-False-torch.float16]": 0.14118510100001913,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-511-False-True-torch.float16]": 0.1559257860000116,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-511-True-False-torch.float16]": 0.16187293100000488,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-511-True-True-torch.float16]": 0.1574601410000298,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-512-False-False-torch.float16]": 0.14867667300003973,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-512-False-True-torch.float16]": 0.14893877899999097,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-512-True-False-torch.float16]": 0.14658159699999374,
+"tests/ops/test_bmm.py::Test_BMM_FWD::test_op[tilecpp-True-4-128-512-512-True-True-torch.float16]": 0.14304042100002334,
+"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill[2048-128-torch.bfloat16]": 0.0023823209999989103,
+"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill[2048-128-torch.float32]": 0.003073423999978786,
+"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill[2048-16-torch.bfloat16]": 0.17624465699998382,
+"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill[2048-16-torch.float32]": 0.2621569510000086,
+"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill[256-7-torch.bfloat16]": 0.2008582889999957,
+"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill[256-7-torch.float32]": 0.1662126850000334,
+"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill[300-33-torch.bfloat16]": 0.19624382399999263,
+"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill[300-33-torch.float32]": 0.2319032039999911,
+"tests/ops/test_causal_conv1d_lfm2.py::test_op_prefill_matches_update_stepwise": 0.17672456400001124,
+"tests/ops/test_causal_conv1d_lfm2.py::test_op_update[2048-torch.bfloat16]": 0.22842502199995351,
+"tests/ops/test_causal_conv1d_lfm2.py::test_op_update[2048-torch.float32]": 0.23967872899999065,
+"tests/ops/test_causal_conv1d_lfm2.py::test_op_update[256-torch.bfloat16]": 0.0021318540000265784,
+"tests/ops/test_causal_conv1d_lfm2.py::test_op_update[256-torch.float32]": 0.0023420360000159235,
+"tests/ops/test_causal_conv1d_lfm2.py::test_op_update[300-torch.bfloat16]": 0.23320846300001108,
+"tests/ops/test_causal_conv1d_lfm2.py::test_op_update[300-torch.float32]": 0.1841406360000235,
+"tests/ops/test_causal_conv1d_lfm2.py::test_op_update_2d_legacy_call[torch.bfloat16]": 0.0025542129999962526,
+"tests/ops/test_causal_conv1d_lfm2.py::test_op_update_2d_legacy_call[torch.float32]": 0.002640290000016421,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-K256_V128-bf16]": 42.909546011,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-K256_V128-fp32]": 0.0014528100000177346,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-K64_V128-bf16]": 5.273690520000031,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-K64_V128-fp32]": 0.002296516999990672,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T100_unaligned-bf16]": 0.18777529800007642,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T100_unaligned-fp32]": 0.0021648090000212505,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T16-bf16]": 7.371221793999979,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T16-fp32]": 0.0015462060000004385,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T2048_long_seq-bf16]": 15.962806328,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T2048_long_seq-fp32]": 0.0028713700000082554,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T32_final_state-bf16]": 26.026194939999954,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T32_final_state-fp32]": 0.0023816139999723873,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T37_prime-bf16]": 0.1829576819999943,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T37_prime-fp32]": 0.0020989679999843247,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T512_long_seq-bf16]": 14.435413848999985,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T512_long_seq-fp32]": 0.0021389739999619906,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T64-bf16]": 0.1827234910000186,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T64-fp32]": 0.0021577119999847127,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T65_off_by_1-bf16]": 34.650554639000006,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-T65_off_by_1-fp32]": 0.002033546000035358,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-chunk32-bf16]": 6.704753653000012,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-chunk32-fp32]": 0.002064903000018603,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-decode-bf16]": 11.276366806000027,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-decode-fp32]": 0.0013054399999816724,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-init_final_state-bf16]": 10.825378040999993,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-init_final_state-fp32]": 0.0021273239999857196,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-l2norm-bf16]": 12.165929416000012,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-l2norm-fp32]": 0.0014527590000170676,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-minimal-bf16]": 15.719284311000024,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-minimal-fp32]": 0.002151413999996521,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-qwen3_5_T10782-bf16]": 3.2421283999999844,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[cutile-qwen3_5_T10782-fp32]": 0.0014117320000082145,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-K256_V128-bf16]": 0.0022450049999065413,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-K256_V128-fp32]": 0.002014783000049647,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-K64_V128-bf16]": 13.038035407999985,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-K64_V128-fp32]": 0.003113435999921421,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T100_unaligned-bf16]": 23.912057498999957,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T100_unaligned-fp32]": 0.002129607000000533,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T16-bf16]": 0.16537801699999477,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T16-fp32]": 0.002530534000015905,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T2048_long_seq-bf16]": 0.0021499449999851095,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T2048_long_seq-fp32]": 0.0019690679999939675,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T32_final_state-bf16]": 14.624063643999989,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T32_final_state-fp32]": 0.0025309199999981047,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T37_prime-bf16]": 0.1400922889999947,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T37_prime-fp32]": 0.0012445810000087931,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T512_long_seq-bf16]": 0.002006980999993857,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T512_long_seq-fp32]": 0.0024551399999950263,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T64-bf16]": 0.17255476399998315,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T64-fp32]": 0.0023651289999975234,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T65_off_by_1-bf16]": 50.95820785699999,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-T65_off_by_1-fp32]": 0.0013118570000187901,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-chunk32-bf16]": 5.933981839999973,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-chunk32-fp32]": 0.0020504759999937505,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-decode-bf16]": 0.16569058999996855,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-decode-fp32]": 0.00259142099997689,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-init_final_state-bf16]": 18.547193610999955,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-init_final_state-fp32]": 0.0021863290000112556,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-l2norm-bf16]": 0.002208590000066124,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-l2norm-fp32]": 0.0013334970000755675,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-minimal-bf16]": 0.2069455170000083,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-minimal-fp32]": 0.002601741000006541,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-qwen3_5_T10782-bf16]": 0.002117755000000443,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op[tilecpp-qwen3_5_T10782-fp32]": 0.002077718999970557,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op_correlated_keys_stable_triangular_solve[cutile-in_kernel_l2]": 4.544428712000013,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op_correlated_keys_stable_triangular_solve[cutile-prenormalized]": 28.093120167999984,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op_correlated_keys_stable_triangular_solve[tilecpp-in_kernel_l2]": 0.0017776619999381182,
+"tests/ops/test_chunk_gated_delta_rule.py::Test_ChunkGatedDeltaRule::test_op_correlated_keys_stable_triangular_solve[tilecpp-prenormalized]": 15.422655924999958,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-1024-0.2-False-False-torch.float32-1e-05]": 0.16640614300001744,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-1024-0.2-False-True-torch.float32-1e-05]": 0.1947943109999528,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-1024-0.2-True-False-torch.float32-1e-05]": 0.17306771599999138,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-1024-0.2-True-True-torch.float32-1e-05]": 0.3258892079999782,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-4096-0.4-False-False-torch.float32-1e-05]": 0.16582739100010713,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-4096-0.4-False-True-torch.float32-1e-05]": 0.1681810740000742,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-4096-0.4-True-False-torch.float32-1e-05]": 0.33514501099995186,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-4096-0.4-True-True-torch.float32-1e-05]": 0.3299875369999654,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-65536-0.8-False-False-torch.float16-1e-05]": 0.173262464000004,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-65536-0.8-False-False-torch.float32-1e-05]": 0.17470351599996548,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-65536-0.8-False-True-torch.float16-1e-05]": 0.16198477899996533,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-65536-0.8-False-True-torch.float32-1e-05]": 0.17179194000004827,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-65536-0.8-True-False-torch.float16-1e-05]": 0.1674301599999808,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-65536-0.8-True-False-torch.float32-1e-05]": 0.17401968299998316,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-65536-0.8-True-True-torch.float16-1e-05]": 0.3389435249999906,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[cutile-65536-0.8-True-True-torch.float32-1e-05]": 0.35654895599998326,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-1024-0.2-False-False-torch.float32-1e-05]": 0.1359185529999536,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-1024-0.2-False-True-torch.float32-1e-05]": 0.20256426400004557,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-1024-0.2-True-False-torch.float32-1e-05]": 0.1767375359999619,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-1024-0.2-True-True-torch.float32-1e-05]": 0.9203006770000002,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-4096-0.4-False-False-torch.float32-1e-05]": 0.1703810629999225,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-4096-0.4-False-True-torch.float32-1e-05]": 0.18348954199996115,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-4096-0.4-True-False-torch.float32-1e-05]": 0.16464686600005507,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-4096-0.4-True-True-torch.float32-1e-05]": 0.9649258020000389,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-65536-0.8-False-False-torch.float16-1e-05]": 0.19165847000004987,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-65536-0.8-False-False-torch.float32-1e-05]": 0.2130497080000282,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-65536-0.8-False-True-torch.float16-1e-05]": 0.18657470900001272,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-65536-0.8-False-True-torch.float32-1e-05]": 0.18914590800000042,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-65536-0.8-True-False-torch.float16-1e-05]": 0.15955533699997204,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-65536-0.8-True-False-torch.float32-1e-05]": 0.17253355600001896,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-65536-0.8-True-True-torch.float16-1e-05]": 0.9456462230000398,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[tilecpp-65536-0.8-True-True-torch.float32-1e-05]": 0.8392195519999746,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-1024-0.2-False-False-torch.float32-1e-05]": 0.12948974700003646,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-1024-0.2-False-True-torch.float32-1e-05]": 0.14123438600000782,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-1024-0.2-True-False-torch.float32-1e-05]": 14.716151400000001,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-1024-0.2-True-True-torch.float32-1e-05]": 3.7071543010000028,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-4096-0.4-False-False-torch.float32-1e-05]": 0.1948086019999664,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-4096-0.4-False-True-torch.float32-1e-05]": 0.17595804099988754,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-4096-0.4-True-False-torch.float32-1e-05]": 0.1766997380000248,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-4096-0.4-True-True-torch.float32-1e-05]": 13.65470158200003,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-65536-0.8-False-False-torch.float16-1e-05]": 0.16556013600006736,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-65536-0.8-False-False-torch.float32-1e-05]": 0.1617137810000031,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-65536-0.8-False-True-torch.float16-1e-05]": 0.16847760599995354,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-65536-0.8-False-True-torch.float32-1e-05]": 0.17333123299994213,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-65536-0.8-True-False-torch.float16-1e-05]": 12.901728124999977,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-65536-0.8-True-False-torch.float32-1e-05]": 0.17162440399999923,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-65536-0.8-True-True-torch.float16-1e-05]": 2.235390151000047,
+"tests/ops/test_dropout.py::Test_Dropout::test_op[triton-65536-0.8-True-True-torch.float32-1e-05]": 2.4398699160000206,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-1-119]": 0.3268022320000341,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-1-256]": 0.8823760659999493,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-1-8192]": 1.8248984240000254,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-1-9]": 0.9882581789999563,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-4-119]": 0.16938884200004622,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-4-256]": 0.7870574519999991,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-4-8192]": 1.930936969999948,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-4-9]": 0.6774728689999847,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-8-119]": 0.17216982400003644,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-8-256]": 0.7422299870000302,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-8-8192]": 1.509780644999978,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[cutile-torch.float16-8-9]": 0.7381128459999786,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-1-119]": 1.596578792999992,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-1-256]": 1.5883803480000438,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-1-8192]": 2.315272772999947,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-1-9]": 2.427883546999965,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-4-119]": 1.5957819210000252,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-4-256]": 1.3704667740000218,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-4-8192]": 2.5834790670000416,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-4-9]": 1.5444705960000533,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-8-119]": 1.3251356839999744,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-8-256]": 1.6608055429999808,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-8-8192]": 2.1958568079999736,
+"tests/ops/test_flash_decode.py::Test_FlashDecode::test_op[tilecpp-torch.float16-8-9]": 1.1215785580000102,
+"tests/ops/test_gemma_attention.py::TestGemmaAttention::test_op[cutile-1-12-4-512-64-256-30.0-True-torch.bfloat16]": 0.001750875000027463,
+"tests/ops/test_gemma_attention.py::TestGemmaAttention::test_op[cutile-1-8-8-128-64-0-None-False-torch.float16]": 0.0017509379999864905,
+"tests/ops/test_gemma_attention.py::TestGemmaAttention::test_op[cutile-1-8-8-128-64-64-10.0-False-torch.float32]": 0.0020495439999876908,
+"tests/ops/test_gemma_attention.py::TestGemmaAttention::test_op[cutile-1-8-8-128-64-64-10.0-True-torch.float16]": 0.001894632000016827,
+"tests/ops/test_gemma_attention.py::TestGemmaAttention::test_op[tilecpp-1-12-4-512-64-256-30.0-True-torch.bfloat16]": 0.0017201640000052976,
+"tests/ops/test_gemma_attention.py::TestGemmaAttention::test_op[tilecpp-1-8-8-128-64-0-None-False-torch.float16]": 0.0017645030000039696,
+"tests/ops/test_gemma_attention.py::TestGemmaAttention::test_op[tilecpp-1-8-8-128-64-64-10.0-False-torch.float32]": 0.002007073999948261,
+"tests/ops/test_gemma_attention.py::TestGemmaAttention::test_op[tilecpp-1-8-8-128-64-64-10.0-True-torch.float16]": 0.0017284650000419788,
+"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[cutile-1-8-4-256-64-64-10.0-torch.float16]": 0.0011973750000606742,
+"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[cutile-1-8-8-10019-64-256-50.0-torch.bfloat16]": 0.0012797460000228966,
+"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[cutile-1-8-8-10024-64-0-None-torch.float16]": 0.0010813359999701788,
+"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[cutile-1-8-8-128-64-0-10.0-torch.float16]": 0.0017169120000062321,
+"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[cutile-1-8-8-128-64-0-None-torch.float16]": 0.0016884930000173881,
+"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[cutile-1-8-8-256-64-64-10.0-torch.float16]": 0.0018512220000275192,
+"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[tilecpp-1-8-4-256-64-64-10.0-torch.float16]": 0.00131218199999239,
+"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[tilecpp-1-8-8-10019-64-256-50.0-torch.bfloat16]": 0.0010668770000279437,
+"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[tilecpp-1-8-8-10024-64-0-None-torch.float16]": 0.0011487530000522383,
+"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[tilecpp-1-8-8-128-64-0-10.0-torch.float16]": 0.0012478550000309951,
+"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[tilecpp-1-8-8-128-64-0-None-torch.float16]": 0.001078182000014749,
+"tests/ops/test_gemma_attention_decode.py::TestGemmaAttentionDecode::test_op[tilecpp-1-8-8-256-64-64-10.0-torch.float16]": 0.001128085999994255,
+"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[1024, 512, 256, 128]-[1024, 512, 256, 128]-[1024, 512, 256, 128]-False-torch.float16]": 6.886716718999992,
+"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[1024, 512, 256, 128]-[1024, 512, 256, 128]-[1024, 512, 256, 128]-True-torch.float16]": 7.036091510999938,
+"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[1024, 512, 256, 128]-[1024, 512, 256, 128]-[128, 128, 128, 128]-False-torch.float16]": 0.32105234800002336,
+"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[1024, 512, 256, 128]-[1024, 512, 256, 128]-[128, 128, 128, 128]-True-torch.float16]": 0.3210227239999881,
+"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[1024, 512, 256, 128]-[128, 128, 128, 128]-[1024, 512, 256, 128]-False-torch.float16]": 0.31790624200004913,
+"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[1024, 512, 256, 128]-[128, 128, 128, 128]-[1024, 512, 256, 128]-True-torch.float16]": 0.31702503799999704,
+"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[1024, 512, 256, 128]-[128, 128, 128, 128]-[128, 128, 128, 128]-False-torch.float16]": 6.905363883999996,
+"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[1024, 512, 256, 128]-[128, 128, 128, 128]-[128, 128, 128, 128]-True-torch.float16]": 3.289589630000023,
+"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[256, 256, 256, 256]-[1024, 512, 256, 128]-[1024, 512, 256, 128]-False-torch.float16]": 0.31555707499995833,
+"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[256, 256, 256, 256]-[1024, 512, 256, 128]-[1024, 512, 256, 128]-True-torch.float16]": 0.3181356090000804,
+"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[256, 256, 256, 256]-[1024, 512, 256, 128]-[128, 128, 128, 128]-False-torch.float16]": 0.30788313699997616,
+"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[256, 256, 256, 256]-[1024, 512, 256, 128]-[128, 128, 128, 128]-True-torch.float16]": 0.31579089099994917,
+"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[256, 256, 256, 256]-[128, 128, 128, 128]-[1024, 512, 256, 128]-False-torch.float16]": 0.36039260500001546,
+"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[256, 256, 256, 256]-[128, 128, 128, 128]-[1024, 512, 256, 128]-True-torch.float16]": 0.30845097899998564,
+"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[256, 256, 256, 256]-[128, 128, 128, 128]-[128, 128, 128, 128]-False-torch.float16]": 6.703885912999965,
+"tests/ops/test_group_gemm.py::Test_GroupGemm::test_op[cutile-[256, 256, 256, 256]-[128, 128, 128, 128]-[128, 128, 128, 128]-True-torch.float16]": 4.589147235999974,
+"tests/ops/test_layer_norm.py::Test_LayerNorm2D::test_op[cutile-1004-2456-1.0-torch.float16]": 0.6310031040000013,
+"tests/ops/test_layer_norm.py::Test_LayerNorm2D::test_op[cutile-256-256-0.0-torch.float16]": 0.4049483280000459,
+"tests/ops/test_layer_norm.py::Test_LayerNorm2D::test_op[cutile-256-256-1.0-torch.float16]": 0.4148091419999673,
+"tests/ops/test_layer_norm.py::Test_LayerNorm2D::test_op[cutile-512-256-0.0-torch.float32]": 0.35450512799997114,
+"tests/ops/test_layer_norm.py::Test_LayerNorm2D::test_op[cutile-9-9-0.0-torch.float16]": 0.42291374400002724,
+"tests/ops/test_layer_norm.py::Test_LayerNormNCHW::test_op[cutile-2-512-256-0.0-torch.float16]": 5.289983520999954,
+"tests/ops/test_layer_norm.py::Test_LayerNormNCHW::test_op[cutile-256-256-128-1.0-torch.float16]": 1.5385725169999205,
+"tests/ops/test_layer_norm.py::Test_LayerNormNCHW::test_op[cutile-512-256-128-0.0-torch.float32]": 1.9614299630000573,
+"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[cutile-256-256-0.0-torch.float16]": 0.39376916999998457,
+"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[cutile-256-256-0.0-torch.float32]": 0.39113765399997646,
+"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[cutile-256-256-1.0-torch.float16]": 0.3828092749999996,
+"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[cutile-9-9-0.0-torch.float16]": 0.38703394999998864,
+"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[tilecpp-256-256-0.0-torch.float16]": 0.9550727139999822,
+"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[tilecpp-256-256-0.0-torch.float32]": 0.8847416450000196,
+"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[tilecpp-256-256-1.0-torch.float16]": 0.17916857499994876,
+"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[tilecpp-9-9-0.0-torch.float16]": 0.9988954309999372,
+"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[triton-256-256-0.0-torch.float16]": 15.640779026000018,
+"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[triton-256-256-0.0-torch.float32]": 16.15364535799995,
+"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[triton-256-256-1.0-torch.float16]": 0.17710689000000457,
+"tests/ops/test_layer_norm_legacy.py::Test_LayerNorm::test_op[triton-9-9-0.0-torch.float16]": 19.559062837000056,
+"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[cutile-1024-1024-torch.bfloat16]": 7.5909948699999745,
+"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[cutile-256-256-torch.bfloat16]": 4.723012463999964,
+"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[cutile-30000-1024-torch.bfloat16]": 8.205230397000037,
+"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[cutile-4096-512-torch.bfloat16]": 7.230829905000064,
+"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[tilecpp-1024-1024-torch.bfloat16]": 0.0015939169999796832,
+"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[tilecpp-256-256-torch.bfloat16]": 12.73870397799999,
+"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[tilecpp-30000-1024-torch.bfloat16]": 0.001539141999955973,
+"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[tilecpp-4096-512-torch.bfloat16]": 5.8214610909999465,
+"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[triton-1024-1024-torch.bfloat16]": 22.08223682199997,
+"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[triton-256-256-torch.bfloat16]": 14.430014552999978,
+"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[triton-30000-1024-torch.bfloat16]": 2.353984422000053,
+"tests/ops/test_layer_norm_legacy.py::Test_PersistentLayerNorm::test_op[triton-4096-512-torch.bfloat16]": 20.591679341999964,
+"tests/ops/test_linear_gluact_linear.py::Test_LinearGluactLinear::test_op[cutile-1024-1024-1023-silu-torch.float32]": 0.0015707160000033582,
+"tests/ops/test_linear_gluact_linear.py::Test_LinearGluactLinear::test_op[cutile-1024-1024-1024-relu-torch.float32]": 0.0018345740000427213,
+"tests/ops/test_linear_gluact_linear.py::Test_LinearGluactLinear::test_op[cutile-1024-1024-1024-silu-torch.float32]": 0.0014997750000134147,
+"tests/ops/test_linear_gluact_linear.py::Test_LinearGluactLinear::test_op[cutile-1024-1024-512-gelu_tanh-torch.float32]": 0.0015689759999872877,
+"tests/ops/test_linear_gluact_linear.py::Test_LinearGluactLinear::test_op[cutile-512-512-256-silu-torch.float16]": 0.0015633649999813315,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-T192-bf16]": 16.440764941999987,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-batch2-bf16]": 253.32257192999998,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-chunk128-ungrouped-bf16]": 1145.207325153,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-groups2-bf16]": 268.45968700699996,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-large-bf16]": 422.1874202299998,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-minimal-bf16]": 221.771849911,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-single-chunk-bf16]": 312.383758036,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-standard-bf16]": 638.3585100250002,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkBackward::test_op[cutile-three-chunks-bf16]": 594.4428256350001,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-T192-bf16]": 42.89020906100001,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-batch2-bf16]": 2.778162144000021,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-chunk128-ungrouped-bf16]": 333.343246516,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-groups2-bf16]": 51.94319316399998,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-large-bf16]": 163.647564042,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-minimal-bf16]": 53.80677443799999,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-single-chunk-bf16]": 55.554044868000005,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-standard-bf16]": 109.19098343500013,
+"tests/ops/test_mamba2.py::Test_Mamba2_ChunkForward::test_op[cutile-three-chunks-bf16]": 124.55637726199996,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=False-1024-1024-1023-0-0-False-False-torch.bfloat16]": 0.0014235450000796845,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=False-1024-1024-1024-0-0-False-False-torch.bfloat16]": 1.2060682849999012,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=False-3072-6144-2720-0-0-False-False-torch.bfloat16]": 0.0028908550000323885,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=False-8-8-8-0-0-False-False-torch.bfloat16]": 1.361771955000222,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=False-8192-8192-8192-0-0-False-False-torch.bfloat16]": 0.0012700009999662143,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=True-1024-1024-1023-0-0-False-False-torch.bfloat16]": 0.0014426599998387246,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=True-1024-1024-1024-0-0-False-False-torch.bfloat16]": 12.805215674000237,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=True-3072-6144-2720-0-0-False-False-torch.bfloat16]": 0.0014244520000374905,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=True-8-8-8-0-0-False-False-torch.bfloat16]": 11.903275265000048,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[cutile-use_tma=True-static_persistent=True-8192-8192-8192-0-0-False-False-torch.bfloat16]": 0.0014179180000155611,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=False-1024-1024-1023-0-0-False-False-torch.bfloat16]": 0.0020956509999905393,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=False-1024-1024-1024-0-0-False-False-torch.bfloat16]": 0.1645430559999852,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=False-3072-6144-2720-0-0-False-False-torch.bfloat16]": 0.0021101730000054886,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=False-8-8-8-0-0-False-False-torch.bfloat16]": 0.1649000970000145,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=False-8192-8192-8192-0-0-False-False-torch.bfloat16]": 0.002017899000009038,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=True-1024-1024-1023-0-0-False-False-torch.bfloat16]": 0.0021729110000308083,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=True-1024-1024-1024-0-0-False-False-torch.bfloat16]": 83.49551602200003,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=True-3072-6144-2720-0-0-False-False-torch.bfloat16]": 0.0021057470000016565,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=True-8-8-8-0-0-False-False-torch.bfloat16]": 0.1684778700000038,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=False-static_persistent=True-8192-8192-8192-0-0-False-False-torch.bfloat16]": 0.002041437999992013,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=False-1024-1024-1023-0-0-False-False-torch.bfloat16]": 0.002147935000010648,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=False-1024-1024-1024-0-0-False-False-torch.bfloat16]": 2.6578335300000333,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=False-3072-6144-2720-0-0-False-False-torch.bfloat16]": 0.002159642999998823,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=False-8-8-8-0-0-False-False-torch.bfloat16]": 2.3299466170000187,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=False-8192-8192-8192-0-0-False-False-torch.bfloat16]": 0.002060707999959277,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=True-1024-1024-1023-0-0-False-False-torch.bfloat16]": 0.0030767089999983455,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=True-1024-1024-1024-0-0-False-False-torch.bfloat16]": 75.0932048799998,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=True-3072-6144-2720-0-0-False-False-torch.bfloat16]": 0.0016036669999834885,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=True-8-8-8-0-0-False-False-torch.bfloat16]": 11.111064156999959,
+"tests/ops/test_matmul.py::Test_Matmul::test_op[tilecpp-use_tma=True-static_persistent=True-8192-8192-8192-0-0-False-False-torch.bfloat16]": 0.002155513999980485,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(2112, 7168)-256]": 51.009376728999996,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(2112, 7168)-32]": 17.41440472599993,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(24576, 1536)-256]": 45.79547172799994,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(24576, 1536)-32]": 19.748002109000254,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(24576, 7168)-256]": 50.56251920999989,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(24576, 7168)-32]": 21.520172590000357,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(3072, 1536)-256]": 42.396649485000125,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(3072, 1536)-32]": 17.274422895000043,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(32768, 512)-256]": 40.794714587000044,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(32768, 512)-32]": 18.50863815899993,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(36864, 7168)-256]": 46.39639672700014,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(36864, 7168)-32]": 18.12684397599992,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(4096, 7168)-256]": 48.23905214199999,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(4096, 7168)-32]": 18.408497686000032,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(576, 7168)-256]": 53.89157612399981,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(576, 7168)-32]": 19.437075839000045,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(7168, 16384)-256]": 52.73270881899998,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(7168, 16384)-32]": 24.21143705999998,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(7168, 18432)-256]": 64.55232849499998,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(7168, 18432)-32]": 25.82071939100024,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(7168, 2048)-256]": 48.38804295700015,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(7168, 2048)-32]": 22.28964718499992,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(7168, 256)-256]": 34.102962714000114,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-128-128-(7168, 256)-32]": 11.90084951499989,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(2112, 7168)-256]": 44.79916375100004,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(2112, 7168)-32]": 17.358172249000063,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(24576, 1536)-256]": 30.828758362000144,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(24576, 1536)-32]": 13.377840394999907,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(24576, 7168)-256]": 46.419465708000075,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(24576, 7168)-32]": 18.152508297000054,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(3072, 1536)-256]": 30.373764799999947,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(3072, 1536)-32]": 13.32867958899999,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(32768, 512)-256]": 32.21574944200006,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(32768, 512)-32]": 14.595762071999957,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(36864, 7168)-256]": 48.14672487300004,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(36864, 7168)-32]": 17.601953686000115,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(4096, 7168)-256]": 47.200766808000026,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(4096, 7168)-32]": 17.27700099500001,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(576, 7168)-256]": 44.509754900000075,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(576, 7168)-32]": 16.09275152099997,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(7168, 16384)-256]": 66.59263509400006,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(7168, 16384)-32]": 23.430203581000058,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(7168, 18432)-256]": 72.66806213200005,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(7168, 18432)-32]": 25.761708424000005,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(7168, 2048)-256]": 28.84236316600004,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(7168, 2048)-32]": 11.69138479500009,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(7168, 256)-256]": 25.49763278799992,
+"tests/ops/test_matmul.py::Test_W8A8BlockFp8Matmul::test_op[cutile-torch.float32-32-32-(7168, 256)-32]": 9.444950798000036,
+"tests/ops/test_mla.py::Test_MLA::test_op[cutile-1-64-64-torch.bfloat16-False]": 0.001577435000172045,
+"tests/ops/test_mla.py::Test_MLA::test_op[cutile-1-64-64-torch.bfloat16-True]": 0.002608299999792507,
+"tests/ops/test_mla.py::Test_MLA::test_op[cutile-4-64-64-torch.bfloat16-False]": 0.0016085720001228765,
+"tests/ops/test_mla.py::Test_MLA::test_op[cutile-4-64-64-torch.bfloat16-True]": 0.0015680110000175773,
+"tests/ops/test_mla.py::Test_MLA::test_op[tilecpp-1-64-64-torch.bfloat16-False]": 0.0015698119998432958,
+"tests/ops/test_mla.py::Test_MLA::test_op[tilecpp-1-64-64-torch.bfloat16-True]": 0.0018964999999298016,
+"tests/ops/test_mla.py::Test_MLA::test_op[tilecpp-4-64-64-torch.bfloat16-False]": 0.0015762550001454656,
+"tests/ops/test_mla.py::Test_MLA::test_op[tilecpp-4-64-64-torch.bfloat16-True]": 0.0016327440000623028,
+"tests/ops/test_mla_decoding.py::Test_MLADecoding::test_op[cutile-512-64-torch.float16-128-False]": 0.0017581489998974575,
+"tests/ops/test_mla_decoding.py::Test_MLADecoding::test_op[cutile-512-64-torch.float16-16-True]": 1.4314949190002153,
+"tests/ops/test_mla_decoding.py::Test_MLADecoding::test_op[cutile-512-64-torch.float16-32-True]": 0.17501831799972933,
+"tests/ops/test_mla_decoding.py::Test_MLADecoding::test_op[cutile-512-64-torch.float16-64-False]": 0.002589053000065178,
+"tests/ops/test_mla_decoding.py::Test_MLADecoding::test_op[tilecpp-512-64-torch.float16-128-False]": 0.15583089200003997,
+"tests/ops/test_mla_decoding.py::Test_MLADecoding::test_op[tilecpp-512-64-torch.float16-16-True]": 2.184913616000358,
+"tests/ops/test_mla_decoding.py::Test_MLADecoding::test_op[tilecpp-512-64-torch.float16-32-True]": 0.17055151400018076,
+"tests/ops/test_mla_decoding.py::Test_MLADecoding::test_op[tilecpp-512-64-torch.float16-64-False]": 0.16141638299995975,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-1024-16]": 8.217994514999873,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-1024-24]": 0.32393222299992885,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-1024-32]": 0.16311396200012496,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-1024-8]": 8.347974648000218,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-11049-16]": 4.341782556999988,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-11049-24]": 0.5376043579999532,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-11049-32]": 0.7657193490001646,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-11049-8]": 4.490133292999872,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-129-16]": 6.52517855300016,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-129-24]": 0.38920027700009996,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-129-32]": 0.33821019899983185,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-129-8]": 6.7987506179999855,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-8192-16]": 8.218592922000198,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-8192-24]": 0.5389320190001854,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-8192-32]": 0.5315033529998345,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-128-8192-8]": 5.338411867000104,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-1024-16]": 5.802147434000005,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-1024-24]": 0.29208919600000627,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-1024-32]": 0.3240518750000092,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-1024-8]": 8.536246603000023,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-11049-16]": 7.3375677719999715,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-11049-24]": 0.5151648350000073,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-11049-32]": 0.4682168200000092,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-11049-8]": 6.749728711999978,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-129-16]": 0.28147885299995323,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-129-24]": 0.3802362379999522,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-129-32]": 5.88389042,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-129-8]": 7.2094662870001684,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-8192-16]": 8.726436527999994,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-8192-24]": 0.3945217790000015,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-8192-32]": 0.3957400329999814,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[cutile-torch.float16-512-8192-8]": 8.674215608999987,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-1024-16]": 14.76153707200001,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-1024-24]": 11.998832469999968,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-1024-32]": 18.294106239999962,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-1024-8]": 14.957994362000079,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-11049-16]": 15.965961027999924,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-11049-24]": 17.488446825999972,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-11049-32]": 14.659640495000076,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-11049-8]": 14.287483003999967,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-129-16]": 17.019989525000028,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-129-24]": 17.802136947999998,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-129-32]": 15.99674772100002,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-129-8]": 14.200805408999969,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-8192-16]": 15.012151170000038,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-8192-24]": 14.38310270300002,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-8192-32]": 15.329581627999914,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-128-8192-8]": 14.090039805999993,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-1024-16]": 10.656618018000017,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-1024-24]": 10.819409657000051,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-1024-32]": 13.34323769599996,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-1024-8]": 14.322517325000035,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-11049-16]": 16.100943171999916,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-11049-24]": 14.280895140000098,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-11049-32]": 13.708630698999968,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-11049-8]": 16.58831898699998,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-129-16]": 18.043329024000002,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-129-24]": 13.81065893799996,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-129-32]": 14.466203002999976,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-129-8]": 17.005992801000048,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-8192-16]": 14.494138815000042,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-8192-24]": 15.072196396000095,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-8192-32]": 11.220680682999955,
+"tests/ops/test_mla_decoding_split_kv.py::Test_MLADecodingSplitKV::test_op[tilecpp-torch.float16-512-8192-8]": 15.049231586000133,
+"tests/ops/test_moe.py::Test_MOE::test_op[cutile-torch.bfloat16-16-512-256-8-2]": 4.970538376000036,
+"tests/ops/test_moe.py::Test_MOE::test_op[cutile-torch.bfloat16-32-1024-512-16-4]": 4.6297666340000205,
+"tests/ops/test_moe.py::Test_MOE::test_op[cutile-torch.bfloat16-64-1024-512-20-8]": 4.058250293000015,
+"tests/ops/test_moe.py::Test_MOE::test_op[cutile-torch.float8_e4m3fn-16-512-256-8-2]": 0.003315835999956107,
+"tests/ops/test_moe.py::Test_MOE::test_op[cutile-torch.float8_e4m3fn-32-1024-512-16-4]": 0.001634234999983164,
+"tests/ops/test_moe.py::Test_MOE::test_op[cutile-torch.float8_e4m3fn-64-1024-512-20-8]": 0.0015840419999904043,
+"tests/ops/test_moe.py::Test_MOE::test_op[tilecpp-torch.bfloat16-16-512-256-8-2]": 7.556035729999962,
+"tests/ops/test_moe.py::Test_MOE::test_op[tilecpp-torch.bfloat16-32-1024-512-16-4]": 7.499185562000093,
+"tests/ops/test_moe.py::Test_MOE::test_op[tilecpp-torch.bfloat16-64-1024-512-20-8]": 6.618254191000119,
+"tests/ops/test_moe.py::Test_MOE::test_op[tilecpp-torch.float8_e4m3fn-16-512-256-8-2]": 5.533458117000123,
+"tests/ops/test_moe.py::Test_MOE::test_op[tilecpp-torch.float8_e4m3fn-32-1024-512-16-4]": 3.8561203989999058,
+"tests/ops/test_moe.py::Test_MOE::test_op[tilecpp-torch.float8_e4m3fn-64-1024-512-20-8]": 4.59931268999992,
+"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-2048-4096-1024-64-4-swiglu]": 0.3867171799998914,
+"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-2048-1408-60-4-swiglu]": 5.715540440000041,
+"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-1-swiglu]": 0.3743374169997651,
+"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-2-swiglu]": 0.3832258400000228,
+"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-4-geglu]": 7.616790687999924,
+"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-4-gelu]": 5.189289527000028,
+"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-4-reglu]": 5.911652697000022,
+"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-4-relu]": 5.087130030000026,
+"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-4-relu_sq]": 5.5223060219999525,
+"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-4-silu]": 5.581913783000118,
+"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-4096-4096-1024-64-4-swiglu]": 6.585608053999977,
+"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-512-4096-1024-64-4-swiglu]": 0.41960518299993055,
+"tests/ops/test_moe_actgrad_bwd.py::Test_MoEActgradBwd::test_op[cutile-torch.bfloat16-512-4096-14336-8-2-swiglu]": 7.482981276999908,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-K64_V128-bf16]": 0.16616915499992047,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-K64_V128-fp32]": 0.17419180199988205,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-T16-bf16]": 1.2445961649999617,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-T16-fp32]": 1.1769618820000005,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-T32_final_state-bf16]": 0.9033302540001387,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-T32_final_state-fp32]": 0.8925231249999115,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-T64-bf16]": 1.0452844599998343,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-T64-fp32]": 1.3339788069998804,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-decode-bf16]": 0.6065378780000401,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-decode-fp32]": 0.5473742010001388,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-init_final_state-bf16]": 1.5000870950000262,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-init_final_state-fp32]": 1.4770382140001175,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-l2norm-bf16]": 7.726437966999924,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-l2norm-fp32]": 7.679393432999859,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-minimal-bf16]": 0.5823412119999603,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-minimal-fp32]": 0.5741998670000612,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-qwen35_cached_decode-bf16]": 0.7504276720000007,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[cutile-qwen35_cached_decode-fp32]": 0.7555993800000351,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-K64_V128-bf16]": 0.16114544699985345,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-K64_V128-fp32]": 0.16972328100018785,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-T16-bf16]": 0.15528954800004158,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-T16-fp32]": 0.1501650159999599,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-T32_final_state-bf16]": 358.535216902,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-T32_final_state-fp32]": 377.27709863999996,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-T64-bf16]": 0.1378721760000019,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-T64-fp32]": 0.1518540449999648,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-decode-bf16]": 0.14935207799999262,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-decode-fp32]": 0.15021733499997936,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-init_final_state-bf16]": 2.0115331470001365,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-init_final_state-fp32]": 1.1133887589999176,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-l2norm-bf16]": 184.22908768300022,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-l2norm-fp32]": 192.46732161399973,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-minimal-bf16]": 18.34018505900002,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-minimal-fp32]": 14.529538559999992,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-qwen35_cached_decode-bf16]": 3.185633171000063,
+"tests/ops/test_recurrent_gated_delta_rule.py::Test_RecurrentGatedDeltaRule::test_op[tilecpp-qwen35_cached_decode-fp32]": 3.6948278319998735,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-None-256-18432-torch.float16]": 0.0012539839999590185,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-None-256-256-torch.float16]": 0.5426267289999771,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-None-256-256-torch.float32]": 0.26741104900000323,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-None-256-768-torch.float16]": 0.2914584359999992,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-None-31072-4096-torch.bfloat16]": 1.055037647000006,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-None-4096-256-torch.bfloat16]": 0.4591973370000062,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_cached-256-18432-torch.float16]": 0.7681705930000078,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_cached-256-256-torch.float16]": 0.27794979600000147,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_cached-256-256-torch.float32]": 0.33248802599999294,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_cached-256-768-torch.float16]": 0.30035313300001576,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_cached-31072-4096-torch.bfloat16]": 0.4104248440000333,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_cached-4096-256-torch.bfloat16]": 0.2900041629999919,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_reload-256-18432-torch.float16]": 0.433514544000019,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_reload-256-256-torch.float16]": 0.16458885099999065,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_reload-256-256-torch.float32]": 0.15609652399999163,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_reload-256-768-torch.float16]": 0.13821124899999404,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_reload-31072-4096-torch.bfloat16]": 0.37756405400003246,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-multi_wave_reload-4096-256-torch.bfloat16]": 0.28639800700003093,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-static_persistent-256-18432-torch.float16]": 0.0013404780000030314,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-static_persistent-256-256-torch.float16]": 0.42664972099998977,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-static_persistent-256-256-torch.float32]": 1.6269931680000127,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-static_persistent-256-768-torch.float16]": 0.7231870359999846,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-static_persistent-31072-4096-torch.bfloat16]": 0.16701645199995596,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[cutile-static_persistent-4096-256-torch.bfloat16]": 0.13754532799998742,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-None-256-18432-torch.float16]": 0.0019690919999675316,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-None-256-256-torch.float16]": 0.6654270170000132,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-None-256-256-torch.float32]": 0.8363777010000035,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-None-256-768-torch.float16]": 0.9360571800000912,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-None-31072-4096-torch.bfloat16]": 1.3571823640000389,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-None-4096-256-torch.bfloat16]": 1.0813614639999969,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_cached-256-18432-torch.float16]": 2.284770045000016,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_cached-256-256-torch.float16]": 0.9650470349999978,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_cached-256-256-torch.float32]": 0.9635174719999782,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_cached-256-768-torch.float16]": 0.9956206310000653,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_cached-31072-4096-torch.bfloat16]": 1.149085292000052,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_cached-4096-256-torch.bfloat16]": 0.936074906999977,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_reload-256-18432-torch.float16]": 1.0680440730000669,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_reload-256-256-torch.float16]": 0.13972056499994778,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_reload-256-256-torch.float32]": 0.1917465709999533,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_reload-256-768-torch.float16]": 0.13508918099995526,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_reload-31072-4096-torch.bfloat16]": 1.0121352549999756,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-multi_wave_reload-4096-256-torch.bfloat16]": 0.9298575779999965,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-static_persistent-256-18432-torch.float16]": 0.0010212419999788835,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-static_persistent-256-256-torch.float16]": 0.0011916060000203288,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-static_persistent-256-256-torch.float32]": 0.0009928380000019388,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-static_persistent-256-768-torch.float16]": 0.0010458200000016404,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-static_persistent-31072-4096-torch.bfloat16]": 0.0010105590000080156,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[tilecpp-static_persistent-4096-256-torch.bfloat16]": 0.0010396810000088408,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-None-256-18432-torch.float16]": 0.0017755510000085906,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-None-256-256-torch.float16]": 1.4924478119999947,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-None-256-256-torch.float32]": 0.2766061799999875,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-None-256-768-torch.float16]": 0.29100361800001906,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-None-31072-4096-torch.bfloat16]": 53.954061851999995,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-None-4096-256-torch.bfloat16]": 19.252041902000002,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_cached-256-18432-torch.float16]": 0.0016450749999989966,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_cached-256-256-torch.float16]": 0.0011748269999998229,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_cached-256-256-torch.float32]": 0.0016230759999871225,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_cached-256-768-torch.float16]": 0.001955914999939523,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_cached-31072-4096-torch.bfloat16]": 0.0016323329999750058,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_cached-4096-256-torch.bfloat16]": 0.0015736189999984163,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_reload-256-18432-torch.float16]": 0.40837753800002474,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_reload-256-256-torch.float16]": 0.164428762,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_reload-256-256-torch.float32]": 0.18564478099995085,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_reload-256-768-torch.float16]": 0.1534653009999829,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_reload-31072-4096-torch.bfloat16]": 0.3467538799999943,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-multi_wave_reload-4096-256-torch.bfloat16]": 0.2998777220000193,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-static_persistent-256-18432-torch.float16]": 0.0012055910000299264,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-static_persistent-256-256-torch.float16]": 21.578596568000023,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-static_persistent-256-256-torch.float32]": 30.755522332000055,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-static_persistent-256-768-torch.float16]": 30.127315220999947,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-static_persistent-31072-4096-torch.bfloat16]": 0.21002041200000576,
+"tests/ops/test_rms_norm.py::Test_RMSNorm::test_op[triton-static_persistent-4096-256-torch.bfloat16]": 0.14477061200000207,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-1-128-32-32-128-0.5]": 0.0010494509999716684,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-1-128-32-32-64-1.0]": 0.0023063559999627614,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-1-128-32-8-256-0.25]": 0.0010734190000221133,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-1-128-32-8-64-1.0]": 0.0010913259999938418,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-2-128-32-32-64-1.0]": 0.0011106640000093648,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-2-128-32-8-128-0.5]": 0.0012349040000003697,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-2-128-32-8-256-0.25]": 0.0010093479999682131,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-2-128-32-8-64-1.0]": 0.0010592510000151378,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-3-423-73-155-92-1.0]": 0.0003710830000045462,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.bfloat16-0.01-0.01-3-423-73-213-92-1.0]": 0.00039361400001780567,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-1-128-32-32-128-0.5]": 0.2910575259999746,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-1-128-32-32-64-1.0]": 0.5385629799999947,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-1-128-32-8-256-0.25]": 0.3246821769999997,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-1-128-32-8-64-1.0]": 0.31307871599994996,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-2-128-32-32-64-1.0]": 1.322349805999977,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-2-128-32-8-128-0.5]": 0.34107055900000205,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-2-128-32-8-256-0.25]": 0.3146801570000548,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-2-128-32-8-64-1.0]": 0.3202611229999661,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-3-423-73-155-92-1.0]": 0.00037655399995628613,
+"tests/ops/test_rope.py::Test_RoPE::test_op[cutile-torch.float32-1e_05-1e_05-3-423-73-213-92-1.0]": 0.00042289900000014313,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-1-128-32-32-128-0.5]": 0.0019128319999595078,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-1-128-32-32-64-1.0]": 0.0017747330000190686,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-1-128-32-8-256-0.25]": 0.0016723770000055538,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-1-128-32-8-64-1.0]": 0.002120500999978958,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-2-128-32-32-64-1.0]": 0.0017060760000049413,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-2-128-32-8-128-0.5]": 0.001772126999981083,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-2-128-32-8-256-0.25]": 0.0016899060000241661,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-2-128-32-8-64-1.0]": 0.0016750660000184325,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-3-423-73-155-92-1.0]": 0.0005815309999945839,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.bfloat16-0.01-0.01-3-423-73-213-92-1.0]": 0.0005928270000197244,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-1-128-32-32-128-0.5]": 1.0423356530000092,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-1-128-32-32-64-1.0]": 1.1275050019999924,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-1-128-32-8-256-0.25]": 1.123919944999983,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-1-128-32-8-64-1.0]": 1.141088112999995,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-2-128-32-32-64-1.0]": 1.136114120999963,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-2-128-32-8-128-0.5]": 1.1237543900000162,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-2-128-32-8-256-0.25]": 1.1194037290000267,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-2-128-32-8-64-1.0]": 1.0961069410000164,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-3-423-73-155-92-1.0]": 0.0006522479999944153,
+"tests/ops/test_rope.py::Test_RoPE::test_op[tilecpp-torch.float32-1e_05-1e_05-3-423-73-213-92-1.0]": 0.0007172780000246348,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-1-128-32-32-128-0.5]": 0.0018669570000042768,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-1-128-32-32-64-1.0]": 0.0018784430000096108,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-1-128-32-8-256-0.25]": 0.001720234000003984,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-1-128-32-8-64-1.0]": 0.0017276769999909902,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-2-128-32-32-64-1.0]": 0.0019520299999840063,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-2-128-32-8-128-0.5]": 0.0016946110000048975,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-2-128-32-8-256-0.25]": 0.0017221550000101615,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-2-128-32-8-64-1.0]": 0.0017559550000214585,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-3-423-73-155-92-1.0]": 0.0005843060000074729,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.bfloat16-0.01-0.01-3-423-73-213-92-1.0]": 0.0006050779999782208,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-1-128-32-32-128-0.5]": 0.3550884399999745,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-1-128-32-32-64-1.0]": 0.14403538200002686,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-1-128-32-8-256-0.25]": 2.3099722709999924,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-1-128-32-8-64-1.0]": 0.5701033720000055,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-2-128-32-32-64-1.0]": 0.675888367999903,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-2-128-32-8-128-0.5]": 0.3575526450000268,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-2-128-32-8-256-0.25]": 0.35119318200000293,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-2-128-32-8-64-1.0]": 0.1481856920000837,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-3-423-73-155-92-1.0]": 0.001122172999941995,
+"tests/ops/test_rope.py::Test_RoPE::test_op[triton-torch.float32-1e_05-1e_05-3-423-73-213-92-1.0]": 0.00040035900002521885,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float16-0.02-0.02-1-128-32-32-64-1.0]": 1.2020774869999968,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float16-0.02-0.02-1-128-32-8-256-0.25]": 1.1655035619999978,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float16-0.02-0.02-1-512-16-16-128-1.0]": 1.0882564550000495,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float16-0.02-0.02-2-128-32-8-128-0.5]": 0.447425351999982,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float16-0.02-0.02-2-128-32-8-64-1.0]": 1.2822573629999852,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float32-1e_05-1e_05-1-128-32-32-64-1.0]": 1.3694001020000144,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float32-1e_05-1e_05-1-128-32-8-256-0.25]": 1.318598193000014,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float32-1e_05-1e_05-1-512-16-16-128-1.0]": 1.483233097999971,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float32-1e_05-1e_05-2-128-32-8-128-0.5]": 0.3969996250000065,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[cutile-torch.float32-1e_05-1e_05-2-128-32-8-64-1.0]": 1.2846271250000143,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float16-0.02-0.02-1-128-32-32-64-1.0]": 1.923103655999995,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float16-0.02-0.02-1-128-32-8-256-0.25]": 1.9263355900000079,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float16-0.02-0.02-1-512-16-16-128-1.0]": 2.0188569389999884,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float16-0.02-0.02-2-128-32-8-128-0.5]": 1.6597726099999477,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float16-0.02-0.02-2-128-32-8-64-1.0]": 2.0428846599999986,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float32-1e_05-1e_05-1-128-32-32-64-1.0]": 1.1172961789999647,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float32-1e_05-1e_05-1-128-32-8-256-0.25]": 1.1002200209999842,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float32-1e_05-1e_05-1-512-16-16-128-1.0]": 1.7174889910000104,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float32-1e_05-1e_05-2-128-32-8-128-0.5]": 1.1081660140000054,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[tilecpp-torch.float32-1e_05-1e_05-2-128-32-8-64-1.0]": 1.1662907900000334,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float16-0.02-0.02-1-128-32-32-64-1.0]": 0.9043353030000105,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float16-0.02-0.02-1-128-32-8-256-0.25]": 0.5584236770000075,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float16-0.02-0.02-1-512-16-16-128-1.0]": 0.9370129660000543,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float16-0.02-0.02-2-128-32-8-128-0.5]": 0.542781749000028,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float16-0.02-0.02-2-128-32-8-64-1.0]": 0.7794265720000055,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float32-1e_05-1e_05-1-128-32-32-64-1.0]": 1.125217436000014,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float32-1e_05-1e_05-1-128-32-8-256-0.25]": 0.31300678499994206,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float32-1e_05-1e_05-1-512-16-16-128-1.0]": 0.9440683970000237,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float32-1e_05-1e_05-2-128-32-8-128-0.5]": 0.34372964100001013,
+"tests/ops/test_rope.py::Test_RoPE::test_op_backward[triton-torch.float32-1e_05-1e_05-2-128-32-8-64-1.0]": 0.631120583000012,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-16-1024-1024-torch.float16]": 0.15874963900000694,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-16-1024-2048-torch.float32]": 0.1624370530000192,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-16-1024-512-torch.float32]": 2.200744071000031,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-16-512-1024-torch.float32]": 0.21874926199998868,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-16-512-512-torch.float16]": 0.18308980400001929,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-32-1024-1024-torch.float32]": 0.171862673000021,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-32-1024-4096-torch.float32]": 2.8109475979999843,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-32-1024-512-torch.float16]": 0.20433570900001996,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-32-512-1024-torch.float16]": 0.18946083700001282,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-32-512-2048-torch.float32]": 0.17053569899999843,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-32-512-512-torch.float32]": 0.14896759100000168,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-8-1024-1024-torch.float32]": 2.074385515000017,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-8-1024-2048-torch.float32]": 0.19440464000001612,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-8-1024-512-torch.float16]": 0.17851306700001146,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-8-512-1024-torch.float16]": 2.034167255,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-8-512-2048-torch.float32]": 3.1707001480000088,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[cutile-8-512-512-torch.float16]": 2.108449830000012,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-16-1024-1024-torch.float16]": 0.19510967799999435,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-16-1024-2048-torch.float32]": 0.21283559500000138,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-16-1024-512-torch.float32]": 1.7167564500000196,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-16-512-1024-torch.float32]": 0.18096109999999044,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-16-512-512-torch.float16]": 0.18664921799998524,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-32-1024-1024-torch.float32]": 0.20812421900001254,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-32-1024-4096-torch.float32]": 2.2478429359999836,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-32-1024-512-torch.float16]": 0.20301833700000316,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-32-512-1024-torch.float16]": 0.22159423599998718,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-32-512-2048-torch.float32]": 0.2074765970000101,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-32-512-512-torch.float32]": 0.1975675939999917,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-8-1024-1024-torch.float32]": 1.434142243000025,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-8-1024-2048-torch.float32]": 0.18912431699999388,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-8-1024-512-torch.float16]": 0.22868750599999998,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-8-512-1024-torch.float16]": 1.2722476840000354,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-8-512-2048-torch.float32]": 1.8962144200000068,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op[tilecpp-8-512-512-torch.float16]": 1.7109953469999937,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-13-256-1024-torch.float32]": 0.18087383300002102,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-13-333-1500-torch.float32]": 0.6867329339999912,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-7-100-1000-torch.float32]": 0.5568470499999876,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-7-512-512-torch.float32]": 0.2512775950000048,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-8-100-512-torch.float32]": 0.177501613999965,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-8-256-1000-torch.float32]": 2.4556671179999796,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-8-256-1500-torch.float32]": 3.8411709430000087,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-8-256-3000-torch.float32]": 3.0196159049999665,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[cutile-8-333-1024-torch.float32]": 0.4775725659999921,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-13-256-1024-torch.float32]": 0.152225331999972,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-13-333-1500-torch.float32]": 0.1290054260000204,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-7-100-1000-torch.float32]": 0.14563463999999726,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-7-512-512-torch.float32]": 0.1687126480000245,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-8-100-512-torch.float32]": 0.14956905900001516,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-8-256-1000-torch.float32]": 0.7584219340000402,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-8-256-1500-torch.float32]": 0.9742191040000137,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-8-256-3000-torch.float32]": 0.993313124000025,
+"tests/ops/test_silu_and_mul.py::Test_SiLUAndMul::test_op_irregular_shapes[tilecpp-8-333-1024-torch.float32]": 0.14699114900002996,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-cutile-256-1009-torch.float16]": 1.1401769180000088,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-cutile-256-2048-torch.float16]": 1.0646575400000131,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-cutile-256-2048-torch.float32]": 1.0408266750000053,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-cutile-256-256-torch.float16]": 0.7579328550000071,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-cutile-256-256-torch.float32]": 0.7871794819999991,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-cutile-256-32768-torch.float32]": 2.7974705700000015,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-cutile-256-9-torch.float32]": 0.9352103929999487,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-tilecpp-256-1009-torch.float16]": 1.2596905329999686,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-tilecpp-256-2048-torch.float16]": 1.022793777000004,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-tilecpp-256-2048-torch.float32]": 1.118681754000022,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-tilecpp-256-256-torch.float16]": 1.0827376340000114,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-tilecpp-256-256-torch.float32]": 1.084295764999979,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-tilecpp-256-32768-torch.float32]": 1.4467600540000092,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[baseline-tilecpp-256-9-torch.float32]": 1.070006535999994,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-cutile-256-1009-torch.float16]": 0.6315102949999982,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-cutile-256-2048-torch.float16]": 0.669066966999992,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-cutile-256-2048-torch.float32]": 0.611014950000083,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-cutile-256-256-torch.float16]": 0.47092844299999115,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-cutile-256-256-torch.float32]": 0.5291841350000936,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-cutile-256-32768-torch.float32]": 1.145617675999972,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-cutile-256-9-torch.float32]": 0.5231570930000657,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-tilecpp-256-1009-torch.float16]": 0.3244006599999807,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-tilecpp-256-2048-torch.float16]": 0.3232680379999806,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-tilecpp-256-2048-torch.float32]": 0.32390979899997774,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-tilecpp-256-256-torch.float16]": 0.32123083699997323,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-tilecpp-256-256-torch.float32]": 0.35996640700005855,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-tilecpp-256-32768-torch.float32]": 0.3322249590000297,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_chunked-tilecpp-256-9-torch.float32]": 0.34866172199997436,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-cutile-256-1009-torch.float16]": 0.5091025400000149,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-cutile-256-2048-torch.float16]": 0.38704238399998303,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-cutile-256-2048-torch.float32]": 0.5026739890000158,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-cutile-256-256-torch.float16]": 0.38819984499991733,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-cutile-256-256-torch.float32]": 0.4632740319999584,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-cutile-256-32768-torch.float32]": 1.031928389000086,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-cutile-256-9-torch.float32]": 0.3969155029999456,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-tilecpp-256-1009-torch.float16]": 0.22547129899999163,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-tilecpp-256-2048-torch.float16]": 0.2512758089999352,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-tilecpp-256-2048-torch.float32]": 0.24841813800003365,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-tilecpp-256-256-torch.float16]": 0.23185727799994993,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-tilecpp-256-256-torch.float32]": 0.2716753019999487,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-tilecpp-256-32768-torch.float32]": 0.24373808600000757,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_multi_wave-tilecpp-256-9-torch.float32]": 0.2536740639999948,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-cutile-256-1009-torch.float16]": 1.876546410000003,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-cutile-256-2048-torch.float16]": 1.3630672699999877,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-cutile-256-2048-torch.float32]": 1.9604606329999967,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-cutile-256-256-torch.float16]": 1.9935898699999939,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-cutile-256-256-torch.float32]": 2.542905179999991,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-cutile-256-32768-torch.float32]": 6.538353380000018,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-cutile-256-9-torch.float32]": 2.068101912999964,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-tilecpp-256-1009-torch.float16]": 0.3069133509999915,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-tilecpp-256-2048-torch.float16]": 0.2978788670000654,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-tilecpp-256-2048-torch.float32]": 0.32375537899997653,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-tilecpp-256-256-torch.float16]": 0.302390111999955,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-tilecpp-256-256-torch.float32]": 0.32828348400005325,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-tilecpp-256-32768-torch.float32]": 0.33605909600004225,
+"tests/ops/test_softmax.py::Test_Softmax::test_op[use_tma-tilecpp-256-9-torch.float32]": 0.29816762699994115,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-1-512-16-1]": 0.34584004900005993,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-1-512-16-2]": 0.3662090899999839,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-1-512-32-1]": 0.357477571000004,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-1-512-32-2]": 0.3059208590000253,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-16-512-16-1]": 0.44810361100002183,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-16-512-16-2]": 0.4499209030000202,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-16-512-32-1]": 0.4477385879999929,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-16-512-32-2]": 0.44555508200005534,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-2-512-16-1]": 0.3713049099999921,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-2-512-16-2]": 0.3858243399999992,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-2-512-32-1]": 0.373924704999979,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-2-512-32-2]": 0.3630158600000186,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-8-512-16-1]": 0.3551683070000422,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-8-512-16-2]": 0.32730095800002346,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-8-512-32-1]": 0.40621599600001446,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.bfloat16-8-512-32-2]": 0.3644911339999908,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-1-512-16-1]": 0.31253336000003173,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-1-512-16-2]": 0.31832237100002203,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-1-512-32-1]": 0.1862677549999603,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-1-512-32-2]": 0.2985655690000044,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-16-512-16-1]": 0.42291093599999385,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-16-512-16-2]": 0.46309799499999826,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-16-512-32-1]": 0.4634819640000387,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-16-512-32-2]": 0.4669788030000177,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-2-512-16-1]": 0.3512683970000694,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-2-512-16-2]": 0.2868126390000043,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-2-512-32-1]": 0.3654950199999121,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-2-512-32-2]": 0.3890564440000617,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-8-512-16-1]": 0.36680172199999106,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-8-512-16-2]": 0.37527972400010867,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-8-512-32-1]": 0.3764671310000267,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[cutile-torch.float16-8-512-32-2]": 0.35086121300003015,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-1-512-16-1]": 1.0040956880000067,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-1-512-16-2]": 1.012875570999995,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-1-512-32-1]": 1.0018410330000052,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-1-512-32-2]": 1.0737679060000573,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-16-512-16-1]": 0.9146165719999999,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-16-512-16-2]": 1.0807356870000149,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-16-512-32-1]": 0.7317455599999789,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-16-512-32-2]": 0.951315700000066,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-2-512-16-1]": 1.0551527789999682,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-2-512-16-2]": 0.9531165689999739,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-2-512-32-1]": 1.0489070099999367,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-2-512-32-2]": 1.004165836000027,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-8-512-16-1]": 1.075929798000061,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-8-512-16-2]": 1.0676013450000141,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-8-512-32-1]": 0.834817141999963,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.bfloat16-8-512-32-2]": 1.0679210329999478,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-1-512-16-1]": 1.0106024619999516,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-1-512-16-2]": 0.990609204000009,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-1-512-32-1]": 1.0006418660000236,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-1-512-32-2]": 1.0299623609999458,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-16-512-16-1]": 1.0983279739999716,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-16-512-16-2]": 1.0974622299999055,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-16-512-32-1]": 1.0907782080000175,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-16-512-32-2]": 1.104858375000049,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-2-512-16-1]": 0.1439652249999881,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-2-512-16-2]": 1.0044969389999778,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-2-512-32-1]": 0.1462585160000458,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-2-512-32-2]": 1.0121299650000424,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-8-512-16-1]": 0.17458873700002187,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-8-512-16-2]": 0.9430250890000025,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-8-512-32-1]": 0.14068316999993158,
+"tests/ops/test_splitk_reduce.py::Test_SplitkReduce::test_op[tilecpp-torch.float16-8-512-32-2]": 1.015973353999982,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[cutile-2-1024-512-torch.float32]": 0.45836916299992936,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[cutile-4-512-1024-torch.float32]": 0.5301632639999525,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[cutile-8-256-2048-torch.float32]": 0.5867733279999925,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[cutile-8-512-1024-torch.float16]": 0.4594767569999476,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[cutile-8-512-512-torch.float16]": 0.400229190999994,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[tilecpp-2-1024-512-torch.float32]": 0.8847595779999438,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[tilecpp-4-512-1024-torch.float32]": 0.997665091999977,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[tilecpp-8-256-2048-torch.float32]": 0.8442136799999957,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[tilecpp-8-512-1024-torch.float16]": 1.80865105700002,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward[tilecpp-8-512-512-torch.float16]": 1.834599867999998,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-13-128-1024]": 0.13831050799996092,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-13-333-1500]": 0.8225903419999554,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-4-128-3000]": 0.8607159530000672,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-4-256-1000]": 0.7004901379999637,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-4-256-1500]": 0.8192049840000095,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-7-100-1000]": 0.7447568769999862,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-7-256-512]": 0.1445811070000218,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-8-100-512]": 0.15877384699996355,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[cutile-8-333-256]": 0.4298265900000615,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-13-128-1024]": 0.1739139480000631,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-13-333-1500]": 0.1713802740000574,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-4-128-3000]": 1.176448281999967,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-4-256-1000]": 0.16526088899996694,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-4-256-1500]": 0.16484434500006273,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-7-100-1000]": 0.18676084800006265,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-7-256-512]": 0.21575042199998506,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-8-100-512]": 0.16802273200005402,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_backward_irregular[tilecpp-8-333-256]": 1.7052026570000294,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward[cutile-2-512-1024-4096]": 0.15746927400005006,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward[cutile-4-1024-2048-8192]": 0.1673893910000288,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward[cutile-8-2048-4096-14336]": 0.3210723040000403,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward[tilecpp-2-512-1024-4096]": 0.16095033799996372,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward[tilecpp-4-1024-2048-8192]": 0.17923468899994077,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward[tilecpp-8-2048-4096-14336]": 0.20826760799997146,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-13-256-2048]": 0.16140244299998585,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-13-333-1500]": 0.19669590900008416,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-7-100-1000]": 0.20152156300002844,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-7-512-1024]": 0.16233994000003804,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-8-100-1024]": 0.1650339400000007,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-8-256-1000]": 0.19377215200006503,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-8-256-1500]": 0.1899424270000054,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-8-256-3000]": 0.1940133329999867,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[cutile-8-333-512]": 0.2546142520000103,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-13-256-2048]": 0.1768233629999827,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-13-333-1500]": 0.17423004600004788,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-7-100-1000]": 0.178094834000035,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-7-512-1024]": 0.16941051899999593,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-8-100-1024]": 0.7903090969999198,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-8-256-1000]": 0.14117290799998727,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-8-256-1500]": 0.9752054420000036,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-8-256-3000]": 1.108794794000005,
+"tests/ops/test_swiglu.py::Test_SwiGLU::test_op_forward_irregular[tilecpp-8-333-512]": 0.9739176749999956,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-1024-1024-torch.float16]": 0.30677372500002775,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-1024-1024-torch.float32]": 0.33147616600001584,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-128-128-torch.float16]": 0.3456768629999942,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-128-128-torch.float32]": 0.35810367099998075,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-16-16-torch.float16]": 0.38308831799997733,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-16-16-torch.float32]": 0.4057929460000196,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-2-2-torch.float16]": 0.45025189700004375,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-2-2-torch.float32]": 0.4243210779999913,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-2048-2048-torch.float16]": 0.347525088999987,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-2048-2048-torch.float32]": 0.357055374999959,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-256-256-torch.float16]": 0.37542383899995,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-256-256-torch.float32]": 0.31466435700008333,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-32-32-torch.float16]": 0.33935328100000106,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-32-32-torch.float32]": 0.324109854000028,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-4-4-torch.float16]": 0.3820920989999763,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-4-4-torch.float32]": 0.4092628670000522,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-4096-4096-torch.float16]": 0.4922090539999999,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-4096-4096-torch.float32]": 0.35345445700005484,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-512-512-torch.float16]": 0.3460189440000363,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-512-512-torch.float32]": 0.35523384299995087,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-64-64-torch.float16]": 0.34444478599994,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-64-64-torch.float32]": 0.3898798129999932,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-8-8-torch.float16]": 0.40832737599998836,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-8-8-torch.float32]": 0.41232155500000545,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-8192-8192-torch.float16]": 0.5191771269999776,
+"tests/ops/test_transpose.py::Test_Transpose::test_op[cutile-8192-8192-torch.float32]": 0.34392569599998524,
+"tests/suites/flashinfer/test_flashinfer_attention.py::Test_FlashInfer_DecodePaged::test_op[cutile-4-1024-128-float16]": 9.224961689999986,
+"tests/suites/flashinfer/test_flashinfer_attention.py::Test_FlashInfer_MLADecodePaged::test_op[cutile-32-4-1024-128-float16]": 0.0020152949998646363,
+"tests/suites/flashinfer/test_flashinfer_attention.py::Test_FlashInfer_PrefillPaged::test_op[cutile-4-1024-128-128-float16]": 28.71088855900001,
+"tests/suites/flashinfer/test_flashinfer_attention.py::Test_FlashInfer_PrefillPaged::test_op[cutile-4-1024-192-128-float16]": 34.345878869000046,
+"tests/suites/flashinfer/test_flashinfer_attention.py::Test_FlashInfer_PrefillRagged::test_op[cutile-4-1024-128-float16]": 145.29569402700002,
+"tests/suites/flashinfer/test_flashinfer_attention.py::Test_FlashInfer_PrefillRagged::test_op[cutile-4-1024-192-float16]": 186.50054191499999,
+"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.0-0.0-False-True-torch.bfloat16-torch.bfloat16-4096-4096-4096-cutile]": 68.59352360599996,
+"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.0-0.0-False-True-torch.bfloat16-torch.bfloat16-8192-8192-8192-cutile]": 3.3925157429998762,
+"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.0-0.0-False-True-torch.float16-torch.float16-4096-4096-4096-cutile]": 62.18301584799997,
+"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.0-0.0-False-True-torch.float16-torch.float16-8192-8192-8192-cutile]": 5.239656554999897,
+"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.0-0.0-False-True-torch.float8_e4m3fn-torch.bfloat16-4096-4096-4096-cutile]": 45.23009381500003,
+"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.0-0.0-False-True-torch.float8_e4m3fn-torch.bfloat16-8192-8192-8192-cutile]": 2.5606278060000704,
+"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.5-2.0-False-True-torch.bfloat16-torch.bfloat16-4096-4096-4096-cutile]": 0.018340832000035334,
+"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.5-2.0-False-True-torch.bfloat16-torch.bfloat16-8192-8192-8192-cutile]": 0.030786922000061168,
+"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.5-2.0-False-True-torch.float16-torch.float16-4096-4096-4096-cutile]": 0.017732929999965563,
+"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.5-2.0-False-True-torch.float16-torch.float16-8192-8192-8192-cutile]": 0.03004212800010464,
+"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.5-2.0-False-True-torch.float8_e4m3fn-torch.bfloat16-4096-4096-4096-cutile]": 0.018641426000044703,
+"tests/suites/flashinfer/test_flashinfer_gemm_alpha_beta.py::Test_FlashInfer_Matmul_Alpha_Beta::test_op[1.5-2.0-False-True-torch.float8_e4m3fn-torch.bfloat16-8192-8192-8192-cutile]": 0.02885706500012475,
+"tests/suites/flashinfer/test_flashinfer_masked_bmm.py::Test_FlashInfer_MaskedBMM::test_op[cutile-False-False-torch.float16-6-4096-512-4096-7168]": 68.83983299500005,
+"tests/suites/flashinfer/test_flashinfer_masked_bmm.py::Test_FlashInfer_MaskedBMM::test_op[cutile-False-False-torch.float16-6-4096-512-7168-2048]": 1.3291189089999307,
+"tests/suites/flashinfer/test_flashinfer_masked_bmm.py::Test_FlashInfer_MaskedBMM::test_op[cutile-False-True-torch.float16-6-4096-512-4096-7168]": 62.81985610100014,
+"tests/suites/flashinfer/test_flashinfer_masked_bmm.py::Test_FlashInfer_MaskedBMM::test_op[cutile-False-True-torch.float16-6-4096-512-7168-2048]": 1.7609901000002992,
+"tests/suites/flashinfer/test_flashinfer_masked_bmm.py::Test_FlashInfer_MaskedBMM::test_op[cutile-True-False-torch.float16-6-4096-512-4096-7168]": 65.24516029600022,
+"tests/suites/flashinfer/test_flashinfer_masked_bmm.py::Test_FlashInfer_MaskedBMM::test_op[cutile-True-False-torch.float16-6-4096-512-7168-2048]": 2.224212360000138,
+"tests/suites/flashinfer/test_flashinfer_masked_bmm.py::Test_FlashInfer_MaskedBMM::test_op[cutile-True-True-torch.float16-6-4096-512-4096-7168]": 68.9875606380001,
+"tests/suites/flashinfer/test_flashinfer_masked_bmm.py::Test_FlashInfer_MaskedBMM::test_op[cutile-True-True-torch.float16-6-4096-512-7168-2048]": 1.9805208379998476,
+"tests/suites/flashinfer/test_flashinfer_ragged_block_scaled_bmm.py::Test_FlashInfer_RaggedBlockScaledBMM::test_op[False-True-torch.float8_e4m3fn-torch.bfloat16-2048-2048-128-16-cutile]": 7.569186443000035,
+"tests/suites/flashinfer/test_flashinfer_ragged_block_scaled_bmm.py::Test_FlashInfer_RaggedBlockScaledBMM::test_op[False-True-torch.float8_e4m3fn-torch.bfloat16-2048-2048-128-4-cutile]": 7.253398421999918,
+"tests/suites/flashinfer/test_flashinfer_ragged_block_scaled_bmm.py::Test_FlashInfer_RaggedBlockScaledBMM::test_op[False-True-torch.float8_e4m3fn-torch.bfloat16-2048-2048-512-16-cutile]": 0.7597765000000436,
+"tests/suites/flashinfer/test_flashinfer_ragged_block_scaled_bmm.py::Test_FlashInfer_RaggedBlockScaledBMM::test_op[False-True-torch.float8_e4m3fn-torch.bfloat16-2048-2048-512-4-cutile]": 0.7511204679997263,
+"tests/suites/flashinfer/test_flashinfer_ragged_block_scaled_bmm.py::Test_FlashInfer_RaggedBlockScaledBMM::test_op[False-True-torch.float8_e4m3fn-torch.bfloat16-4096-4096-128-16-cutile]": 0.7790407280001546,
+"tests/suites/flashinfer/test_flashinfer_ragged_block_scaled_bmm.py::Test_FlashInfer_RaggedBlockScaledBMM::test_op[False-True-torch.float8_e4m3fn-torch.bfloat16-4096-4096-128-4-cutile]": 0.7327263170000151,
+"tests/suites/flashinfer/test_flashinfer_ragged_block_scaled_bmm.py::Test_FlashInfer_RaggedBlockScaledBMM::test_op[False-True-torch.float8_e4m3fn-torch.bfloat16-4096-4096-512-16-cutile]": 0.664383356999906,
+"tests/suites/flashinfer/test_flashinfer_ragged_block_scaled_bmm.py::Test_FlashInfer_RaggedBlockScaledBMM::test_op[False-True-torch.float8_e4m3fn-torch.bfloat16-4096-4096-512-4-cutile]": 0.7305581749999419,
+"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_dtypes[torch.bfloat16-4-256-256-256-cutile]": 0.004278494999880422,
+"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_dtypes[torch.float8_e4m3fn-4-256-256-256-cutile]": 23.870699403000117,
+"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_num_groups[1-256-256-256-torch.bfloat16-cutile]": 0.5707847490000404,
+"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_num_groups[4-256-256-256-torch.bfloat16-cutile]": 0.003975593000177469,
+"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_num_groups[8-256-256-256-torch.bfloat16-cutile]": 0.5718257989999529,
+"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_shapes[2-128-128-128-torch.bfloat16-False-False-cutile]": 0.0012983650001388014,
+"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_shapes[2-128-128-128-torch.bfloat16-True-False-cutile]": 0.8874211019997347,
+"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_shapes[4-256-256-256-torch.bfloat16-False-False-cutile]": 0.0010827519997747004,
+"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_shapes[4-256-256-256-torch.bfloat16-True-False-cutile]": 40.080893322000065,
+"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_shapes[4-512-512-512-torch.bfloat16-False-False-cutile]": 0.001783298999953331,
+"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_shapes[4-512-512-512-torch.bfloat16-True-False-cutile]": 0.7858422350000183,
+"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_transpose[False-False-4-256-256-256-torch.bfloat16-cutile]": 0.001758929999823522,
+"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_transpose[False-True-4-256-256-256-torch.bfloat16-cutile]": 0.00167406800005665,
+"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_transpose[True-False-4-256-256-256-torch.bfloat16-cutile]": 0.00579673700008243,
+"tests/suites/flashinfer/test_flashinfer_ragged_bmm.py::Test_FlashInfer_RaggedBMM::test_op_transpose[True-True-4-256-256-256-torch.bfloat16-cutile]": 0.0016638669997064426,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.float8_e4m3fn-128-2048-512]": 0.15136190200018973,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.float8_e4m3fn-128-2048-513]": 0.003699754000081157,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.float8_e4m3fn-16-2048-512]": 0.15264045300000362,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.float8_e4m3fn-16-2048-513]": 0.0037008909998803574,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.float8_e4m3fn-32-2048-512]": 0.13275304500007223,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.float8_e4m3fn-32-2048-513]": 0.003722071000311189,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.float8_e4m3fn-64-2048-512]": 0.15273432100002537,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.float8_e4m3fn-64-2048-513]": 0.003669943999966563,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.int8-128-2048-512]": 0.15631839700017736,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.int8-128-2048-513]": 0.0035463419999359758,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.int8-16-2048-512]": 0.14619444300001305,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.int8-16-2048-513]": 0.0035470390002956265,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.int8-32-2048-512]": 0.14894261299969003,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.int8-32-2048-513]": 0.003507640999941941,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.int8-64-2048-512]": 0.13857701000006273,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-False-torch.int8-64-2048-513]": 0.0039392199998928845,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.float8_e4m3fn-128-2048-512]": 0.033456610000030196,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.float8_e4m3fn-128-2048-513]": 0.003635182999914832,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.float8_e4m3fn-16-2048-512]": 0.039803978999998435,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.float8_e4m3fn-16-2048-513]": 0.003493486000024859,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.float8_e4m3fn-32-2048-512]": 0.03630224700009421,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.float8_e4m3fn-32-2048-513]": 0.0035512510000899056,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.float8_e4m3fn-64-2048-512]": 0.035572941000054925,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.float8_e4m3fn-64-2048-513]": 0.003565712999943571,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.int8-128-2048-512]": 0.16387308999992456,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.int8-128-2048-513]": 0.003551898000011988,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.int8-16-2048-512]": 0.1495977460000404,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.int8-16-2048-513]": 0.0035375860000499415,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.int8-32-2048-512]": 0.14399577799991903,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.int8-32-2048-513]": 0.003546814000003451,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.int8-64-2048-512]": 0.217400342000019,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-False-True-torch.int8-64-2048-513]": 0.003570757000034064,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.float8_e4m3fn-128-2048-512]": 0.0018638110000210872,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.float8_e4m3fn-128-2048-513]": 0.0018020639999463128,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.float8_e4m3fn-16-2048-512]": 0.0018489540000814486,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.float8_e4m3fn-16-2048-513]": 0.0019175470000050154,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.float8_e4m3fn-32-2048-512]": 0.0017828959999519611,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.float8_e4m3fn-32-2048-513]": 0.0018939439999599017,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.float8_e4m3fn-64-2048-512]": 0.0018637719999787805,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.float8_e4m3fn-64-2048-513]": 0.0017910320000282809,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.int8-128-2048-512]": 0.0017999339999619224,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.int8-128-2048-513]": 0.001852184999961537,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.int8-16-2048-512]": 0.001845051000032072,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.int8-16-2048-513]": 0.001791051000054722,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.int8-32-2048-512]": 0.0018296000000077584,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.int8-32-2048-513]": 0.001773427999978594,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.int8-64-2048-512]": 0.0018113309999989724,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-False-torch.int8-64-2048-513]": 0.0018420430000105625,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.float8_e4m3fn-128-2048-512]": 0.14800296300001037,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.float8_e4m3fn-128-2048-513]": 0.003321263000032104,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.float8_e4m3fn-16-2048-512]": 0.12832302499998605,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.float8_e4m3fn-16-2048-513]": 0.0055573310000340825,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.float8_e4m3fn-32-2048-512]": 0.14528636900001857,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.float8_e4m3fn-32-2048-513]": 0.003769400000010137,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.float8_e4m3fn-64-2048-512]": 0.14083780299995396,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.float8_e4m3fn-64-2048-513]": 0.0036917150000022048,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.int8-128-2048-512]": 0.0031415380000225923,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.int8-128-2048-513]": 0.0031689969999320056,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.int8-16-2048-512]": 0.0034770820000176172,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.int8-16-2048-513]": 0.003185577000010653,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.int8-32-2048-512]": 0.0032141149999347363,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.int8-32-2048-513]": 0.0031765330001007896,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.int8-64-2048-512]": 0.0030983279999645674,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-False-True-True-torch.int8-64-2048-513]": 0.003168211000058818,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.float8_e4m3fn-128-2048-512]": 0.0018457740000030753,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.float8_e4m3fn-128-2048-513]": 0.0017866630000753503,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.float8_e4m3fn-16-2048-512]": 0.0018102470000371795,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.float8_e4m3fn-16-2048-513]": 0.0020019210001009924,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.float8_e4m3fn-32-2048-512]": 0.0018051880000484744,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.float8_e4m3fn-32-2048-513]": 0.0018501580000247486,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.float8_e4m3fn-64-2048-512]": 0.0018679919999726735,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.float8_e4m3fn-64-2048-513]": 0.001823993000016344,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.int8-128-2048-512]": 0.00183018899997478,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.int8-128-2048-513]": 0.0017954129999679935,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.int8-16-2048-512]": 0.001843384000039805,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.int8-16-2048-513]": 0.001815439000040442,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.int8-32-2048-512]": 0.0018081590000065262,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.int8-32-2048-513]": 0.0018366730000138887,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.int8-64-2048-512]": 0.0017756879999524244,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-False-torch.int8-64-2048-513]": 0.001844286000050488,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.float8_e4m3fn-128-2048-512]": 0.0021869690000357878,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.float8_e4m3fn-128-2048-513]": 0.0021606710000696694,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.float8_e4m3fn-16-2048-512]": 0.01363270199993849,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.float8_e4m3fn-16-2048-513]": 0.002414777000012691,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.float8_e4m3fn-32-2048-512]": 0.0023265680000008615,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.float8_e4m3fn-32-2048-513]": 0.0021606370000881725,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.float8_e4m3fn-64-2048-512]": 0.0025278000000525935,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.float8_e4m3fn-64-2048-513]": 0.0022689279999212886,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.int8-128-2048-512]": 0.00263363900000968,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.int8-128-2048-513]": 0.0025029799999742863,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.int8-16-2048-512]": 0.0023000070000875894,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.int8-16-2048-513]": 0.0021255300000007082,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.int8-32-2048-512]": 0.0021345239999845944,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.int8-32-2048-513]": 0.002362454000035541,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.int8-64-2048-512]": 0.002121052999996209,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-False-True-torch.int8-64-2048-513]": 0.0021341490000281738,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.float8_e4m3fn-128-2048-512]": 0.0018070429999852422,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.float8_e4m3fn-128-2048-513]": 0.0018586620000746734,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.float8_e4m3fn-16-2048-512]": 0.001229525000042031,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.float8_e4m3fn-16-2048-513]": 0.0012567210000042905,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.float8_e4m3fn-32-2048-512]": 0.0011619939999718554,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.float8_e4m3fn-32-2048-513]": 0.0011364869999397342,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.float8_e4m3fn-64-2048-512]": 0.0019377209999902334,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.float8_e4m3fn-64-2048-513]": 0.0018581800000561088,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.int8-128-2048-512]": 0.001792646999945191,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.int8-128-2048-513]": 0.0018427699999961078,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.int8-16-2048-512]": 0.0018638349999378079,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.int8-16-2048-513]": 0.0018510149999997338,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.int8-32-2048-512]": 0.0018400440000050367,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.int8-32-2048-513]": 0.0018392810000023019,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.int8-64-2048-512]": 0.0018538249999551226,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-False-torch.int8-64-2048-513]": 0.001796950999960245,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.float8_e4m3fn-128-2048-512]": 0.0020754680000436565,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.float8_e4m3fn-128-2048-513]": 0.0020830449999493794,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.float8_e4m3fn-16-2048-512]": 0.0037525479999658273,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.float8_e4m3fn-16-2048-513]": 0.0034185790000265115,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.float8_e4m3fn-32-2048-512]": 0.0033656799999448594,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.float8_e4m3fn-32-2048-513]": 0.0035806879999995544,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.float8_e4m3fn-64-2048-512]": 0.0023976900000093337,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.float8_e4m3fn-64-2048-513]": 0.002057660999980726,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.int8-128-2048-512]": 0.0022848440000302617,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.int8-128-2048-513]": 0.002045398999939607,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.int8-16-2048-512]": 0.0022150980000219533,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.int8-16-2048-513]": 0.002050213000018175,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.int8-32-2048-512]": 0.00203449800000044,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.int8-32-2048-513]": 0.0035369059999652563,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.int8-64-2048-512]": 0.00223681300002454,
+"tests/suites/flashinfer/test_per_token_group_quant_8bit.py::Test_Per_Token_Group_Quant_8bit::test_op[cutile-True-True-True-torch.int8-64-2048-513]": 0.002296849000003931,
+"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e4m3fn-torch.bfloat16-False-10000-4096-64-576-128-128]": 6.310105793000048,
+"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e4m3fn-torch.bfloat16-False-10000-4096-64-576-128-19]": 6.126968955999985,
+"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e4m3fn-torch.bfloat16-False-10000-4096-64-576-128-1]": 9.894507069999975,
+"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e4m3fn-torch.float16-False-10000-4096-64-576-128-128]": 6.662821817999998,
+"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e4m3fn-torch.float16-False-10000-4096-64-576-128-19]": 6.531725861999917,
+"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e4m3fn-torch.float16-False-10000-4096-64-576-128-1]": 10.181068974000027,
+"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e5m2-torch.bfloat16-False-10000-4096-64-576-128-128]": 6.55493346999998,
+"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e5m2-torch.bfloat16-False-10000-4096-64-576-128-19]": 6.841540538999936,
+"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e5m2-torch.bfloat16-False-10000-4096-64-576-128-1]": 9.642607552999948,
+"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e5m2-torch.float16-False-10000-4096-64-576-128-128]": 6.548356052000031,
+"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e5m2-torch.float16-False-10000-4096-64-576-128-19]": 6.237101551000023,
+"tests/suites/flashinfer/test_rope_quantize_fp8.py::Test_FlashInfer_RopeQuantizeFp8::test_op[cutile-torch.float8_e5m2-torch.float16-False-10000-4096-64-576-128-1]": 9.7000522620001,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.0-4-128-torch.bfloat16]": 0.5427869479999572,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.0-4-128-torch.float16]": 0.5832742239999789,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.0-6-300-torch.float32]": 0.5854848159999619,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.0-63-41-torch.float32]": 0.5325411989999793,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.0-8-512-torch.float32]": 4.072010297999896,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.1-4-128-torch.bfloat16]": 0.5026158679999071,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.1-4-128-torch.float16]": 0.565684613000144,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.1-6-300-torch.float32]": 0.651045687000078,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.1-63-41-torch.float32]": 0.5712752489999957,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_backward[cutile-0.1-8-512-torch.float32]": 1.1420998580000514,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-mean-1269-32000-torch.float32]": 4.1678225669999165,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-mean-4-128-torch.bfloat16]": 3.2958957769999415,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-mean-4-128-torch.float16]": 3.946440853000013,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-mean-4-256-torch.float32]": 5.220671379999999,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-mean-6-300-torch.float32]": 2.1785900760000914,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-mean-63-41-torch.float32]": 5.307194140000092,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-mean-8-1024-torch.float32]": 2.6279429329999857,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-sum-1269-32000-torch.float32]": 0.6473794440000802,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-sum-4-128-torch.bfloat16]": 0.44770185500010484,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-sum-4-128-torch.float16]": 0.3652050699998881,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-sum-4-256-torch.float32]": 0.40434822599991094,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-sum-6-300-torch.float32]": 0.3756432919999497,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-sum-63-41-torch.float32]": 0.4391969880000488,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_forward[cutile-sum-8-1024-torch.float32]": 0.7014299829997981,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_ignore_index[cutile-4-256-torch.float32]": 0.5367086829999153,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_ignore_index[cutile-8-512-torch.float32]": 0.17952380999986417,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_predicted_tokens[cutile-4-128-torch.float32]": 0.458277267000085,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_predicted_tokens[cutile-8-256-torch.float32]": 0.49107440799991764,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_reduction_none[cutile-4-128-torch.bfloat16]": 0.14918931900001553,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_reduction_none[cutile-4-128-torch.float16]": 0.16087599400009367,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_reduction_none[cutile-8-512-torch.float32]": 0.6977096670000265,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_softcap[cutile-30.0-4-128-torch.float32]": 4.720004853000091,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_softcap[cutile-30.0-8-512-torch.float32]": 0.7225029430001086,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_softcap[cutile-50.0-4-128-torch.float32]": 0.14645457599999645,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_softcap[cutile-50.0-8-512-torch.float32]": 0.15874205899990557,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_token_accuracy[cutile-4-128-torch.float32]": 0.4106128509999962,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_token_accuracy[cutile-8-256-torch.float32]": 0.40853438400006326,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_weight[cutile-4-128-torch.bfloat16]": 0.4113553799999181,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_weight[cutile-8-512-torch.float32]": 0.7969610970000076,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_z_loss[cutile-0.0001-4-128-torch.float32]": 0.36734328500006086,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_z_loss[cutile-0.0001-8-512-torch.float32]": 0.6820129139998699,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_z_loss[cutile-0.01-4-128-torch.float32]": 0.3776880549997941,
+"tests/suites/liger/test_cross_entropy.py::Test_Liger_CrossEntropy::test_op_z_loss[cutile-0.01-8-512-torch.float32]": 0.6163128910000069,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(1, 1, 1023)-torch.float32]": 0.3952105500001153,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(2, 4, 512)-torch.float32]": 0.3109536649999427,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(2, 8, 4096)-torch.float32]": 0.3820316950000233,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(3, 7, 256)-torch.float32]": 0.1778618749999623,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(4, 16, 2048)-torch.float32]": 0.35241570099992714,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(4, 256)-torch.bfloat16]": 0.2844847030000892,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(4, 256)-torch.float16]": 0.3165936939999483,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(4, 256)-torch.float32]": 0.2976260699999784,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-False-(4, 300)-torch.float32]": 0.38132432400004745,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(1, 1, 1023)-torch.float32]": 0.4174841970000216,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(2, 4, 512)-torch.float32]": 0.32123453099995913,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(2, 8, 4096)-torch.float32]": 0.41145672200002537,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(3, 7, 256)-torch.float32]": 0.18685050900000988,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(4, 16, 2048)-torch.float32]": 0.3262757390001525,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(4, 256)-torch.bfloat16]": 0.2983034950000274,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(4, 256)-torch.float16]": 0.31655468400003883,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(4, 256)-torch.float32]": 0.29091455299999325,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op[cutile-True-(4, 300)-torch.float32]": 0.39859853699999803,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-False-(1, 1, 1023)-torch.float32]": 0.40336928999988686,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-False-(2, 8, 4096)-torch.float32]": 0.3518990190001432,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-False-(3, 7, 256)-torch.float32]": 0.3806481859999167,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-False-(4, 16, 2048)-torch.float32]": 0.4062699830000156,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-False-(4, 256)-torch.bfloat16]": 0.40656577599986576,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-False-(4, 256)-torch.float16]": 0.38621353099995304,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-False-(4, 256)-torch.float32]": 0.390291972,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-False-(8, 512)-torch.float32]": 0.38284947700003613,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-True-(1, 1, 1023)-torch.float32]": 0.5288780170001246,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-True-(2, 8, 4096)-torch.float32]": 0.39870208299987553,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-True-(3, 7, 256)-torch.float32]": 0.3914490309999792,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-True-(4, 16, 2048)-torch.float32]": 0.40855121399999916,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-True-(4, 256)-torch.bfloat16]": 0.39118114500001866,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-True-(4, 256)-torch.float16]": 0.4037389220000023,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-True-(4, 256)-torch.float32]": 0.39610159999995176,
+"tests/suites/liger/test_dyt.py::Test_Liger_DyT::test_op_backward[cutile-True-(8, 512)-torch.float32]": 0.3957006589998855,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-0.0-(2, 128, 512)-torch.float32]": 0.37120980699990014,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-0.0-(4, 256)-torch.bfloat16]": 0.5716984099999536,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-0.0-(4, 256)-torch.float16]": 0.5758084469999858,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-0.0-(4, 256)-torch.float32]": 0.5103983030001018,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-0.0-(8, 512)-torch.float32]": 0.5788448099999641,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-0.0-(9, 7, 41)-torch.float32]": 0.5630617409999559,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-1.0-(2, 128, 512)-torch.float32]": 0.30474669299985635,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-1.0-(4, 256)-torch.bfloat16]": 0.292273788999978,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-1.0-(4, 256)-torch.float16]": 0.3022528630000352,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-1.0-(4, 256)-torch.float32]": 0.35937263599998914,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-1.0-(8, 512)-torch.float32]": 0.34040893900009905,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_backward[cutile-1.0-(9, 7, 41)-torch.float32]": 0.3323152339999069,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_casting_modes[cutile-gemma]": 1.2051929259999952,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_casting_modes[cutile-llama]": 0.1602785990000939,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_casting_modes[cutile-none]": 1.1717748100001018,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(16, 1024)-torch.float32]": 1.6453711210000392,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(2, 128, 512)-torch.float32]": 0.31049846700000217,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(2, 4, 512)-torch.float32]": 0.17858486100010396,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(4, 256)-torch.bfloat16]": 1.33939320900015,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(4, 256)-torch.float16]": 1.5354459840000345,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(4, 256)-torch.float32]": 1.3766439219998574,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(4, 300)-torch.float32]": 2.0565448140001763,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(8, 512)-torch.float32]": 1.458238275999861,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-0.0-(9, 7, 41)-torch.float32]": 1.4013581030000068,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(16, 1024)-torch.float32]": 0.37888714099995013,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(2, 128, 512)-torch.float32]": 0.3017581980001296,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(2, 4, 512)-torch.float32]": 0.1486691280000514,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(4, 256)-torch.bfloat16]": 0.31535688899987235,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(4, 256)-torch.float16]": 0.3312659059999987,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(4, 256)-torch.float32]": 0.31556880799996634,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(4, 300)-torch.float32]": 0.42882165300011366,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(8, 512)-torch.float32]": 0.3174087410000084,
+"tests/suites/liger/test_fused_add_rms_norm.py::Test_Liger_FusedAddRMSNorm::test_op_forward[cutile-1.0-(9, 7, 41)-torch.float32]": 0.3099025099999153,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_backward[cutile-4-64-128-torch.bfloat16]": 0.5509060750000572,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_backward[cutile-4-64-128-torch.float16]": 0.5562174159999813,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_backward[cutile-4-64-128-torch.float32]": 0.5497027410000328,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_backward[cutile-63-41-41-torch.float32]": 0.19016590799992628,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_backward[cutile-8-128-256-torch.float32]": 0.19161776500004635,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_features[cutile]": 1.0894288010000537,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-mean-1024-1024-4096-torch.float32]": 3.7482965149999927,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-mean-4-64-128-torch.bfloat16]": 0.6377513039999485,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-mean-4-64-128-torch.float16]": 0.5621459080001614,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-mean-4-64-128-torch.float32]": 0.436934009999959,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-mean-6-64-150-torch.float32]": 3.526323486000024,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-mean-63-41-41-torch.float32]": 0.15476190800006862,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-mean-8-128-256-torch.float32]": 0.1855326159999322,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-sum-1024-1024-4096-torch.float32]": 0.4773823089999496,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-sum-4-64-128-torch.bfloat16]": 0.45345141099994635,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-sum-4-64-128-torch.float16]": 0.45620586799986995,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-sum-4-64-128-torch.float32]": 0.47484029099996405,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-sum-6-64-150-torch.float32]": 0.4973025010000356,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-sum-63-41-41-torch.float32]": 0.18272134899996217,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_forward[cutile-sum-8-128-256-torch.float32]": 0.18405978199996298,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_with_bias[cutile-4-64-128-torch.float16]": 0.17428206600004614,
+"tests/suites/liger/test_fused_linear_cross_entropy.py::Test_Liger_FusedLinearCrossEntropy::test_op_with_bias[cutile-4-64-128-torch.float32]": 0.17483348600012505,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_backward[cutile-1-64-128-torch.float32]": 0.19076388799999222,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_backward[cutile-188-31-123-torch.float32]": 0.17747659200006183,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_backward[cutile-4-64-128-torch.bfloat16]": 0.17396575599991593,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_backward[cutile-4-64-128-torch.float16]": 0.17924839899978906,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_backward[cutile-4-64-128-torch.float32]": 0.18427982299988344,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_backward[cutile-63-41-41-torch.float32]": 0.1734771420000243,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_backward[cutile-8-128-256-torch.float32]": 0.1745217050000747,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-1-64-128-torch.float32]": 0.18323821299998144,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-188-31-123-torch.float32]": 0.575813384000071,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-3-7-11-torch.float32]": 0.5892864930000314,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-4-64-128-torch.bfloat16]": 0.16832486800001334,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-4-64-128-torch.float16]": 0.18203620900010264,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-4-64-128-torch.float32]": 0.33024181899997984,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-4-64-32768-torch.float32]": 1.1033686759999455,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-63-41-41-torch.float32]": 0.3669135400000414,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.1-8-128-256-torch.float32]": 0.4387354799999912,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-1-64-128-torch.float32]": 0.15819126299993513,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-188-31-123-torch.float32]": 0.5269869340000923,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-3-7-11-torch.float32]": 0.5967225699999972,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-4-64-128-torch.bfloat16]": 0.17404505599995446,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-4-64-128-torch.float16]": 0.181602157000043,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-4-64-128-torch.float32]": 0.4043905949999953,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-4-64-32768-torch.float32]": 0.7869259280000733,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-63-41-41-torch.float32]": 0.2966517039999417,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_forward[cutile-0.5-8-128-256-torch.float32]": 0.4412395250000145,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-0.5-0.2-1024-1024-4096-torch.bfloat16]": 0.9275168209999265,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-0.5-0.2-1024-1024-4096-torch.float32]": 0.19938850599999114,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-0.5-0.2-141-31-123-torch.bfloat16]": 0.5834716260001187,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-0.5-0.2-141-31-123-torch.float32]": 0.1977391489999718,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-1.0-0.5-1024-1024-4096-torch.bfloat16]": 1.2395952260000058,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-1.0-0.5-1024-1024-4096-torch.float32]": 0.1657110209998791,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-1.0-0.5-141-31-123-torch.bfloat16]": 0.5359508199999254,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-1.0-0.5-141-31-123-torch.float32]": 0.16661331900013465,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-2.0-0.8-1024-1024-4096-torch.bfloat16]": 1.2229365270000017,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-2.0-0.8-1024-1024-4096-torch.float32]": 0.19713633499986827,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-2.0-0.8-141-31-123-torch.bfloat16]": 0.5534038180001062,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-42-2.0-0.8-141-31-123-torch.float32]": 0.1899720650000063,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-0.5-0.2-1024-1024-4096-torch.bfloat16]": 1.199740396999914,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-0.5-0.2-1024-1024-4096-torch.float32]": 0.16538340299996435,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-0.5-0.2-141-31-123-torch.bfloat16]": 0.6338645519999773,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-0.5-0.2-141-31-123-torch.float32]": 0.19164638799998102,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-1.0-0.5-1024-1024-4096-torch.bfloat16]": 0.8883177690000821,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-1.0-0.5-1024-1024-4096-torch.float32]": 0.15820466000002398,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-1.0-0.5-141-31-123-torch.bfloat16]": 0.5501847190000717,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-1.0-0.5-141-31-123-torch.float32]": 0.159155037000005,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-2.0-0.8-1024-1024-4096-torch.bfloat16]": 0.8703965610001205,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-2.0-0.8-1024-1024-4096-torch.float32]": 0.1691457489998811,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-2.0-0.8-141-31-123-torch.bfloat16]": 0.5585395560000279,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_liger_shapes[cutile-_100-2.0-0.8-141-31-123-torch.float32]": 0.15672429099993224,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_with_labels[cutile-16-64-128-torch.float32]": 0.15345520699997905,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_with_labels[cutile-4-64-128-torch.float32]": 0.4010212049998927,
+"tests/suites/liger/test_fused_linear_jsd.py::Test_Liger_FusedLinearJSD::test_op_with_labels[cutile-8-128-256-torch.float32]": 0.4586958309998863,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.bfloat16-3-(1, 1, 16, 32)]": 15.664570917999981,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.bfloat16-3-(1, 2, 64, 32)]": 1.1819314930000928,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.bfloat16-3-(2, 4, 32, 64)]": 15.878804452000054,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.bfloat16-7-(1, 1, 16, 32)]": 0.8673312520000991,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.bfloat16-7-(1, 2, 64, 32)]": 0.25823285399997076,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.bfloat16-7-(2, 4, 32, 64)]": 0.9458949129998473,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float16-3-(1, 1, 16, 32)]": 15.04870649399993,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float16-3-(1, 2, 64, 32)]": 1.1954982439998503,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float16-3-(2, 4, 32, 64)]": 15.80990158499992,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float16-7-(1, 1, 16, 32)]": 0.8374241940000502,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float16-7-(1, 2, 64, 32)]": 0.2698904470000798,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float16-7-(2, 4, 32, 64)]": 0.7232847760001277,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float32-3-(1, 1, 16, 32)]": 21.673485145999962,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float32-3-(1, 2, 64, 32)]": 1.193601627000021,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float32-3-(2, 4, 32, 64)]": 31.214243715000066,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float32-7-(1, 1, 16, 32)]": 1.7312733190000245,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float32-7-(1, 2, 64, 32)]": 1.2360603719999972,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op[cutile-torch.float32-7-(2, 4, 32, 64)]": 2.2309870299999375,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_backward[cutile-(1, 1, 16, 32)]": 0.1797801249999793,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_backward[cutile-(1, 2, 64, 32)]": 0.1600291089999928,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_backward[cutile-(2, 4, 32, 64)]": 0.1610723249999637,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_dilation[cutile-1-(1, 1, 16, 32)]": 0.1822450540000773,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_dilation[cutile-1-(1, 2, 64, 32)]": 0.18271216800008006,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_dilation[cutile-1-(2, 4, 32, 64)]": 0.18005888699997286,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_dilation[cutile-2-(1, 1, 16, 32)]": 1.2927995070000406,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_dilation[cutile-2-(1, 2, 64, 32)]": 0.9291647659999853,
+"tests/suites/liger/test_fused_neighborhood_attention.py::Test_Liger_FusedNeighborhoodAttention::test_op_dilation[cutile-2-(2, 4, 32, 64)]": 2.292167694999989,
+"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_backward[cutile-(4, 256)-torch.bfloat16]": 0.3116517969999677,
+"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_backward[cutile-(4, 256)-torch.float16]": 0.29831528900012927,
+"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_backward[cutile-(4, 256)-torch.float32]": 0.3084483460000911,
+"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_backward[cutile-(8, 512)-torch.float32]": 0.3268278120000332,
+"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_backward[cutile-(9, 41, 341)-torch.float32]": 0.4175646050000523,
+"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_backward[cutile-(9, 7, 41)-torch.float32]": 0.29015121800000543,
+"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(16, 1024)-torch.float32]": 0.30437797799993405,
+"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(2, 2, 8)-torch.float32]": 0.29784083499998815,
+"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(2, 4, 512)-torch.float32]": 0.1802428719998943,
+"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(4, 256)-torch.bfloat16]": 0.31760841200014056,
+"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(4, 256)-torch.float16]": 0.3038644089999707,
+"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(4, 256)-torch.float32]": 0.3011692339998717,
+"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(4, 300)-torch.float32]": 0.3354858879999938,
+"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(8, 512)-torch.float32]": 0.33060680699986733,
+"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(9, 41, 341)-torch.float32]": 0.3617368320000196,
+"tests/suites/liger/test_geglu.py::Test_Liger_GEGLU::test_op_forward[cutile-(9, 7, 41)-torch.float32]": 0.3236115959999779,
+"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_backward[cutile-2-4-2-8-torch.bfloat16]": 0.4468094840000276,
+"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_backward[cutile-2-4-2-8-torch.float16]": 0.41930708699999286,
+"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_backward[cutile-2-4-2-8-torch.float32]": 0.4386187479999535,
+"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_backward[cutile-2-63-21-2163-torch.float32]": 1.0481197509999447,
+"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_backward[cutile-4-8-4-16-torch.float32]": 0.42314908399998785,
+"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_forward[cutile-16-32-1-4096-torch.float32]": 0.700078377000068,
+"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_forward[cutile-2-4-2-8-torch.bfloat16]": 0.41707922800014785,
+"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_forward[cutile-2-4-2-8-torch.float16]": 0.37087033800003155,
+"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_forward[cutile-2-4-2-8-torch.float32]": 0.4157012800000075,
+"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_forward[cutile-2-6-3-10-torch.float32]": 0.36646162599981835,
+"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_forward[cutile-2-63-21-2163-torch.float32]": 0.8138064959998701,
+"tests/suites/liger/test_group_norm.py::Test_Liger_GroupNorm::test_op_forward[cutile-4-8-4-16-torch.float32]": 0.379707010000061,
+"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op[cutile-cispo-1-4-128]": 1.2484261649999553,
+"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op[cutile-cispo-2-8-256]": 1.160424474000024,
+"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op[cutile-cispo-4-16-512]": 1.2874565339998298,
+"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op[cutile-grpo-1-4-128]": 1.1423652700000275,
+"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op[cutile-grpo-2-8-256]": 1.2198565229999758,
+"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op[cutile-grpo-4-16-512]": 1.2230601190000243,
+"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op_backward[cutile]": 1.5959779600000275,
+"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op_with_kl[cutile-2-8-256]": 0.48554686400007085,
+"tests/suites/liger/test_grpo_loss.py::Test_Liger_GrpoLoss::test_op_with_mask[cutile-2-8-256]": 1.5340754750000087,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op[cutile-16-1024-0.5-torch.float32]": 0.7598814640000455,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op[cutile-4-256-0.0-torch.float32]": 0.4027549510000199,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op[cutile-4-256-0.5-torch.bfloat16]": 0.46787760199981676,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op[cutile-4-256-0.5-torch.float16]": 0.4249749010000414,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op[cutile-4-256-0.5-torch.float32]": 0.19500575500001105,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op[cutile-4-256-1.0-torch.float32]": 0.4075266689999353,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op[cutile-8-512-0.3-torch.float32]": 0.4823644659999218,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op[cutile-8-512-0.5-torch.float32]": 0.48262270699990495,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_all_ignored[cutile-20-32-torch.bfloat16]": 0.19509843200000887,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_all_ignored[cutile-20-32-torch.float32]": 0.20296779799991782,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_nondefault_ignore_index[cutile-16-1024-0.3-2-torch.bfloat16]": 0.9530986550000762,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_nondefault_ignore_index[cutile-8-512-0.5-2-torch.float32]": 0.4983328700000129,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_nondefault_ignore_index[cutile-8-512-0.5-42-torch.float32]": 0.5306825509999271,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_not_last_layer[cutile-16441-1271-0.5-torch.float32]": 1.029170127000043,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_not_last_layer[cutile-2048-3200-0.5-torch.bfloat16]": 1.382423086000017,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_not_last_layer[cutile-2048-3200-0.5-torch.float32]": 1.0317882470000086,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_not_last_layer[cutile-8-512-0.5-torch.float16]": 0.49892776399985905,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_symmetry[cutile-16441-1271-torch.float32]": 0.19019926900000428,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_symmetry[cutile-2048-3200-torch.bfloat16]": 0.19586262999996507,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_symmetry[cutile-2048-3200-torch.float32]": 0.19496145300013268,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_with_labels[cutile-16-1024-0.5-torch.float32]": 0.883881247000204,
+"tests/suites/liger/test_jsd.py::Test_Liger_JSD::test_op_with_labels[cutile-8-512-0.5-torch.float32]": 0.54006295399995,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-batchmean-(16, 1024)-torch.float32]": 0.18752352700005304,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-batchmean-(4, 256)-torch.bfloat16]": 0.1845544660000087,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-batchmean-(4, 256)-torch.float16]": 0.19184303700001237,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-batchmean-(4, 256)-torch.float32]": 0.18990032299996074,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-batchmean-(4, 300)-torch.float32]": 0.17781266199995116,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-batchmean-(8, 512)-torch.float32]": 0.18435174299997925,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-mean-(16, 1024)-torch.float32]": 0.19061806099989553,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-mean-(4, 256)-torch.bfloat16]": 0.19662213299989162,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-mean-(4, 256)-torch.float16]": 0.1872432799999615,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-mean-(4, 256)-torch.float32]": 0.19835106200014252,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-mean-(4, 300)-torch.float32]": 0.19405557900006443,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-mean-(8, 512)-torch.float32]": 0.19676518099993245,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-none-(16, 1024)-torch.float32]": 0.35575107499994374,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-none-(4, 256)-torch.bfloat16]": 0.3252213269998947,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-none-(4, 256)-torch.float16]": 0.31510828200009655,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-none-(4, 256)-torch.float32]": 0.3229825720000008,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-none-(4, 300)-torch.float32]": 0.41242635300000074,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-none-(8, 512)-torch.float32]": 0.3325248069999134,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-sum-(16, 1024)-torch.float32]": 0.38000193399989257,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-sum-(4, 256)-torch.bfloat16]": 0.34252799000012146,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-sum-(4, 256)-torch.float16]": 0.32621078600004694,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-sum-(4, 256)-torch.float32]": 0.3379667069999641,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-sum-(4, 300)-torch.float32]": 0.4447356049998916,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-False-sum-(8, 512)-torch.float32]": 0.33414937900010955,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-batchmean-(16, 1024)-torch.float32]": 0.18578904200001034,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-batchmean-(4, 256)-torch.bfloat16]": 0.1844564300000684,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-batchmean-(4, 256)-torch.float16]": 0.183164153000007,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-batchmean-(4, 256)-torch.float32]": 0.1865173539999887,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-batchmean-(4, 300)-torch.float32]": 0.17923661599991192,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-batchmean-(8, 512)-torch.float32]": 0.1923770919999015,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-mean-(16, 1024)-torch.float32]": 0.1841740170000321,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-mean-(4, 256)-torch.bfloat16]": 0.18276485699993827,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-mean-(4, 256)-torch.float16]": 0.19096683000009307,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-mean-(4, 256)-torch.float32]": 0.19969272899982116,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-mean-(4, 300)-torch.float32]": 0.19006947000002583,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-mean-(8, 512)-torch.float32]": 0.1901713329998529,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-none-(16, 1024)-torch.float32]": 0.3384280170000693,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-none-(4, 256)-torch.bfloat16]": 0.3148332370000162,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-none-(4, 256)-torch.float16]": 0.31012092200001007,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-none-(4, 256)-torch.float32]": 0.30104148699990674,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-none-(4, 300)-torch.float32]": 0.39805588000001535,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-none-(8, 512)-torch.float32]": 0.3117231630000106,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-sum-(16, 1024)-torch.float32]": 0.3138567560000638,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-sum-(4, 256)-torch.bfloat16]": 0.3273964889999661,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-sum-(4, 256)-torch.float16]": 0.31872251000004326,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-sum-(4, 256)-torch.float32]": 0.32673740500001713,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-sum-(4, 300)-torch.float32]": 0.3811630419999119,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op[cutile-True-sum-(8, 512)-torch.float32]": 0.33534929600000396,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-batchmean-(16, 1024)-torch.float32]": 0.18444160100000317,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-batchmean-(4, 256)-torch.bfloat16]": 0.1801535170001216,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-batchmean-(4, 256)-torch.float16]": 0.19049427700008437,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-batchmean-(4, 256)-torch.float32]": 0.18492954999999256,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-batchmean-(4, 300)-torch.float32]": 0.18686688299999332,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-batchmean-(8, 512)-torch.float32]": 0.1891274849998581,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-mean-(16, 1024)-torch.float32]": 0.1926992580000615,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-mean-(4, 256)-torch.bfloat16]": 0.1902950819999205,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-mean-(4, 256)-torch.float16]": 0.19185948500000904,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-mean-(4, 256)-torch.float32]": 0.19513591999998425,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-mean-(4, 300)-torch.float32]": 0.19137434200001735,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-mean-(8, 512)-torch.float32]": 0.19058327099992312,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-sum-(16, 1024)-torch.float32]": 0.32274827200012624,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-sum-(4, 256)-torch.bfloat16]": 0.30773376199988434,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-sum-(4, 256)-torch.float16]": 0.31023531199991794,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-sum-(4, 256)-torch.float32]": 0.3012760330001356,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-sum-(4, 300)-torch.float32]": 0.3273675280000816,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-False-sum-(8, 512)-torch.float32]": 0.3196663749998834,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-batchmean-(16, 1024)-torch.float32]": 0.19570834500018464,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-batchmean-(4, 256)-torch.bfloat16]": 0.18697758400003295,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-batchmean-(4, 256)-torch.float16]": 0.19141099899991332,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-batchmean-(4, 256)-torch.float32]": 0.18436576299995977,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-batchmean-(4, 300)-torch.float32]": 0.1952662919999284,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-batchmean-(8, 512)-torch.float32]": 0.18951558999992812,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-mean-(16, 1024)-torch.float32]": 0.18115926999996645,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-mean-(4, 256)-torch.bfloat16]": 0.1755609160001086,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-mean-(4, 256)-torch.float16]": 0.17338709000000563,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-mean-(4, 256)-torch.float32]": 0.19229830300014328,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-mean-(4, 300)-torch.float32]": 0.1821928530000605,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-mean-(8, 512)-torch.float32]": 0.18694922500003486,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-sum-(16, 1024)-torch.float32]": 0.3303857819998939,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-sum-(4, 256)-torch.bfloat16]": 0.3190864530000681,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-sum-(4, 256)-torch.float16]": 0.3183966229998987,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-sum-(4, 256)-torch.float32]": 0.30141784000011285,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-sum-(4, 300)-torch.float32]": 0.36355716699995355,
+"tests/suites/liger/test_kl_div.py::Test_Liger_KLDiv::test_op_backward[cutile-True-sum-(8, 512)-torch.float32]": 0.2893917290000445,
+"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op[cutile-(16, 1024)-torch.float32]": 0.44101088899992646,
+"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op[cutile-(2, 4, 512)-torch.float32]": 0.19196954699998514,
+"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op[cutile-(4, 256)-torch.bfloat16]": 0.3411783150000929,
+"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op[cutile-(4, 256)-torch.float16]": 0.34855548899997757,
+"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op[cutile-(4, 256)-torch.float32]": 0.3827630699998963,
+"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op[cutile-(4, 300)-torch.float32]": 0.4078530110000429,
+"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op[cutile-(8, 512)-torch.float32]": 0.40225511400001324,
+"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op_backward[cutile-(16, 1024)-torch.float32]": 0.4159130369999957,
+"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op_backward[cutile-(2, 4, 512)-torch.float32]": 0.1970986600000515,
+"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op_backward[cutile-(4, 256)-torch.bfloat16]": 0.4359263160000637,
+"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op_backward[cutile-(4, 256)-torch.float16]": 0.39889114300012807,
+"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op_backward[cutile-(4, 256)-torch.float32]": 0.4392361619999292,
+"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op_backward[cutile-(4, 300)-torch.float32]": 0.4050831949999747,
+"tests/suites/liger/test_layer_norm.py::Test_Liger_LayerNorm::test_op_backward[cutile-(8, 512)-torch.float32]": 0.4145495429999073,
+"tests/suites/liger/test_llama4_rope.py::Test_Liger_Llama4Rope::test_op[cutile-(1, 16, 8, 128)-torch.float32]": 0.3827910460001931,
+"tests/suites/liger/test_llama4_rope.py::Test_Liger_Llama4Rope::test_op[cutile-(1, 4, 2, 48)-torch.float32]": 0.38134586000012405,
+"tests/suites/liger/test_llama4_rope.py::Test_Liger_Llama4Rope::test_op[cutile-(2, 8, 4, 64)-torch.bfloat16]": 0.37255919599999743,
+"tests/suites/liger/test_llama4_rope.py::Test_Liger_Llama4Rope::test_op[cutile-(2, 8, 4, 64)-torch.float16]": 0.3839855090000128,
+"tests/suites/liger/test_llama4_rope.py::Test_Liger_Llama4Rope::test_op[cutile-(2, 8, 4, 64)-torch.float32]": 0.3858536490000688,
+"tests/suites/liger/test_llama4_rope.py::Test_Liger_Llama4Rope::test_op_backward[cutile-(2, 8, 4, 64)-torch.bfloat16]": 0.3699848499999234,
+"tests/suites/liger/test_llama4_rope.py::Test_Liger_Llama4Rope::test_op_backward[cutile-(2, 8, 4, 64)-torch.float16]": 0.38336822200017195,
+"tests/suites/liger/test_llama4_rope.py::Test_Liger_Llama4Rope::test_op_backward[cutile-(2, 8, 4, 64)-torch.float32]": 0.3518563660001064,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.bfloat16-(64, 512)-per_token]": 0.36507842399998935,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.bfloat16-(64, 512)-shared]": 0.3360640150000336,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.bfloat16-(8, 128)-per_token]": 0.34639637299994774,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.bfloat16-(8, 128)-shared]": 0.3623474349999469,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.float16-(64, 512)-per_token]": 0.35572408500001984,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.float16-(64, 512)-shared]": 0.37158303400019577,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.float16-(8, 128)-per_token]": 0.3577198660000249,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.float16-(8, 128)-shared]": 0.3724312890001329,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.float32-(64, 512)-per_token]": 0.3525167519999286,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.float32-(64, 512)-shared]": 0.3573840830000563,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.float32-(8, 128)-per_token]": 0.3925095609999971,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op[cutile-torch.float32-(8, 128)-shared]": 0.36143752999998924,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op_backward[cutile-torch.float32-(64, 512)-per_token]": 0.42236594300004526,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op_backward[cutile-torch.float32-(64, 512)-shared]": 0.42434999399995377,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op_backward[cutile-torch.float32-(8, 128)-per_token]": 0.39040892899993196,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op_backward[cutile-torch.float32-(8, 128)-shared]": 0.3601701170000524,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op_casting_modes[cutile-gemma]": 0.36574850900012734,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op_casting_modes[cutile-llama]": 0.19583109200004856,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op_casting_modes[cutile-none]": 0.3523849099999552,
+"tests/suites/liger/test_modulated_rms_norm.py::Test_Liger_ModulatedRMSNorm::test_op_no_shift[cutile]": 0.5865159180001456,
+"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op[cutile-(1, 2, 16)-torch.float32]": 0.4613247690000435,
+"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op[cutile-(2, 1, 8)-torch.bfloat16]": 0.7701759490000768,
+"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op[cutile-(2, 1, 8)-torch.float16]": 0.19442416499998671,
+"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op[cutile-(2, 1, 8)-torch.float32]": 0.5400408689999949,
+"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_backward[cutile-(1, 1, 16)-torch.float32]": 0.46690168899988294,
+"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_backward[cutile-(2, 1, 8)-torch.float32]": 0.46484240000006594,
+"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_backward_conv_params[cutile-2-1-8-1-torch.float32]": 0.21149914700004047,
+"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_backward_conv_params[cutile-2-2-8-1-torch.float32]": 0.20287371300003088,
+"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_backward_conv_params[cutile-2-2-8-2-torch.bfloat16]": 0.4663739350000924,
+"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_backward_conv_params[cutile-2-2-8-2-torch.float32]": 0.20027638800002023,
+"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_with_bias[cutile-(2, 1, 8)-torch.bfloat16]": 0.19544439799994962,
+"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_with_bias[cutile-(2, 1, 8)-torch.float16]": 0.18526842100004615,
+"tests/suites/liger/test_multi_token_attention.py::Test_Liger_MultiTokenAttention::test_op_with_bias[cutile-(2, 1, 8)-torch.float32]": 0.19706477100010034,
+"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op[cutile-(16, 1024)-torch.float32]": 0.6903593929998806,
+"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op[cutile-(2, 4, 512)-torch.float32]": 0.20493697800009159,
+"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op[cutile-(4, 256)-torch.bfloat16]": 0.6601027150001073,
+"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op[cutile-(4, 256)-torch.float16]": 0.6249749750001001,
+"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op[cutile-(4, 256)-torch.float32]": 0.659142007000014,
+"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op[cutile-(4, 300)-torch.float32]": 0.7487954749999517,
+"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op[cutile-(8, 512)-torch.float32]": 0.5826281419999759,
+"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op_backward[cutile-(4, 256)-torch.bfloat16]": 0.4593909469999744,
+"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op_backward[cutile-(4, 256)-torch.float16]": 0.518339830000059,
+"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op_backward[cutile-(4, 256)-torch.float32]": 0.49875404500005516,
+"tests/suites/liger/test_poly_norm.py::Test_Liger_PolyNorm::test_op_backward[cutile-(8, 512)-torch.float32]": 0.45827190900001824,
+"tests/suites/liger/test_qwen2vl_mrope.py::Test_Liger_Qwen2VLMRope::test_op[cutile-(1, 2, 4, 48)-torch.float32]": 0.44062144300005457,
+"tests/suites/liger/test_qwen2vl_mrope.py::Test_Liger_Qwen2VLMRope::test_op[cutile-(1, 4, 8, 128)-torch.float32]": 0.502649073999919,
+"tests/suites/liger/test_qwen2vl_mrope.py::Test_Liger_Qwen2VLMRope::test_op[cutile-(2, 8, 16, 64)-torch.bfloat16]": 0.5290817620000325,
+"tests/suites/liger/test_qwen2vl_mrope.py::Test_Liger_Qwen2VLMRope::test_op[cutile-(2, 8, 16, 64)-torch.float16]": 0.5156687920000422,
+"tests/suites/liger/test_qwen2vl_mrope.py::Test_Liger_Qwen2VLMRope::test_op[cutile-(2, 8, 16, 64)-torch.float32]": 0.5064099930000339,
+"tests/suites/liger/test_qwen2vl_mrope.py::Test_Liger_Qwen2VLMRope::test_op_backward[cutile-(2, 8, 16, 64)-torch.bfloat16]": 0.7972793289999345,
+"tests/suites/liger/test_qwen2vl_mrope.py::Test_Liger_Qwen2VLMRope::test_op_backward[cutile-(2, 8, 16, 64)-torch.float16]": 0.7605072439998821,
+"tests/suites/liger/test_qwen2vl_mrope.py::Test_Liger_Qwen2VLMRope::test_op_backward[cutile-(2, 8, 16, 64)-torch.float32]": 0.7755541979998952,
+"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(1, 1, 1023)-torch.float32]": 0.3608286069999167,
+"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(2, 4, 512)-torch.float32]": 0.28442271799997343,
+"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(2, 8, 4096)-torch.float32]": 0.3275831710000148,
+"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(3, 7, 256)-torch.float32]": 0.20484393899994302,
+"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(4, 16, 2048)-torch.float32]": 0.32333736500004306,
+"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(4, 256)-torch.bfloat16]": 0.30858040900000105,
+"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(4, 256)-torch.float16]": 0.3086490139997977,
+"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(4, 256)-torch.float32]": 0.3096484289998216,
+"tests/suites/liger/test_relu_squared.py::Test_Liger_ReLUSquared::test_op[cutile-(4, 300)-torch.float32]": 0.34494879399994716,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op[cutile-(16, 1024)-torch.float32]": 0.3917609730000322,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op[cutile-(2, 4, 512)-torch.float32]": 0.19636868099996718,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op[cutile-(4, 256)-torch.bfloat16]": 0.35750292099987746,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op[cutile-(4, 256)-torch.float16]": 0.358765725000012,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op[cutile-(4, 256)-torch.float32]": 0.357877042000041,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op[cutile-(4, 300)-torch.float32]": 0.3400099359998876,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op[cutile-(8, 512)-torch.float32]": 0.3585236359999726,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_backward[cutile-(4, 256)-torch.bfloat16]": 0.3790616669998599,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_backward[cutile-(4, 256)-torch.float16]": 0.38609160300006806,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_backward[cutile-(4, 256)-torch.float32]": 0.37537229400004435,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_backward[cutile-(8, 512)-torch.float32]": 0.3926283750000721,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_casting_modes[cutile-(4, 256)-torch.bfloat16-gemma]": 0.33102795899992543,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_casting_modes[cutile-(4, 256)-torch.bfloat16-llama]": 0.19961401299997306,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_casting_modes[cutile-(4, 256)-torch.bfloat16-none]": 0.3619740889998866,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_casting_modes[cutile-(4, 256)-torch.float32-gemma]": 0.33643768599995383,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_casting_modes[cutile-(4, 256)-torch.float32-llama]": 0.20490970000003017,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_casting_modes[cutile-(4, 256)-torch.float32-none]": 0.3287884159999521,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_no_affine[cutile-(4, 256)-torch.bfloat16]": 0.35515634200010027,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_no_affine[cutile-(4, 256)-torch.float16]": 0.34153273100002934,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_no_affine[cutile-(4, 256)-torch.float32]": 0.34579088699990734,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_no_affine[cutile-(8, 512)-torch.float32]": 0.3361719179999909,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_with_offset[cutile-0.0-(4, 256)-torch.bfloat16]": 0.17860751600005642,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_with_offset[cutile-0.0-(4, 256)-torch.float16]": 0.18431906800014985,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_with_offset[cutile-0.0-(4, 256)-torch.float32]": 0.19346212399989327,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_with_offset[cutile-1.0-(4, 256)-torch.bfloat16]": 0.1861001549999628,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_with_offset[cutile-1.0-(4, 256)-torch.float16]": 0.1908787229999689,
+"tests/suites/liger/test_rms_norm.py::Test_Liger_RMSNorm::test_op_with_offset[cutile-1.0-(4, 256)-torch.float32]": 0.18643316800000775,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-False-(1, 2, 4, 48)-torch.float32]": 0.20277592199988703,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-False-(1, 4, 8, 128)-torch.float32]": 0.2053445340000053,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-False-(2, 8, 16, 64)-torch.bfloat16]": 0.461758684000074,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-False-(2, 8, 16, 64)-torch.float16]": 0.4773911209999824,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-False-(2, 8, 16, 64)-torch.float32]": 0.44657733500002905,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-False-(3, 4, 7, 64)-torch.float32]": 0.4272855489999756,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-True-(1, 2, 4, 48)-torch.float32]": 0.4451259410000148,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-True-(1, 4, 8, 128)-torch.float32]": 0.47467739400008213,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-True-(2, 8, 16, 64)-torch.bfloat16]": 0.4628692349998573,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-True-(2, 8, 16, 64)-torch.float16]": 0.4618117659999825,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-True-(2, 8, 16, 64)-torch.float32]": 0.4465178110000352,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op[cutile-True-(3, 4, 7, 64)-torch.float32]": 0.4208680369999911,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op_backward[cutile-(1, 2, 4, 48)-torch.float32]": 0.7318740789999083,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op_backward[cutile-(1, 4, 8, 128)-torch.float32]": 0.6634033999998792,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op_backward[cutile-(2, 8, 16, 64)-torch.bfloat16]": 0.7056307689999812,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op_backward[cutile-(2, 8, 16, 64)-torch.float16]": 0.7075262600000087,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op_backward[cutile-(2, 8, 16, 64)-torch.float32]": 0.6457688680000047,
+"tests/suites/liger/test_rope.py::Test_Liger_Rope::test_op_backward[cutile-(3, 4, 7, 64)-torch.float32]": 0.7113963799999965,
+"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op[cutile-(1, 1023)-torch.float32]": 0.7732838279998759,
+"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op[cutile-(1, 2, 4096)-torch.float32]": 0.20220886800007065,
+"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op[cutile-(1, 4096)-torch.float32]": 0.85520792300008,
+"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op[cutile-(2, 8)-torch.bfloat16]": 0.322386490000099,
+"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op[cutile-(2, 8)-torch.float16]": 0.5799584630000254,
+"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op[cutile-(2, 8)-torch.float32]": 0.612193599999955,
+"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op[cutile-(3, 7, 256)-torch.float32]": 0.6672786809999707,
+"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op[cutile-(4, 16)-torch.float32]": 0.5934020900000405,
+"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op_backward[cutile-(2, 8)-torch.bfloat16]": 0.31893527800014,
+"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op_backward[cutile-(2, 8)-torch.float16]": 0.29742422300000726,
+"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op_backward[cutile-(2, 8)-torch.float32]": 1.4608582899999192,
+"tests/suites/liger/test_softmax.py::Test_Liger_Softmax::test_op_backward[cutile-(4, 16)-torch.float32]": 1.5669187220000822,
+"tests/suites/liger/test_sparsemax.py::Test_Liger_Sparsemax::test_op[cutile-1-(2, 128, 512)-torch.float32]": 1.2475533479999967,
+"tests/suites/liger/test_sparsemax.py::Test_Liger_Sparsemax::test_op[cutile-1-(4, 256)-torch.float32]": 0.16637465000007978,
+"tests/suites/liger/test_sparsemax.py::Test_Liger_Sparsemax::test_op[cutile-1-(5, 123, 123)-torch.float32]": 0.1791932999999517,
+"tests/suites/liger/test_sparsemax.py::Test_Liger_Sparsemax::test_op[cutile-_1-(2, 128, 512)-torch.float32]": 1.643453905000115,
+"tests/suites/liger/test_sparsemax.py::Test_Liger_Sparsemax::test_op[cutile-_1-(4, 256)-torch.float32]": 1.3706467900000234,
+"tests/suites/liger/test_sparsemax.py::Test_Liger_Sparsemax::test_op[cutile-_1-(5, 123, 123)-torch.float32]": 1.3146719129999838,
+"tests/suites/liger/test_sparsemax.py::Test_Liger_Sparsemax::test_op_backward[cutile-(2, 128, 512)-torch.float32]": 1.25955106899994,
+"tests/suites/liger/test_sparsemax.py::Test_Liger_Sparsemax::test_op_backward[cutile-(4, 256)-torch.float32]": 1.1488202309998314,
+"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_backward[cutile-(2, 256, 256)-torch.float32]": 0.29973330500001794,
+"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_backward[cutile-(2, 8, 8)-torch.float32]": 0.3419750789998943,
+"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_backward[cutile-(4, 256)-torch.bfloat16]": 0.3067458970001553,
+"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_backward[cutile-(4, 256)-torch.float16]": 0.3232307730000912,
+"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_backward[cutile-(4, 256)-torch.float32]": 0.3327670290000242,
+"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_backward[cutile-(6, 42, 123)-torch.float32]": 0.33662258100002873,
+"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_backward[cutile-(9, 7, 41)-torch.float32]": 0.3532426360000045,
+"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_forward[cutile-(2, 256, 256)-torch.float32]": 0.3164182809999829,
+"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_forward[cutile-(2, 8, 8)-torch.float32]": 0.29211850799993044,
+"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_forward[cutile-(4, 256)-torch.bfloat16]": 0.29498136400013664,
+"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_forward[cutile-(4, 256)-torch.float16]": 0.31558937800002695,
+"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_forward[cutile-(4, 256)-torch.float32]": 0.32530632999987574,
+"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_forward[cutile-(4, 300)-torch.float32]": 0.4029209869997885,
+"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_forward[cutile-(6, 42, 123)-torch.float32]": 0.3329654380000875,
+"tests/suites/liger/test_swiglu.py::Test_Liger_SwiGLU::test_op_forward[cutile-(9, 7, 41)-torch.float32]": 0.2911379610001177,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-2-1-1024-128]": 0.17493426999999429,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-2-2-512-512]": 0.1649926279999363,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-2-4-127-128]": 0.1670583830000396,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-4-1-1024-128]": 0.16733543500015458,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-4-2-512-512]": 0.16978048499993292,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-4-4-127-128]": 0.17531779600017217,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-None-1-1024-128]": 0.1825178940000569,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-None-2-512-512]": 0.16670351800007666,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_backward[cutile-None-4-127-128]": 0.16666688300006172,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-2-1-1024-128]": 0.17363360799993188,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-2-1-1024-256]": 0.16934920500011685,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-2-2-1024-64]": 0.17099841300000662,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-2-2-512-512]": 0.15678927199996906,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-2-4-127-128]": 0.161265745000037,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-4-1-1024-128]": 0.16487629300013396,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-4-1-1024-256]": 0.17402220199994645,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-4-2-1024-64]": 0.16388415900007658,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-4-2-512-512]": 0.17239195700005894,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-4-4-127-128]": 0.1648518200000808,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-None-1-1024-128]": 0.16425074099993253,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-None-1-1024-256]": 0.1653650909998987,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-None-2-1024-64]": 0.17004540699997506,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-None-2-512-512]": 0.162907339999947,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-False-None-4-127-128]": 0.1626947280000195,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-2-1-1024-128]": 0.17611794499998723,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-2-1-1024-256]": 0.1648509260000992,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-2-2-1024-64]": 0.16526223499988646,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-2-2-512-512]": 0.16172714000003907,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-2-4-127-128]": 0.16586896900003012,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-4-1-1024-128]": 0.16559911000001648,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-4-1-1024-256]": 0.16011759499997424,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-4-2-1024-64]": 0.17604952200008483,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-4-2-512-512]": 0.1640162880000844,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-4-4-127-128]": 0.16660751099982463,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-None-1-1024-128]": 0.17451587100003962,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-None-1-1024-256]": 0.1726395360000197,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-None-2-1024-64]": 0.1713366120000046,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-None-2-512-512]": 0.16480561499997748,
+"tests/suites/liger/test_tiled_mlp.py::Test_Liger_TiledMLP::test_op_forward[cutile-True-None-4-127-128]": 0.1702519170000869,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-batchmean-1-4096-32000]": 4.712295294000114,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-batchmean-3-423-32000]": 3.1280336779999516,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-batchmean-41-401-1271]": 0.7144255919999978,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-mean-1-4096-32000]": 4.722373873000038,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-mean-3-423-32000]": 4.787947915999894,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-mean-41-401-1271]": 0.7179839340000171,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-none-1-4096-32000]": 1.3721157300000186,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-none-3-423-32000]": 1.31884279399992,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-none-41-401-1271]": 0.6439159939999399,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-sum-1-4096-32000]": 3.041589644999931,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-sum-3-423-32000]": 4.678627348000077,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.bfloat16-0.001-0.001-sum-41-401-1271]": 0.7062969219999786,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-batchmean-1-4096-32000]": 5.574871802000075,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-batchmean-3-423-32000]": 6.219412809000005,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-batchmean-41-401-1271]": 0.4258487160000186,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-mean-1-4096-32000]": 5.9956709149998915,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-mean-3-423-32000]": 5.991142064999963,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-mean-41-401-1271]": 0.5802070870001899,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-none-1-4096-32000]": 0.7583121249999749,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-none-3-423-32000]": 0.6361222399999633,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-none-41-401-1271]": 0.581173576999845,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-sum-1-4096-32000]": 5.857412423000142,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-sum-3-423-32000]": 3.8159431700000823,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op[cutile-torch.float32-1e_05-1e_05-sum-41-401-1271]": 0.5399985279999555,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-batchmean-1-4096-32000]": 0.23580633699998543,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-batchmean-3-423-32000]": 0.25839498600009847,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-batchmean-41-401-1271]": 0.23268810699994447,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-mean-1-4096-32000]": 0.23692259300003116,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-mean-3-423-32000]": 0.22104155500005618,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-mean-41-401-1271]": 0.23187645900009102,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-sum-1-4096-32000]": 0.23428132600008666,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-sum-3-423-32000]": 0.2336667319997332,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_backward[cutile-sum-41-401-1271]": 0.2372952330000544,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-batchmean-1-4096-32000]": 3.152960420000113,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-batchmean-3-423-32000]": 3.1506786190000184,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-batchmean-41-401-1271]": 0.5805732140000828,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-mean-1-4096-32000]": 3.1505796530000225,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-mean-3-423-32000]": 3.0856750530000454,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-mean-41-401-1271]": 0.5362900520000267,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-none-1-4096-32000]": 1.0944998669999677,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-none-3-423-32000]": 0.986068562000014,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-none-41-401-1271]": 0.6006757459998653,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-sum-1-4096-32000]": 3.054538505999858,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-sum-3-423-32000]": 3.1433088649999945,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-0-sum-41-401-1271]": 0.5923983350000981,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-batchmean-1-4096-32000]": 3.165341187999843,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-batchmean-3-423-32000]": 3.103124026000046,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-batchmean-41-401-1271]": 0.5379576919998499,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-mean-1-4096-32000]": 2.0136416290000625,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-mean-3-423-32000]": 3.167686166999829,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-mean-41-401-1271]": 0.5871527729998434,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-none-1-4096-32000]": 1.039304549000235,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-none-3-423-32000]": 0.980366641000046,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-none-41-401-1271]": 0.6022642810000889,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-sum-1-4096-32000]": 3.179876357000012,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-sum-3-423-32000]": 2.232598606999886,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-1-sum-41-401-1271]": 0.5993498909999744,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-batchmean-1-4096-32000]": 3.1999989079999978,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-batchmean-3-423-32000]": 2.8830745970001317,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-batchmean-41-401-1271]": 0.5818477519998169,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-mean-1-4096-32000]": 3.1333782030001203,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-mean-3-423-32000]": 3.113541666000174,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-mean-41-401-1271]": 0.5997808880000548,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-none-1-4096-32000]": 0.9846903289999318,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-none-3-423-32000]": 0.9841488350000418,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-none-41-401-1271]": 0.5689308180000126,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-sum-1-4096-32000]": 3.122651090999966,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-sum-3-423-32000]": 3.027133260000028,
+"tests/suites/liger/test_tvd.py::Test_Liger_TVD::test_op_with_ignore_index[cutile-_100-sum-41-401-1271]": 0.5148457999999891,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_all_ignored[cutile]": 0.1500835500003177,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_backward[cutile-torch.bfloat16-1-64-32000]": 0.17003692299999784,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_backward[cutile-torch.bfloat16-2-32-1024]": 0.14900372999977662,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_backward[cutile-torch.float16-1-64-32000]": 0.1573612750000848,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_backward[cutile-torch.float16-2-32-1024]": 0.18441934899988155,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_backward_large_vocab[cutile-torch.bfloat16-1-16-100000]": 0.17175464799993279,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_backward_large_vocab[cutile-torch.bfloat16-2-8-128256]": 0.1600116090000938,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_backward_logit_scaling[cutile-torch.bfloat16-2-16-1024]": 0.14237281300006543,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_backward_softcapping[cutile-torch.bfloat16-2-16-1024]": 0.19212035800001104,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_basic[cutile-torch.bfloat16-1-128-50257]": 1.6208994279998024,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_basic[cutile-torch.bfloat16-2-64-1024]": 0.14367587399988224,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_basic[cutile-torch.bfloat16-4-32-32000]": 0.4152990589998353,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_basic[cutile-torch.float16-1-128-50257]": 2.377372328999627,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_basic[cutile-torch.float16-2-64-1024]": 0.1771534199999678,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_basic[cutile-torch.float16-4-32-32000]": 0.41306939899982353,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_large_vocab[cutile-torch.bfloat16-1-16-100000]": 0.15508102699982373,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_large_vocab[cutile-torch.bfloat16-2-8-128256]": 0.21440859200015439,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_logit_scaling[cutile-torch.bfloat16-2-16-1024]": 0.16080087699992873,
+"tests/suites/unsloth/test_cross_entropy_loss.py::Test_Unsloth_CrossEntropyLoss::test_op_softcapping[cutile-torch.bfloat16-2-16-1024]": 0.17162074100019709,
+"tests/suites/unsloth/test_fp8.py::Test_Unsloth_ActQuant::test_op[cutile-(1024, 4096)-128]": 0.005637577000015881,
+"tests/suites/unsloth/test_fp8.py::Test_Unsloth_ActQuant::test_op[cutile-(128, 512)-128]": 0.0031562489998577803,
+"tests/suites/unsloth/test_fp8.py::Test_Unsloth_ActQuant::test_op[cutile-(2048, 8192)-128]": 0.006351143000074444,
+"tests/suites/unsloth/test_fp8.py::Test_Unsloth_ActQuant::test_op[cutile-(256, 256)-128]": 0.1337379389999569,
+"tests/suites/unsloth/test_fp8.py::Test_Unsloth_ActQuant::test_op[cutile-(64, 1024)-128]": 0.0033916240004145948,
+"tests/suites/unsloth/test_fp8.py::Test_Unsloth_ActQuant::test_op_non_square[cutile-(128, 384)-128]": 0.002976902999989761,
+"tests/suites/unsloth/test_fp8.py::Test_Unsloth_ActQuant::test_op_non_square[cutile-(256, 128)-128]": 0.0029698660002850374,
+"tests/suites/unsloth/test_fp8.py::Test_Unsloth_W8A8_FP8_Matmul::test_op[cutile-1024-2048-1024-[128, 128]]": 1.31681470000035,
+"tests/suites/unsloth/test_fp8.py::Test_Unsloth_W8A8_FP8_Matmul::test_op[cutile-128-128-128-[128, 128]]": 0.4688188650000029,
+"tests/suites/unsloth/test_fp8.py::Test_Unsloth_W8A8_FP8_Matmul::test_op[cutile-256-256-256-[128, 128]]": 0.9235641510001642,
+"tests/suites/unsloth/test_fp8.py::Test_Unsloth_W8A8_FP8_Matmul::test_op[cutile-512-1024-512-[128, 128]]": 1.16430599000023,
+"tests/suites/unsloth/test_fp8.py::Test_Unsloth_WeightDequant::test_op[cutile-2048-4096-128]": 0.5339988489999996,
+"tests/suites/unsloth/test_fp8.py::Test_Unsloth_WeightDequant::test_op[cutile-256-256-128]": 0.4819045719998485,
+"tests/suites/unsloth/test_fp8.py::Test_Unsloth_WeightDequant::test_op[cutile-384-768-128]": 0.5136039689998597,
+"tests/suites/unsloth/test_fp8.py::Test_Unsloth_WeightDequant::test_op[cutile-4096-8192-128]": 0.601548245999993,
+"tests/suites/unsloth/test_fp8.py::Test_Unsloth_WeightDequant::test_op[cutile-512-1024-128]": 0.5088960139999017,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Backward::test_op[cutile-torch.bfloat16-1024-1024]": 0.1840933909998057,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Backward::test_op[cutile-torch.bfloat16-2048-5120]": 0.14594711800009463,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Backward::test_op[cutile-torch.bfloat16-256-512]": 0.1775873999997657,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Backward::test_op[cutile-torch.bfloat16-4096-8192]": 0.22675418100016032,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Backward::test_op[cutile-torch.float16-1024-1024]": 0.1433381310000641,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Backward::test_op[cutile-torch.float16-2048-5120]": 0.17590174099996148,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Backward::test_op[cutile-torch.float16-256-512]": 0.17860355600009825,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Backward::test_op[cutile-torch.float16-4096-8192]": 0.2029162789997372,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.bfloat16-(1, 1, 1024)]": 0.489243450999993,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.bfloat16-(2, 1024, 8192)]": 0.5644367779998447,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.bfloat16-(2, 128, 256)]": 0.5549066359999415,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.bfloat16-(2, 256, 4096)]": 0.5333630350000931,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.bfloat16-(4, 512, 5120)]": 0.5601696739997806,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.bfloat16-(4, 64, 512)]": 0.5505521750001208,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.float16-(1, 1, 1024)]": 0.5223439059998327,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.float16-(2, 1024, 8192)]": 0.5105633150001267,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.float16-(2, 128, 256)]": 0.5103497989998687,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.float16-(2, 256, 4096)]": 0.5697062379999807,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.float16-(4, 512, 5120)]": 0.5902446539998891,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Approx_Forward::test_op[cutile-torch.float16-(4, 64, 512)]": 0.556120349000139,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Backward::test_op[cutile-torch.bfloat16-1024-1024]": 0.2063122460001523,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Backward::test_op[cutile-torch.bfloat16-2048-5120]": 0.21028575300010743,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Backward::test_op[cutile-torch.bfloat16-256-512]": 0.1731246930000907,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Backward::test_op[cutile-torch.bfloat16-4096-8192]": 0.21270800699994652,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Backward::test_op[cutile-torch.float16-1024-1024]": 0.2027825870002289,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Backward::test_op[cutile-torch.float16-2048-5120]": 0.20374202599987257,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Backward::test_op[cutile-torch.float16-256-512]": 0.18787014099984845,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Backward::test_op[cutile-torch.float16-4096-8192]": 0.1917063959999723,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.bfloat16-(1, 1, 1024)]": 0.6242708369998127,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.bfloat16-(2, 1024, 8192)]": 0.6490382190002038,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.bfloat16-(2, 128, 256)]": 0.6394138929999826,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.bfloat16-(2, 256, 4096)]": 0.6225823629999923,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.bfloat16-(4, 512, 5120)]": 0.6039742759999172,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.bfloat16-(4, 64, 512)]": 0.5766622930000267,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.float16-(1, 1, 1024)]": 0.6055443339998874,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.float16-(2, 1024, 8192)]": 0.6340466839999408,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.float16-(2, 128, 256)]": 0.7168839549997301,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.float16-(2, 256, 4096)]": 0.6724914210001316,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.float16-(4, 512, 5120)]": 0.6981359300002623,
+"tests/suites/unsloth/test_geglu.py::Test_Unsloth_GEGLU_Exact_Forward::test_op[cutile-torch.float16-(4, 64, 512)]": 0.6793683329997293,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_W_grad[cutile-torch.bfloat16-4-16-64-64-1]": 0.004310525000164489,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_W_grad[cutile-torch.bfloat16-4-32-128-128-1]": 0.0041871430000810506,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_W_grad[cutile-torch.bfloat16-8-64-512-256-1]": 0.0048980560002291895,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_W_grad[cutile-torch.float16-4-16-64-64-1]": 0.00511947300014981,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_W_grad[cutile-torch.float16-4-32-128-128-1]": 0.004519653999977891,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_W_grad[cutile-torch.float16-8-64-512-256-1]": 0.004948314000102982,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_W_grad_topk4[cutile-torch.bfloat16-4-16-128-128-4]": 0.005731827000090561,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_W_grad_topk4[cutile-torch.bfloat16-8-8-256-128-4]": 0.005271031000120274,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_X_grad[cutile-torch.bfloat16-4-16-64-64-1]": 10.89242393099994,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_X_grad[cutile-torch.bfloat16-4-32-128-128-1]": 9.856002469000032,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_X_grad[cutile-torch.bfloat16-8-64-512-256-1]": 14.824474498999962,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_X_grad[cutile-torch.float16-4-16-64-64-1]": 12.046316986000193,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_X_grad[cutile-torch.float16-4-32-128-128-1]": 8.916284281999879,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_X_grad[cutile-torch.float16-8-64-512-256-1]": 11.867364591000069,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_X_grad_topk4[cutile-torch.bfloat16-4-16-128-128-4]": 9.41593453099972,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_backward_X_grad_topk4[cutile-torch.bfloat16-8-8-256-128-4]": 11.030241812999748,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_deterministic[cutile-8-32-256-128]": 3.0150253269998757,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_edge_cases[cutile-torch.bfloat16-[0, 0, 0, 64]-64-64]": 0.0032965020000119694,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_edge_cases[cutile-torch.bfloat16-[128]-256-128]": 3.7352756039999804,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_edge_cases[cutile-torch.bfloat16-[32, 16, 8, 8]-64-64]": 0.0034016699999028788,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_edge_cases[cutile-torch.bfloat16-[64, 0, 32, 0]-128-128]": 2.7955372280000574,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.bfloat16-16-16-768-512-1]": 7.367953390999901,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.bfloat16-16-32-2048-1024-1]": 14.361707386000035,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.bfloat16-4-16-64-64-1]": 1.476268416000039,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.bfloat16-4-32-128-128-1]": 2.1682778029996825,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.bfloat16-8-16-128-128-2]": 2.743319328000098,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.bfloat16-8-64-512-256-1]": 5.019522788000131,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.bfloat16-8-8-64-128-1]": 1.7902406600001086,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.float16-16-16-768-512-1]": 7.719074948000298,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.float16-16-32-2048-1024-1]": 16.967831626999896,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.float16-4-16-64-64-1]": 1.8829819460002,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.float16-4-32-128-128-1]": 2.327845694000189,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.float16-8-16-128-128-2]": 2.600064420000308,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.float16-8-64-512-256-1]": 4.405418885000017,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward[cutile-torch.float16-8-8-64-128-1]": 2.3336202299999513,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward_large[cutile-torch.bfloat16-128-8-768-512-1]": 11.720894618000102,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward_topk4[cutile-torch.bfloat16-4-16-128-128-4]": 2.898677513000166,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward_topk4[cutile-torch.bfloat16-8-8-256-128-4]": 2.6139918810001745,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward_upstream_model[cutile-torch.bfloat16-128-8-768-2048-8]": 8.337506396000208,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_forward_upstream_model[cutile-torch.bfloat16-16-64-8192-5120-1]": 4.73672951399999,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_non_power_of_2[cutile-torch.bfloat16-4-20-300-192]": 4.882536112999787,
+"tests/suites/unsloth/test_grouped_gemm.py::Test_Unsloth_GroupedGemm::test_op_non_power_of_2[cutile-torch.bfloat16-8-13-100-96]": 3.399145462000206,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.bfloat16-(16, 1024)]": 0.7182513649997873,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.bfloat16-(2, 4, 512)]": 0.004056658999843421,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.bfloat16-(21, 349, 1024)]": 0.7398809720002646,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.bfloat16-(32, 4096)]": 1.1759080489998723,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.bfloat16-(4, 256)]": 0.7132079749997047,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.bfloat16-(4, 300)]": 1.2468562239996572,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.bfloat16-(8, 512)]": 0.6969587870000851,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.float16-(16, 1024)]": 0.7019365190001281,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.float16-(2, 4, 512)]": 0.004192770000145174,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.float16-(21, 349, 1024)]": 0.7710498789997473,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.float16-(32, 4096)]": 1.0440413509998052,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.float16-(4, 256)]": 0.6945462000001044,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.float16-(4, 300)]": 1.2362111910001659,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op[cutile-torch.float16-(8, 512)]": 0.7272594359999403,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op_backward[cutile-(2, 4, 512)-torch.bfloat16]": 0.005433294999875216,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op_backward[cutile-(4, 2048)-torch.bfloat16]": 1.834468194999772,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op_backward[cutile-(4, 256)-torch.float16]": 0.8848309149998386,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op_backward[cutile-(4, 256)-torch.float32]": 1.414844075000019,
+"tests/suites/unsloth/test_layernorm.py::Test_Unsloth_Layernorm::test_op_backward[cutile-(8, 512)-torch.bfloat16]": 0.7173705130001053,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward[cutile-(2, 4, 512)-torch.bfloat16]": 0.006152160999818079,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward[cutile-(4, 2048)-torch.bfloat16]": 1.5755701359998966,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward[cutile-(4, 256)-torch.float16]": 0.6261514590000843,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward[cutile-(4, 256)-torch.float32]": 1.2609097519998613,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward[cutile-(8, 512)-torch.bfloat16]": 0.8636905160001334,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward_gemma[cutile-(16, 1024)-torch.bfloat16]": 0.706549639000059,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward_gemma[cutile-(2, 4, 512)-torch.bfloat16]": 0.006578468999805409,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward_gemma[cutile-(4, 256)-torch.float16]": 0.6797182580000936,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_backward_gemma[cutile-(8, 512)-torch.bfloat16]": 0.6777787909998096,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_gemma[cutile-torch.bfloat16-(16, 1024)]": 0.6484336480000366,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_gemma[cutile-torch.bfloat16-(4, 256)]": 0.6375791150001078,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_gemma[cutile-torch.bfloat16-(8, 512)]": 0.6237344859998757,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_gemma[cutile-torch.float16-(16, 1024)]": 0.6479642649999278,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_gemma[cutile-torch.float16-(4, 256)]": 0.6182493329999943,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_gemma[cutile-torch.float16-(8, 512)]": 0.5891814709998471,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.bfloat16-(16, 1024)]": 0.6514104099999258,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.bfloat16-(2, 4, 512)]": 0.004359965999583437,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.bfloat16-(21, 349, 1024)]": 0.6129418579998855,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.bfloat16-(32, 4096)]": 0.7918365489997541,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.bfloat16-(4, 256)]": 0.5992806580002252,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.bfloat16-(4, 300)]": 1.0282178870002099,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.bfloat16-(8, 512)]": 0.6432343330000094,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.float16-(16, 1024)]": 0.6270632270000078,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.float16-(2, 4, 512)]": 0.003894088999913947,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.float16-(21, 349, 1024)]": 0.6414760130001014,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.float16-(32, 4096)]": 0.7711700269999255,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.float16-(4, 256)]": 0.6073270919998777,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.float16-(4, 300)]": 0.9481650619998163,
+"tests/suites/unsloth/test_rms_layernorm.py::Test_Unsloth_RMSLayernorm::test_op_standard[cutile-torch.float16-(8, 512)]": 0.5829998469998827,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op[cutile-torch.bfloat16-1-128-4-128]": 0.5879165169999396,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op[cutile-torch.bfloat16-2-64-8-64]": 0.5743185009998797,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op[cutile-torch.bfloat16-4-32-16-64]": 0.5867083170001024,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op[cutile-torch.float16-1-128-4-128]": 0.5827101059999222,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op[cutile-torch.float16-2-64-8-64]": 0.5785357270001441,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op[cutile-torch.float16-4-32-16-64]": 0.5688274209999236,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op_backward[cutile-torch.bfloat16-1-128-4-128]": 0.29133492800019667,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op_backward[cutile-torch.bfloat16-2-64-8-64]": 0.14127327599999262,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op_backward[cutile-torch.float16-1-128-4-128]": 0.14103484199995364,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding::test_op_backward[cutile-torch.float16-2-64-8-64]": 0.2869288970005073,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op[cutile-torch.bfloat16-1-32-8-128-128]": 0.7454382539999642,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op[cutile-torch.bfloat16-2-8-8-64-64]": 0.7359485999998014,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op[cutile-torch.bfloat16-4-16-4-32-64]": 0.7167601929998,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op[cutile-torch.float16-1-32-8-128-128]": 0.7453873649997149,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op[cutile-torch.float16-2-8-8-64-64]": 0.7175122740000006,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op[cutile-torch.float16-4-16-4-32-64]": 1.0285976100003609,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op_backward[cutile-torch.bfloat16-1-32-8-128-128]": 0.1933366329999444,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op_backward[cutile-torch.bfloat16-2-8-8-64-64]": 0.17466796500002602,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op_backward[cutile-torch.float16-1-32-8-128-128]": 0.41468371299970386,
+"tests/suites/unsloth/test_rope_embedding.py::Test_Unsloth_RoPE_Embedding_QK::test_op_backward[cutile-torch.float16-2-8-8-64-64]": 0.16442170099981013,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.bfloat16-1024-1024]": 0.16950670400024137,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.bfloat16-2048-5120]": 0.17051741600016612,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.bfloat16-256-512]": 0.16179069800023171,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.bfloat16-4096-8192]": 0.20020528800000648,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.bfloat16-512-4096]": 0.17137989599996217,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.float16-1024-1024]": 0.16914017399994918,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.float16-2048-5120]": 0.16259149600023193,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.float16-256-512]": 0.1789819409998472,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.float16-4096-8192]": 0.20282407400009106,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Backward::test_op[cutile-torch.float16-512-4096]": 0.17278373199997077,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.bfloat16-(1, 1, 1024)]": 0.5636113690000002,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.bfloat16-(2, 1024, 8192)]": 0.598257945000114,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.bfloat16-(2, 128, 256)]": 0.5941371979999985,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.bfloat16-(2, 256, 4096)]": 0.583311558999867,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.bfloat16-(4, 512, 5120)]": 0.5593433520002691,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.bfloat16-(4, 64, 512)]": 0.5828949229999125,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.float16-(1, 1, 1024)]": 0.5862362660002418,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.float16-(2, 1024, 8192)]": 0.6174577150000005,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.float16-(2, 128, 256)]": 0.6039500359997874,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.float16-(2, 256, 4096)]": 0.5805169420000311,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.float16-(4, 512, 5120)]": 0.5842543880000903,
+"tests/suites/unsloth/test_swiglu.py::Test_Unsloth_SwiGLU_Forward::test_op[cutile-torch.float16-(4, 64, 512)]": 0.5870494030000373
 }