From e4d2e810869e0a8e4e3c920a429d841b083734e6 Mon Sep 17 00:00:00 2001 From: PyTorch UpdateBot Date: Mon, 16 Dec 2024 12:40:40 +0000 Subject: [PATCH] Update slow tests (#143278) This PR is auto-generated weekly by [this action](https://github.com/pytorch/pytorch/blob/main/.github/workflows/weekly.yml). Update the list of slow tests. Pull Request resolved: https://github.com/pytorch/pytorch/pull/143278 Approved by: https://github.com/pytorchbot --- test/slow_tests.json | 646 ++++++++++++++++++++++--------------------- 1 file changed, 326 insertions(+), 320 deletions(-) diff --git a/test/slow_tests.json b/test/slow_tests.json index 7fd363065ea82..a88e373d8af6b 100644 --- a/test/slow_tests.json +++ b/test/slow_tests.json @@ -1,322 +1,328 @@ { - "EndToEndLSTM (__main__.RNNTest)": 220.1296641031901, - "MultiheadAttention (__main__.ModulesTest)": 173.19766743977866, - "test_AllenaiLongformerBase_repro_cpu (__main__.CpuHalideTests)": 215.52533467610678, - "test__adaptive_avg_pool2d (__main__.CPUReproTests)": 179.0608901977539, - "test_adaptive_max_pool2d1_cpu (__main__.CpuHalideTests)": 113.46566772460938, - "test_after_aot_cpu_runtime_error (__main__.MinifierIsolateTests)": 66.69855668809679, - "test_alexnet_prefix_cpu (__main__.CpuHalideTests)": 191.67499796549478, - "test_aot_autograd_exhaustive_nn_functional_max_pool2d_cpu_float32 (__main__.TestEagerFusionOpInfoCPU)": 68.1894998550415, - "test_aot_autograd_symbolic_exhaustive_nn_functional_max_pool1d_cpu_float32 (__main__.TestEagerFusionOpInfoCPU)": 97.54266866048177, - "test_aot_autograd_symbolic_exhaustive_nn_functional_max_pool2d_cpu_float32 (__main__.TestEagerFusionOpInfoCPU)": 148.51333618164062, - "test_aot_autograd_symbolic_exhaustive_nn_functional_max_pool3d_cpu_float32 (__main__.TestEagerFusionOpInfoCPU)": 96.68933359781902, - "test_aot_autograd_symbolic_exhaustive_svd_cpu_float32 (__main__.TestEagerFusionOpInfoCPU)": 63.17433420817057, - "test_aot_autograd_symbolic_module_exhaustive_nn_TransformerDecoderLayer_cpu_float32 (__main__.TestEagerFusionModuleInfoCPU)": 115.37100219726562, - "test_avg_pool3d_backward2_cpu (__main__.CpuTests)": 497.1377766927083, - "test_avg_pool3d_backward2_cuda (__main__.GPUTests)": 89.84316762288411, - "test_avg_pool3d_backward2_dynamic_shapes_cpu (__main__.DynamicShapesCodegenCpuTests)": 512.6926676432291, - "test_avg_pool3d_backward2_dynamic_shapes_cpu (__main__.DynamicShapesCpuTests)": 504.09523518880206, - "test_avg_pool3d_backward2_dynamic_shapes_cuda (__main__.DynamicShapesGPUTests)": 128.62566630045572, - "test_avg_pool3d_backward_cpu (__main__.CpuHalideTests)": 62.03799947102865, - "test_backward_nn_functional_multi_head_attention_forward_cpu_float32 (__main__.TestCompositeComplianceCPU)": 69.20466613769531, - "test_backward_nn_functional_multi_head_attention_forward_cuda_float32 (__main__.TestCompositeComplianceCUDA)": 63.945332845052086, - "test_basic_cpu (__main__.EfficientConvBNEvalCpuTests)": 272.8941124810113, - "test_basic_cuda (__main__.EfficientConvBNEvalGpuTests)": 193.1396687825521, - "test_builtin_equivalent_funcs (__main__.TorchFunctionModeTests)": 107.44954588918975, - "test_collect_callgrind (__main__.TestBenchmarkUtils)": 364.11266750759546, - "test_comprehensive_constant_pad_nd_cpu_float16 (__main__.TestInductorOpInfoCPU)": 82.01866912841797, - "test_comprehensive_constant_pad_nd_cpu_float32 (__main__.TestInductorOpInfoCPU)": 81.21833038330078, - "test_comprehensive_constant_pad_nd_cpu_float64 (__main__.TestInductorOpInfoCPU)": 83.85000101725261, - "test_comprehensive_constant_pad_nd_cpu_int32 (__main__.TestInductorOpInfoCPU)": 77.375, - "test_comprehensive_constant_pad_nd_cpu_int64 (__main__.TestInductorOpInfoCPU)": 84.44499969482422, - "test_comprehensive_diff_cpu_bool (__main__.TestInductorOpInfoCPU)": 118.20800018310547, - "test_comprehensive_diff_cpu_float32 (__main__.TestInductorOpInfoCPU)": 121.04333241780598, - "test_comprehensive_diff_cpu_float64 (__main__.TestInductorOpInfoCPU)": 118.59633382161458, - "test_comprehensive_diff_cpu_int32 (__main__.TestInductorOpInfoCPU)": 113.44066619873047, - "test_comprehensive_diff_cpu_int64 (__main__.TestInductorOpInfoCPU)": 117.9990005493164, - "test_comprehensive_diff_cuda_complex128 (__main__.TestDecompCUDA)": 76.80516624450684, - "test_comprehensive_diff_cuda_complex64 (__main__.TestDecompCUDA)": 124.16683578491211, - "test_comprehensive_diff_cuda_float32 (__main__.TestDecompCUDA)": 91.94649759928386, - "test_comprehensive_dist_cpu_float16 (__main__.TestInductorOpInfoCPU)": 96.91400146484375, - "test_comprehensive_dist_cpu_float32 (__main__.TestInductorOpInfoCPU)": 94.62099965413411, - "test_comprehensive_dist_cpu_float64 (__main__.TestInductorOpInfoCPU)": 92.66366831461589, - "test_comprehensive_eye_cpu_bool (__main__.TestInductorOpInfoCPU)": 147.22899881998697, - "test_comprehensive_eye_cpu_float16 (__main__.TestInductorOpInfoCPU)": 142.98733520507812, - "test_comprehensive_eye_cpu_float32 (__main__.TestInductorOpInfoCPU)": 146.37999979654947, - "test_comprehensive_eye_cpu_float64 (__main__.TestInductorOpInfoCPU)": 143.72832743326822, - "test_comprehensive_eye_cpu_int32 (__main__.TestInductorOpInfoCPU)": 142.44133504231772, - "test_comprehensive_eye_cpu_int64 (__main__.TestInductorOpInfoCPU)": 141.59266662597656, - "test_comprehensive_grid_sampler_2d_cpu_bfloat16 (__main__.TestDecompCPU)": 81.18483310275607, - "test_comprehensive_grid_sampler_2d_cpu_float16 (__main__.TestDecompCPU)": 84.32133271959093, - "test_comprehensive_grid_sampler_2d_cpu_float16 (__main__.TestInductorOpInfoCPU)": 78.75933329264323, - "test_comprehensive_grid_sampler_2d_cpu_float32 (__main__.TestDecompCPU)": 355.1566670735677, - "test_comprehensive_grid_sampler_2d_cpu_float32 (__main__.TestInductorOpInfoCPU)": 85.89933268229167, - "test_comprehensive_grid_sampler_2d_cpu_float64 (__main__.TestDecompCPU)": 338.6666666666667, - "test_comprehensive_grid_sampler_2d_cpu_float64 (__main__.TestInductorOpInfoCPU)": 86.28433481852214, - "test_comprehensive_grid_sampler_2d_cuda_bfloat16 (__main__.TestDecompCUDA)": 218.52383677164713, - "test_comprehensive_grid_sampler_2d_cuda_float16 (__main__.TestDecompCUDA)": 211.49849700927734, - "test_comprehensive_grid_sampler_2d_cuda_float32 (__main__.TestDecompCUDA)": 742.4113362630209, - "test_comprehensive_grid_sampler_2d_cuda_float64 (__main__.TestDecompCUDA)": 759.7113342285156, - "test_comprehensive_linalg_lu_solve_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 71.63033345540364, - "test_comprehensive_linalg_lu_solve_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 66.56185722351074, - "test_comprehensive_linalg_norm_cpu_float32 (__main__.TestInductorOpInfoCPU)": 61.718714577811106, - "test_comprehensive_linalg_norm_cpu_float64 (__main__.TestInductorOpInfoCPU)": 62.04695256551107, - "test_comprehensive_linalg_solve_triangular_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 66.42046686808268, - "test_comprehensive_linalg_solve_triangular_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 63.64026590983073, - "test_comprehensive_linalg_svd_cuda_complex64 (__main__.TestDecompCUDA)": 109.45283381144206, - "test_comprehensive_linalg_vector_norm_cpu_float16 (__main__.TestInductorOpInfoCPU)": 230.26200358072916, - "test_comprehensive_linalg_vector_norm_cpu_float32 (__main__.TestInductorOpInfoCPU)": 229.3096669514974, - "test_comprehensive_linalg_vector_norm_cpu_float64 (__main__.TestInductorOpInfoCPU)": 229.84033711751303, - "test_comprehensive_linalg_vector_norm_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 63.27973302205404, - "test_comprehensive_linalg_vector_norm_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 65.21060028076172, - "test_comprehensive_logspace_cpu_float32 (__main__.TestInductorOpInfoCPU)": 490.0329996744792, - "test_comprehensive_logspace_cpu_float64 (__main__.TestInductorOpInfoCPU)": 495.6479899088542, - "test_comprehensive_logspace_cpu_int32 (__main__.TestInductorOpInfoCPU)": 485.4586588541667, - "test_comprehensive_logspace_cpu_int64 (__main__.TestInductorOpInfoCPU)": 484.2929992675781, - "test_comprehensive_logspace_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 66.42806727091471, - "test_comprehensive_logspace_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 65.60626703898112, - "test_comprehensive_masked_amax_cpu_float16 (__main__.TestInductorOpInfoCPU)": 113.08300018310547, - "test_comprehensive_masked_amax_cpu_float32 (__main__.TestInductorOpInfoCPU)": 112.02633412679036, - "test_comprehensive_masked_amax_cpu_float64 (__main__.TestInductorOpInfoCPU)": 108.22500101725261, - "test_comprehensive_masked_amax_cpu_int32 (__main__.TestInductorOpInfoCPU)": 110.25033315022786, - "test_comprehensive_masked_amax_cpu_int64 (__main__.TestInductorOpInfoCPU)": 103.6489995320638, - "test_comprehensive_masked_amin_cpu_float16 (__main__.TestInductorOpInfoCPU)": 111.12166849772136, - "test_comprehensive_masked_amin_cpu_float32 (__main__.TestInductorOpInfoCPU)": 106.92000071207683, - "test_comprehensive_masked_amin_cpu_float64 (__main__.TestInductorOpInfoCPU)": 107.7066650390625, - "test_comprehensive_masked_amin_cpu_int32 (__main__.TestInductorOpInfoCPU)": 102.8219985961914, - "test_comprehensive_masked_amin_cpu_int64 (__main__.TestInductorOpInfoCPU)": 104.96033223470052, - "test_comprehensive_masked_argmax_cpu_float16 (__main__.TestInductorOpInfoCPU)": 61.22890490577335, - "test_comprehensive_masked_argmax_cpu_float32 (__main__.TestInductorOpInfoCPU)": 61.557095482235866, - "test_comprehensive_masked_argmax_cpu_float64 (__main__.TestInductorOpInfoCPU)": 61.328714461553666, - "test_comprehensive_masked_argmax_cpu_int32 (__main__.TestInductorOpInfoCPU)": 60.466380709693546, - "test_comprehensive_masked_argmin_cpu_float16 (__main__.TestInductorOpInfoCPU)": 61.34333310808454, - "test_comprehensive_masked_argmin_cpu_float32 (__main__.TestInductorOpInfoCPU)": 60.81590452648344, - "test_comprehensive_masked_argmin_cpu_float64 (__main__.TestInductorOpInfoCPU)": 61.18657157534645, - "test_comprehensive_masked_argmin_cpu_int32 (__main__.TestInductorOpInfoCPU)": 60.31990450904483, - "test_comprehensive_masked_argmin_cpu_int64 (__main__.TestInductorOpInfoCPU)": 60.0750474475679, - "test_comprehensive_masked_mean_cpu_bool (__main__.TestInductorOpInfoCPU)": 108.3163350423177, - "test_comprehensive_masked_mean_cpu_float16 (__main__.TestInductorOpInfoCPU)": 115.56999969482422, - "test_comprehensive_masked_mean_cpu_float32 (__main__.TestInductorOpInfoCPU)": 106.30500030517578, - "test_comprehensive_masked_mean_cpu_float64 (__main__.TestInductorOpInfoCPU)": 105.89099884033203, - "test_comprehensive_masked_mean_cpu_int32 (__main__.TestInductorOpInfoCPU)": 111.5923360188802, - "test_comprehensive_masked_mean_cpu_int64 (__main__.TestInductorOpInfoCPU)": 105.85166676839192, - "test_comprehensive_masked_norm_cpu_float16 (__main__.TestInductorOpInfoCPU)": 549.9986572265625, - "test_comprehensive_masked_norm_cpu_float32 (__main__.TestInductorOpInfoCPU)": 542.0886840820312, - "test_comprehensive_masked_norm_cpu_float64 (__main__.TestInductorOpInfoCPU)": 535.2476603190104, - "test_comprehensive_masked_norm_cuda_float16 (__main__.TestInductorOpInfoCUDA)": 152.52733357747397, - "test_comprehensive_masked_norm_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 152.16699981689453, - "test_comprehensive_masked_norm_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 154.0283317565918, - "test_comprehensive_masked_prod_cpu_bool (__main__.TestInductorOpInfoCPU)": 105.93000030517578, - "test_comprehensive_masked_prod_cpu_float16 (__main__.TestInductorOpInfoCPU)": 109.06000010172527, - "test_comprehensive_masked_prod_cpu_float32 (__main__.TestInductorOpInfoCPU)": 108.55166625976562, - "test_comprehensive_masked_prod_cpu_float64 (__main__.TestInductorOpInfoCPU)": 113.97900136311848, - "test_comprehensive_masked_prod_cpu_int32 (__main__.TestInductorOpInfoCPU)": 104.53033447265625, - "test_comprehensive_masked_prod_cpu_int64 (__main__.TestInductorOpInfoCPU)": 107.53666687011719, - "test_comprehensive_masked_sum_cpu_bool (__main__.TestInductorOpInfoCPU)": 108.38800303141277, - "test_comprehensive_masked_sum_cpu_float16 (__main__.TestInductorOpInfoCPU)": 104.35866800944011, - "test_comprehensive_masked_sum_cpu_float32 (__main__.TestInductorOpInfoCPU)": 108.9250005086263, - "test_comprehensive_masked_sum_cpu_float64 (__main__.TestInductorOpInfoCPU)": 106.02133178710938, - "test_comprehensive_masked_sum_cpu_int32 (__main__.TestInductorOpInfoCPU)": 104.08533223470052, - "test_comprehensive_masked_sum_cpu_int64 (__main__.TestInductorOpInfoCPU)": 105.32666778564453, - "test_comprehensive_nn_functional_gaussian_nll_loss_cuda_float32 (__main__.TestDecompCUDA)": 100.34150060017903, - "test_comprehensive_nn_functional_gaussian_nll_loss_cuda_float64 (__main__.TestDecompCUDA)": 77.54533386230469, - "test_comprehensive_nn_functional_glu_cpu_float16 (__main__.TestInductorOpInfoCPU)": 86.9943339029948, - "test_comprehensive_nn_functional_glu_cpu_float32 (__main__.TestInductorOpInfoCPU)": 85.09500122070312, - "test_comprehensive_nn_functional_glu_cpu_float64 (__main__.TestInductorOpInfoCPU)": 83.10499827067058, - "test_comprehensive_nn_functional_grid_sample_cpu_float32 (__main__.TestDecompCPU)": 96.50466664632161, - "test_comprehensive_nn_functional_grid_sample_cpu_float64 (__main__.TestDecompCPU)": 84.42766571044922, - "test_comprehensive_nn_functional_grid_sample_cuda_float32 (__main__.TestDecompCUDA)": 171.1691640218099, - "test_comprehensive_nn_functional_grid_sample_cuda_float64 (__main__.TestDecompCUDA)": 163.23983256022134, - "test_comprehensive_nn_functional_interpolate_bicubic_cpu_uint8 (__main__.TestInductorOpInfoCPU)": 60.09000015258789, - "test_comprehensive_nn_functional_interpolate_bicubic_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 68.31720021565755, - "test_comprehensive_nn_functional_interpolate_bicubic_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 70.2181989034017, - "test_comprehensive_nn_functional_interpolate_trilinear_cuda_float32 (__main__.TestDecompCUDA)": 60.08900006612142, - "test_comprehensive_nn_functional_max_pool1d_cpu_float16 (__main__.TestInductorOpInfoCPU)": 193.65399678548178, - "test_comprehensive_nn_functional_max_pool1d_cpu_float32 (__main__.TestInductorOpInfoCPU)": 196.34233601888022, - "test_comprehensive_nn_functional_max_pool1d_cpu_float64 (__main__.TestInductorOpInfoCPU)": 198.0546671549479, - "test_comprehensive_nn_functional_max_pool2d_cpu_float16 (__main__.TestInductorOpInfoCPU)": 985.8349812825521, - "test_comprehensive_nn_functional_max_pool2d_cpu_float32 (__main__.TestInductorOpInfoCPU)": 945.8166707356771, - "test_comprehensive_nn_functional_max_pool2d_cpu_float64 (__main__.TestInductorOpInfoCPU)": 997.4046427408854, - "test_comprehensive_nn_functional_max_pool2d_cpu_int32 (__main__.TestInductorOpInfoCPU)": 900.6526692708334, - "test_comprehensive_nn_functional_max_pool2d_cpu_int64 (__main__.TestInductorOpInfoCPU)": 948.7556559244791, - "test_comprehensive_nn_functional_max_pool2d_cuda_float16 (__main__.TestInductorOpInfoCUDA)": 880.8436686197916, - "test_comprehensive_nn_functional_max_pool2d_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 881.2054951985677, - "test_comprehensive_nn_functional_max_pool2d_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 881.2033284505209, - "test_comprehensive_nn_functional_max_pool3d_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 60.17899983723958, - "test_comprehensive_nn_functional_max_unpool2d_cpu_float16 (__main__.TestInductorOpInfoCPU)": 234.34032694498697, - "test_comprehensive_nn_functional_max_unpool2d_cpu_float32 (__main__.TestInductorOpInfoCPU)": 232.85466512044272, - "test_comprehensive_nn_functional_max_unpool2d_cpu_float64 (__main__.TestInductorOpInfoCPU)": 229.92400105794272, - "test_comprehensive_nn_functional_max_unpool2d_cuda_float16 (__main__.TestInductorOpInfoCUDA)": 62.94233322143555, - "test_comprehensive_nn_functional_max_unpool2d_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 105.60166549682617, - "test_comprehensive_nn_functional_max_unpool2d_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 105.9643325805664, - "test_comprehensive_nn_functional_max_unpool3d_cpu_float16 (__main__.TestInductorOpInfoCPU)": 147.85733540852866, - "test_comprehensive_nn_functional_max_unpool3d_cpu_float32 (__main__.TestInductorOpInfoCPU)": 150.83533732096353, - "test_comprehensive_nn_functional_max_unpool3d_cpu_float64 (__main__.TestInductorOpInfoCPU)": 148.5616658528646, - "test_comprehensive_nn_functional_pad_constant_cpu_float16 (__main__.TestInductorOpInfoCPU)": 80.99933369954427, - "test_comprehensive_nn_functional_pad_constant_cpu_float32 (__main__.TestInductorOpInfoCPU)": 83.31099955240886, - "test_comprehensive_nn_functional_pad_constant_cpu_float64 (__main__.TestInductorOpInfoCPU)": 80.20166524251302, - "test_comprehensive_nn_functional_pad_constant_cpu_int64 (__main__.TestInductorOpInfoCPU)": 79.47300211588542, - "test_comprehensive_nn_functional_poisson_nll_loss_cpu_float16 (__main__.TestInductorOpInfoCPU)": 132.75399780273438, - "test_comprehensive_nn_functional_poisson_nll_loss_cpu_float32 (__main__.TestInductorOpInfoCPU)": 133.9810028076172, - "test_comprehensive_nn_functional_poisson_nll_loss_cpu_float64 (__main__.TestInductorOpInfoCPU)": 139.02233378092447, - "test_comprehensive_nn_functional_poisson_nll_loss_cpu_int32 (__main__.TestInductorOpInfoCPU)": 127.22966766357422, - "test_comprehensive_nn_functional_poisson_nll_loss_cpu_int64 (__main__.TestInductorOpInfoCPU)": 127.71566772460938, - "test_comprehensive_nn_functional_unfold_cpu_bool (__main__.TestInductorOpInfoCPU)": 143.62633260091147, - "test_comprehensive_nn_functional_unfold_cpu_float16 (__main__.TestInductorOpInfoCPU)": 271.15667724609375, - "test_comprehensive_nn_functional_unfold_cpu_float32 (__main__.TestInductorOpInfoCPU)": 271.4836730957031, - "test_comprehensive_nn_functional_unfold_cpu_float64 (__main__.TestInductorOpInfoCPU)": 266.7383321126302, - "test_comprehensive_nn_functional_unfold_cuda_float16 (__main__.TestInductorOpInfoCUDA)": 72.5027333577474, - "test_comprehensive_nn_functional_unfold_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 71.69033406575521, - "test_comprehensive_nn_functional_unfold_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 77.27240041097005, - "test_comprehensive_ormqr_cpu_complex64 (__main__.TestDecompCPU)": 60.75344467163086, - "test_comprehensive_ormqr_cuda_complex128 (__main__.TestDecompCUDA)": 91.64133326212566, - "test_comprehensive_ormqr_cuda_complex64 (__main__.TestDecompCUDA)": 104.39499918619792, - "test_comprehensive_ormqr_cuda_float32 (__main__.TestDecompCUDA)": 61.26383399963379, - "test_comprehensive_ormqr_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 120.8076680501302, - "test_comprehensive_ormqr_cuda_float64 (__main__.TestDecompCUDA)": 66.78233295016818, - "test_comprehensive_ormqr_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 92.47266642252605, - "test_comprehensive_pca_lowrank_cuda_complex128 (__main__.TestDecompCUDA)": 63.50544357299805, - "test_comprehensive_svd_cuda_complex64 (__main__.TestDecompCUDA)": 61.66049893697103, - "test_cond_autograd_nested (__main__.TestControlFlow)": 75.12622282240126, - "test_constructor_autograd_SparseBSC_cuda (__main__.TestSparseAnyCUDA)": 92.76866658528645, - "test_constructor_autograd_SparseBSR_cuda (__main__.TestSparseAnyCUDA)": 81.77316729227702, - "test_constructor_autograd_SparseCSC_cuda (__main__.TestSparseAnyCUDA)": 69.01483408610027, - "test_conv1d_basic (__main__.TestXNNPACKConv1dTransformPass)": 235.45655483669705, - "test_conv1d_with_relu_fc (__main__.TestXNNPACKConv1dTransformPass)": 418.4912295871311, - "test_conv2d_unary_cpu_cpp_wrapper (__main__.TestCppWrapper)": 277.8283386230469, - "test_conv_transpose_with_output_size_and_no_batch_dim_ConvTranspose3d_cuda (__main__.TestConvolutionNNDeviceTypeCUDA)": 74.30288975696183, - "test_correctness_AdamW_use_closure_False_cuda_float32 (__main__.CompiledOptimizerParityTestsCUDA)": 63.32433255513509, - "test_correctness_AdamW_use_closure_True_cuda_float32 (__main__.CompiledOptimizerParityTestsCUDA)": 118.41116333007812, - "test_correctness_Adam_use_closure_False_cuda_float32 (__main__.CompiledOptimizerParityTestsCUDA)": 63.608333587646484, - "test_correctness_Adam_use_closure_True_cuda_float32 (__main__.CompiledOptimizerParityTestsCUDA)": 118.49500147501628, - "test_correctness_NAdam_use_closure_True_cuda_float32 (__main__.CompiledOptimizerParityTestsCUDA)": 73.29433377583821, - "test_count_nonzero_all (__main__.TestBool)": 618.1965535481771, - "test_custom_module_lstm (__main__.TestQuantizedOps)": 451.38599480523004, - "test_dispatch_symbolic_meta_outplace_all_strides_nn_functional_gaussian_nll_loss_cuda_float32 (__main__.TestMetaCUDA)": 78.72033055623372, - "test_dtensor_op_db_nn_functional_gaussian_nll_loss_cpu_float32 (__main__.TestDTensorOpsCPU)": 95.11700185139973, - "test_eig_check_magma_cuda_float32 (__main__.TestLinalgCUDA)": 74.53150256474812, - "test_fn_fwgrad_bwgrad_cumprod_cuda_complex128 (__main__.TestFwdGradientsCUDA)": 65.97416750590007, - "test_fn_gradgrad_cumprod_cuda_complex128 (__main__.TestBwdGradientsCUDA)": 77.60833358764648, - "test_fn_gradgrad_map_nested_cpu_float64 (__main__.TestBwdGradientsCPU)": 81.31566874186198, - "test_fn_gradgrad_map_triple_nested_cpu_float64 (__main__.TestBwdGradientsCPU)": 485.8236592610677, - "test_fn_gradgrad_map_triple_nested_cuda_float64 (__main__.TestBwdGradientsCUDA)": 320.9843343098958, - "test_fuse_large_params_cpu (__main__.CpuTests)": 62.273110707600914, - "test_fuse_large_params_dynamic_shapes_cpu (__main__.DynamicShapesCodegenCpuTests)": 88.22533416748047, - "test_fuse_large_params_dynamic_shapes_cpu (__main__.DynamicShapesCpuTests)": 88.22255622016058, - "test_fuse_large_params_dynamic_shapes_cuda (__main__.DynamicShapesCodegenGPUTests)": 71.54350026448567, - "test_fuse_large_params_dynamic_shapes_cuda (__main__.DynamicShapesGPUTests)": 70.71649932861328, - "test_grad_nn_Transformer_cuda_float64 (__main__.TestModuleCUDA)": 78.63566716512044, - "test_gradgrad_nn_LSTM_eval_mode_cuda_float64 (__main__.TestModuleCUDA)": 108.80233510335286, - "test_gradgrad_nn_LSTM_train_mode_cuda_float64 (__main__.TestModuleCUDA)": 92.26133219401042, - "test_gradgrad_nn_TransformerDecoderLayer_cuda_float64 (__main__.TestModuleCUDA)": 183.13116709391275, - "test_gradgrad_nn_TransformerEncoder_eval_mode_cuda_float64 (__main__.TestModuleCUDA)": 107.90300114949544, - "test_gradgrad_nn_TransformerEncoder_train_mode_cuda_float64 (__main__.TestModuleCUDA)": 116.88283284505208, - "test_grid_sampler_2d_cpu (__main__.CpuHalideTests)": 185.94166564941406, - "test_indexing (__main__.TestAutogradWithCompiledAutograd)": 67.41155582004123, - "test_indirect_device_assert (__main__.TritonCodeGenTests)": 192.8030039469401, - "test_inductor_no_recursionerror_on_for_loops_dynamic_shapes (__main__.DynamicShapesReproTests)": 66.80844412909613, - "test_inplace_gradgrad_cumprod_cuda_complex128 (__main__.TestBwdGradientsCUDA)": 76.77566719055176, - "test_inputs_overlapping_with_mutation_stress_dynamic_shapes (__main__.DynamicShapesAotAutogradFallbackTests)": 148.20703983306885, - "test_jit_cuda_archflags (__main__.TestCppExtensionJIT)": 115.02666727701823, - "test_linalg_solve_triangular_large_cuda_complex128 (__main__.TestLinalgCUDA)": 654.1164906819662, - "test_linalg_solve_triangular_large_cuda_complex64 (__main__.TestLinalgCUDA)": 75.0540002187093, - "test_linear (__main__.TestStaticQuantizedModule)": 137.5438872443305, - "test_linear_binary_cpp_wrapper (__main__.TestCppWrapper)": 169.01766459147134, - "test_linear_binary_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 177.24600219726562, - "test_linear_binary_folding_cpu (__main__.FreezingCpuTests)": 60.86720863978068, - "test_linear_packed_cpp_wrapper (__main__.TestCppWrapper)": 81.38533274332683, - "test_linear_packed_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 79.2270024617513, - "test_linear_relu (__main__.TestStaticQuantizedModule)": 60.32511181301541, - "test_lobpcg_ortho_cuda_float64 (__main__.TestLinalgCUDA)": 72.1783332824707, - "test_lstm_cpu (__main__.TestMkldnnCPU)": 69.95388910505507, - "test_many_overlapping_inputs_does_not_explode_guards_dynamic_shapes (__main__.DynamicShapesReproTests)": 109.84462547302246, - "test_max_autotune_cutlass_backend_addmm_dynamic_False_max_autotune_gemm_backends_ATen,Triton,CUTLASS (__main__.TestCutlassBackend)": 85.27133178710938, - "test_max_pool2d_with_indices_backward4_dynamic_shapes_cpu (__main__.DynamicShapesCodegenCpuTests)": 63.74255498250326, - "test_max_pool2d_with_indices_backward4_dynamic_shapes_cpu (__main__.DynamicShapesCpuTests)": 63.035000271267364, - "test_memory_format_operators_cpu (__main__.TestTorchDeviceTypeCPU)": 68.88351071957085, - "test_memory_format_operators_cuda (__main__.TestTorchDeviceTypeCUDA)": 66.61290535330772, - "test_mixed_mm_exhaustive_dtypes (__main__.TestPatternMatcher)": 62.811166763305664, - "test_proper_exit (__main__.TestDataLoader)": 218.88583374023438, - "test_proper_exit (__main__.TestDataLoaderPersistentWorkers)": 216.4471689860026, - "test_qconv2d_add_cpu_cpp_wrapper (__main__.TestCppWrapper)": 62.98266728719076, - "test_qconv2d_add_cpu_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 65.98199971516927, - "test_qconv2d_add_relu_cpu_cpp_wrapper (__main__.TestCppWrapper)": 63.97133255004883, - "test_qconv2d_add_relu_cpu_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 65.84966532389323, - "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_False_is_dynamic_False (__main__.TestPatternMatcher)": 73.92566935221355, - "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_False_is_dynamic_False_cpp_wrapper (__main__.TestCppWrapper)": 118.22500101725261, - "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_False_is_dynamic_False_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 115.06033325195312, - "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_False_is_dynamic_True_cpp_wrapper (__main__.TestCppWrapper)": 123.65566507975261, - "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_False_is_dynamic_True_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 115.93033091227214, - "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_True_is_dynamic_False_cpp_wrapper (__main__.TestCppWrapper)": 111.03600056966145, - "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_True_is_dynamic_False_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 112.9943339029948, - "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_True_is_dynamic_True_cpp_wrapper (__main__.TestCppWrapper)": 113.60833485921223, - "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_True_is_dynamic_True_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 114.89666748046875, - "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_False_is_dynamic_False (__main__.TestPatternMatcher)": 72.23033142089844, - "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_False_is_dynamic_False_cpp_wrapper (__main__.TestCppWrapper)": 113.50899759928386, - "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_False_is_dynamic_False_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 114.69200134277344, - "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_False_is_dynamic_True (__main__.TestPatternMatcher)": 72.91333516438802, - "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_False_is_dynamic_True_cpp_wrapper (__main__.TestCppWrapper)": 115.02433268229167, - "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_False_is_dynamic_True_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 114.52066802978516, - "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_True_is_dynamic_False_cpp_wrapper (__main__.TestCppWrapper)": 113.70533243815105, - "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_True_is_dynamic_False_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 124.88433583577473, - "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_True_is_dynamic_True (__main__.TestPatternMatcher)": 72.26066589355469, - "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_True_is_dynamic_True_cpp_wrapper (__main__.TestCppWrapper)": 133.6240005493164, - "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_True_is_dynamic_True_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 114.92199961344402, - "test_quick_core_backward__unsafe_masked_index_cpu_float64 (__main__.TestDecompCPU)": 361.30833943684894, - "test_quick_core_backward__unsafe_masked_index_cuda_float64 (__main__.TestDecompCUDA)": 630.3721669514974, - "test_quick_core_backward__unsafe_masked_index_put_accumulate_cpu_float64 (__main__.TestDecompCPU)": 579.5943400065104, - "test_quick_core_backward__unsafe_masked_index_put_accumulate_cuda_float64 (__main__.TestDecompCUDA)": 931.8911641438802, - "test_quick_core_backward_expand_copy_cuda_float64 (__main__.TestDecompCUDA)": 67.74916712443034, - "test_quick_core_backward_nn_functional_max_unpool3d_grad_cpu_float64 (__main__.TestDecompCPU)": 67.52499898274739, - "test_quick_core_backward_nn_functional_max_unpool3d_grad_cuda_float64 (__main__.TestDecompCUDA)": 186.8096669514974, - "test_quick_core_backward_roll_cpu_float64 (__main__.TestDecompCPU)": 83.84033203125, - "test_quick_core_backward_roll_cuda_float64 (__main__.TestDecompCUDA)": 151.91033172607422, - "test_quick_core_backward_select_scatter_cpu_float64 (__main__.TestDecompCPU)": 64.96833377414279, - "test_quick_core_backward_select_scatter_cuda_float64 (__main__.TestDecompCUDA)": 98.79216639200847, - "test_quick_core_backward_split_with_sizes_copy_cpu_float64 (__main__.TestDecompCPU)": 66.73466491699219, - "test_quick_core_backward_split_with_sizes_copy_cuda_float64 (__main__.TestDecompCUDA)": 120.15350087483723, - "test_quick_core_backward_std_cuda_float64 (__main__.TestDecompCUDA)": 75.97200012207031, - "test_replicatepad_64bit_indexing_cuda_float16 (__main__.TestNNDeviceTypeCUDA)": 315.7243347167969, - "test_reveal_module_list.py (__main__.TestTyping)": 74.03495261782692, - "test_rosenbrock_sparse_with_lrsched_False_SGD_cuda_float64 (__main__.TestOptimRenewedCUDA)": 67.20899868011475, - "test_rosenbrock_sparse_with_lrsched_True_SGD_cuda_float64 (__main__.TestOptimRenewedCUDA)": 77.53233353296916, - "test_save_load_large_string_attribute (__main__.TestSaveLoad)": 99.41733296712239, - "test_shuffler_iterdatapipe (__main__.IntegrationTestDataLoaderDataPipe)": 162.41844346788196, - "test_slow_tasks (__main__.TestFunctionalAutogradBenchmark)": 125.97477934095595, - "test_sort_stable_cpu (__main__.CpuTritonTests)": 75.2596664428711, - "test_split_cumsum_cpu (__main__.CpuTritonTests)": 89.3259989420573, - "test_svd_lowrank_cuda_complex128 (__main__.TestLinalgCUDA)": 212.526008049647, - "test_terminate_handler_on_crash (__main__.TestTorch)": 70.78766715526581, - "test_terminate_signal (__main__.ForkTest)": 105.19766641490989, - "test_terminate_signal (__main__.ParallelForkServerShouldWorkTest)": 105.57222100264497, - "test_terminate_signal (__main__.SpawnTest)": 109.08366616566975, - "test_transformer_backend_inductor_fullgraph_True (__main__.TestFullyShardCompile)": 111.81922234429254, - "test_transpose_copy (__main__.CPUReproTests)": 69.31099955240886, - "test_triton_bsr_scatter_mm_blocksize_64_cuda_bfloat16 (__main__.TestSparseCompressedTritonKernelsCUDA)": 67.60133171081543, - "test_triton_bsr_scatter_mm_blocksize_64_cuda_float16 (__main__.TestSparseCompressedTritonKernelsCUDA)": 66.02666727701823, - "test_triton_bsr_scatter_mm_blocksize_64_cuda_float32 (__main__.TestSparseCompressedTritonKernelsCUDA)": 73.4166653951009, - "test_triton_bsr_softmax_cuda_bfloat16 (__main__.TestSparseCompressedTritonKernelsCUDA)": 150.1351687113444, - "test_triton_bsr_softmax_cuda_float16 (__main__.TestSparseCompressedTritonKernelsCUDA)": 136.73966852823892, - "test_triton_bsr_softmax_cuda_float32 (__main__.TestSparseCompressedTritonKernelsCUDA)": 118.15099970499675, - "test_unary_ops (__main__.TestTEFuserDynamic)": 200.08266872829861, - "test_unary_ops (__main__.TestTEFuserStatic)": 185.98278141021729, - "test_upsample_bicubic2d_cpu (__main__.CpuHalideTests)": 96.19566599527995, - "test_variant_consistency_jit_nn_functional_max_pool2d_cpu_float32 (__main__.TestJitCPU)": 84.51966603597005, - "test_variant_consistency_jit_nn_functional_max_pool2d_cuda_float32 (__main__.TestJitCUDA)": 70.91866683959961, - "test_vec_bitwise (__main__.CPUReproTests)": 64.92966667811076, - "test_views1_dynamic_shapes_cpu (__main__.DynamicShapesCodegenCpuTests)": 61.070791721343994, - "test_views1_dynamic_shapes_cpu (__main__.DynamicShapesCpuTests)": 60.762874921162926, - "test_vmapjvpvjp_linalg_lu_solve_cuda_float32 (__main__.TestOperatorsCUDA)": 75.43333307902019, - "test_vmapjvpvjp_linalg_svd_cuda_float32 (__main__.TestOperatorsCUDA)": 60.10366694132487, - "test_vmapjvpvjp_max_pool2d_with_indices_backward_cpu_float32 (__main__.TestOperatorsCPU)": 63.91166559855143, - "test_vmapjvpvjp_max_pool2d_with_indices_backward_cuda_float32 (__main__.TestOperatorsCUDA)": 67.3610013326009, - "test_vmapjvpvjp_nn_functional_conv2d_cpu_float32 (__main__.TestOperatorsCPU)": 66.57408332824707, - "test_vmapjvpvjp_nn_functional_max_pool2d_cpu_float32 (__main__.TestOperatorsCPU)": 96.97366714477539, - "test_vmapjvpvjp_nn_functional_max_pool2d_cuda_float32 (__main__.TestOperatorsCUDA)": 73.43949953715007, - "test_vmapjvpvjp_svd_cuda_float32 (__main__.TestOperatorsCUDA)": 62.89116605122884, - "test_vmapjvpvjp_unbind_cpu_float32 (__main__.TestOperatorsCPU)": 66.54466756184895, - "test_vmapjvpvjp_unbind_cuda_float32 (__main__.TestOperatorsCUDA)": 81.35183270772298, - "test_vmapvjpvjp_meshgrid_list_of_tensors_cuda_float32 (__main__.TestOperatorsCUDA)": 72.34049987792969, - "test_vmapvjpvjp_meshgrid_variadic_tensors_cuda_float32 (__main__.TestOperatorsCUDA)": 69.77216529846191, - "test_vmapvjpvjp_nn_functional_bilinear_cuda_float32 (__main__.TestOperatorsCUDA)": 136.2606658935547 + "EndToEndLSTM (__main__.RNNTest)": 224.2776641845703, + "MultiheadAttention (__main__.ModulesTest)": 170.8386688232422, + "test_AllenaiLongformerBase_repro_cpu (__main__.CpuHalideTests)": 215.9730021158854, + "test__adaptive_avg_pool2d (__main__.CPUReproTests)": 190.66033342149524, + "test_adaptive_max_pool2d1_cpu (__main__.CpuHalideTests)": 114.27133433024089, + "test_after_aot_cpu_runtime_error (__main__.MinifierIsolateTests)": 71.38377804226346, + "test_alexnet_prefix_cpu (__main__.CpuHalideTests)": 191.32866414388022, + "test_aot_autograd_exhaustive_nn_functional_max_pool2d_cpu_float32 (__main__.TestEagerFusionOpInfoCPU)": 67.06866836547852, + "test_aot_autograd_symbolic_exhaustive_linalg_svd_cpu_float32 (__main__.TestEagerFusionOpInfoCPU)": 60.106476193382626, + "test_aot_autograd_symbolic_exhaustive_nn_functional_max_pool1d_cpu_float32 (__main__.TestEagerFusionOpInfoCPU)": 92.41733296712239, + "test_aot_autograd_symbolic_exhaustive_nn_functional_max_pool2d_cpu_float32 (__main__.TestEagerFusionOpInfoCPU)": 153.84766642252603, + "test_aot_autograd_symbolic_exhaustive_nn_functional_max_pool3d_cpu_float32 (__main__.TestEagerFusionOpInfoCPU)": 85.03233337402344, + "test_aot_autograd_symbolic_module_exhaustive_nn_TransformerDecoderLayer_cpu_float32 (__main__.TestEagerFusionModuleInfoCPU)": 84.20766703287761, + "test_avg_pool3d_backward2_cpu (__main__.CpuTests)": 510.78777398003473, + "test_avg_pool3d_backward2_cuda (__main__.GPUTests)": 97.41633224487305, + "test_avg_pool3d_backward2_dynamic_shapes_cpu (__main__.DynamicShapesCodegenCpuTests)": 510.7107747395833, + "test_avg_pool3d_backward2_dynamic_shapes_cpu (__main__.DynamicShapesCpuTests)": 507.0798814561632, + "test_avg_pool3d_backward2_dynamic_shapes_cuda (__main__.DynamicShapesGPUTests)": 124.66699981689453, + "test_avg_pool3d_backward_cpu (__main__.CpuHalideTests)": 62.02766672770182, + "test_backward_nn_functional_multi_head_attention_forward_cpu_float32 (__main__.TestCompositeComplianceCPU)": 65.31933339436848, + "test_backward_nn_functional_multi_head_attention_forward_cuda_float32 (__main__.TestCompositeComplianceCUDA)": 64.8836669921875, + "test_basic_cpu (__main__.EfficientConvBNEvalCpuTests)": 279.31766594780817, + "test_basic_cuda (__main__.EfficientConvBNEvalGpuTests)": 196.89432907104492, + "test_builtin_equivalent_funcs (__main__.TorchFunctionModeTests)": 104.88957105364118, + "test_checkpoint_cast (__main__.TestFxToOnnx)": 363.49000040690106, + "test_collect_callgrind (__main__.TestBenchmarkUtils)": 374.91833157009546, + "test_comprehensive_constant_pad_nd_cpu_float16 (__main__.TestInductorOpInfoCPU)": 82.49566650390625, + "test_comprehensive_constant_pad_nd_cpu_float32 (__main__.TestInductorOpInfoCPU)": 82.13433329264323, + "test_comprehensive_constant_pad_nd_cpu_float64 (__main__.TestInductorOpInfoCPU)": 81.46333312988281, + "test_comprehensive_constant_pad_nd_cpu_int32 (__main__.TestInductorOpInfoCPU)": 86.26499938964844, + "test_comprehensive_constant_pad_nd_cpu_int64 (__main__.TestInductorOpInfoCPU)": 78.45066579182942, + "test_comprehensive_diff_cpu_bool (__main__.TestInductorOpInfoCPU)": 120.62533315022786, + "test_comprehensive_diff_cpu_float32 (__main__.TestInductorOpInfoCPU)": 120.60333251953125, + "test_comprehensive_diff_cpu_float64 (__main__.TestInductorOpInfoCPU)": 112.93033345540364, + "test_comprehensive_diff_cpu_int32 (__main__.TestInductorOpInfoCPU)": 118.08499908447266, + "test_comprehensive_diff_cpu_int64 (__main__.TestInductorOpInfoCPU)": 123.97566731770833, + "test_comprehensive_diff_cuda_complex128 (__main__.TestDecompCUDA)": 81.66700108846028, + "test_comprehensive_diff_cuda_complex64 (__main__.TestDecompCUDA)": 79.4769999186198, + "test_comprehensive_diff_cuda_float64 (__main__.TestDecompCUDA)": 64.5281252861023, + "test_comprehensive_dist_cpu_float16 (__main__.TestInductorOpInfoCPU)": 96.58866628011067, + "test_comprehensive_dist_cpu_float32 (__main__.TestInductorOpInfoCPU)": 91.72833251953125, + "test_comprehensive_dist_cpu_float64 (__main__.TestInductorOpInfoCPU)": 93.73600006103516, + "test_comprehensive_eye_cpu_bool (__main__.TestInductorOpInfoCPU)": 151.32967122395834, + "test_comprehensive_eye_cpu_float16 (__main__.TestInductorOpInfoCPU)": 143.57733662923178, + "test_comprehensive_eye_cpu_float32 (__main__.TestInductorOpInfoCPU)": 141.9923299153646, + "test_comprehensive_eye_cpu_float64 (__main__.TestInductorOpInfoCPU)": 142.34900410970053, + "test_comprehensive_eye_cpu_int32 (__main__.TestInductorOpInfoCPU)": 143.60033671061197, + "test_comprehensive_eye_cpu_int64 (__main__.TestInductorOpInfoCPU)": 149.10132853190103, + "test_comprehensive_grid_sampler_2d_cpu_bfloat16 (__main__.TestDecompCPU)": 81.41033426920573, + "test_comprehensive_grid_sampler_2d_cpu_float16 (__main__.TestDecompCPU)": 81.97200012207031, + "test_comprehensive_grid_sampler_2d_cpu_float16 (__main__.TestInductorOpInfoCPU)": 84.08200073242188, + "test_comprehensive_grid_sampler_2d_cpu_float32 (__main__.TestDecompCPU)": 347.4583333333333, + "test_comprehensive_grid_sampler_2d_cpu_float32 (__main__.TestInductorOpInfoCPU)": 83.02366638183594, + "test_comprehensive_grid_sampler_2d_cpu_float64 (__main__.TestDecompCPU)": 328.58299763997394, + "test_comprehensive_grid_sampler_2d_cpu_float64 (__main__.TestInductorOpInfoCPU)": 92.41300201416016, + "test_comprehensive_grid_sampler_2d_cuda_bfloat16 (__main__.TestDecompCUDA)": 206.7806625366211, + "test_comprehensive_grid_sampler_2d_cuda_float16 (__main__.TestDecompCUDA)": 201.07150268554688, + "test_comprehensive_grid_sampler_2d_cuda_float32 (__main__.TestDecompCUDA)": 838.8346659342448, + "test_comprehensive_grid_sampler_2d_cuda_float64 (__main__.TestDecompCUDA)": 846.4761555989584, + "test_comprehensive_linalg_lu_solve_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 87.954833984375, + "test_comprehensive_linalg_lu_solve_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 71.02933311462402, + "test_comprehensive_linalg_norm_cpu_float32 (__main__.TestInductorOpInfoCPU)": 64.14100011189778, + "test_comprehensive_linalg_norm_cpu_float64 (__main__.TestInductorOpInfoCPU)": 63.77633412679037, + "test_comprehensive_linalg_solve_triangular_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 71.52133178710938, + "test_comprehensive_linalg_solve_triangular_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 64.97216606140137, + "test_comprehensive_linalg_svd_cuda_complex128 (__main__.TestDecompCUDA)": 65.3273328145345, + "test_comprehensive_linalg_vector_norm_cpu_float16 (__main__.TestInductorOpInfoCPU)": 227.3056640625, + "test_comprehensive_linalg_vector_norm_cpu_float32 (__main__.TestInductorOpInfoCPU)": 217.87433369954428, + "test_comprehensive_linalg_vector_norm_cpu_float64 (__main__.TestInductorOpInfoCPU)": 235.36133321126303, + "test_comprehensive_linalg_vector_norm_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 69.16516621907552, + "test_comprehensive_linalg_vector_norm_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 70.0791670481364, + "test_comprehensive_logspace_cpu_float32 (__main__.TestInductorOpInfoCPU)": 471.6073303222656, + "test_comprehensive_logspace_cpu_float64 (__main__.TestInductorOpInfoCPU)": 482.71632893880206, + "test_comprehensive_logspace_cpu_int32 (__main__.TestInductorOpInfoCPU)": 484.3889973958333, + "test_comprehensive_logspace_cpu_int64 (__main__.TestInductorOpInfoCPU)": 478.03033447265625, + "test_comprehensive_logspace_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 83.64666748046875, + "test_comprehensive_logspace_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 81.86583391825359, + "test_comprehensive_masked_amax_cpu_float16 (__main__.TestInductorOpInfoCPU)": 109.95899963378906, + "test_comprehensive_masked_amax_cpu_float32 (__main__.TestInductorOpInfoCPU)": 107.42000071207683, + "test_comprehensive_masked_amax_cpu_float64 (__main__.TestInductorOpInfoCPU)": 112.86100006103516, + "test_comprehensive_masked_amax_cpu_int32 (__main__.TestInductorOpInfoCPU)": 111.37566630045573, + "test_comprehensive_masked_amax_cpu_int64 (__main__.TestInductorOpInfoCPU)": 106.12166849772136, + "test_comprehensive_masked_amin_cpu_float16 (__main__.TestInductorOpInfoCPU)": 110.41200002034505, + "test_comprehensive_masked_amin_cpu_float32 (__main__.TestInductorOpInfoCPU)": 107.4979985555013, + "test_comprehensive_masked_amin_cpu_float64 (__main__.TestInductorOpInfoCPU)": 108.29533386230469, + "test_comprehensive_masked_amin_cpu_int32 (__main__.TestInductorOpInfoCPU)": 101.06366729736328, + "test_comprehensive_masked_amin_cpu_int64 (__main__.TestInductorOpInfoCPU)": 111.66866556803386, + "test_comprehensive_masked_argmax_cpu_float16 (__main__.TestInductorOpInfoCPU)": 65.20666631062825, + "test_comprehensive_masked_argmax_cpu_float32 (__main__.TestInductorOpInfoCPU)": 66.2469991048177, + "test_comprehensive_masked_argmax_cpu_float64 (__main__.TestInductorOpInfoCPU)": 62.26433308919271, + "test_comprehensive_masked_argmax_cpu_int32 (__main__.TestInductorOpInfoCPU)": 65.90833282470703, + "test_comprehensive_masked_argmin_cpu_float16 (__main__.TestInductorOpInfoCPU)": 67.25066884358723, + "test_comprehensive_masked_argmin_cpu_float32 (__main__.TestInductorOpInfoCPU)": 63.0856679280599, + "test_comprehensive_masked_argmin_cpu_float64 (__main__.TestInductorOpInfoCPU)": 64.3423334757487, + "test_comprehensive_masked_argmin_cpu_int32 (__main__.TestInductorOpInfoCPU)": 62.418999989827476, + "test_comprehensive_masked_argmin_cpu_int64 (__main__.TestInductorOpInfoCPU)": 61.3716672261556, + "test_comprehensive_masked_mean_cpu_bool (__main__.TestInductorOpInfoCPU)": 107.6316655476888, + "test_comprehensive_masked_mean_cpu_float16 (__main__.TestInductorOpInfoCPU)": 105.95433298746745, + "test_comprehensive_masked_mean_cpu_float32 (__main__.TestInductorOpInfoCPU)": 108.12066650390625, + "test_comprehensive_masked_mean_cpu_float64 (__main__.TestInductorOpInfoCPU)": 107.28466542561848, + "test_comprehensive_masked_mean_cpu_int32 (__main__.TestInductorOpInfoCPU)": 107.82733408610027, + "test_comprehensive_masked_mean_cpu_int64 (__main__.TestInductorOpInfoCPU)": 103.07533264160156, + "test_comprehensive_masked_norm_cpu_float16 (__main__.TestInductorOpInfoCPU)": 530.9063313802084, + "test_comprehensive_masked_norm_cpu_float32 (__main__.TestInductorOpInfoCPU)": 531.5923258463541, + "test_comprehensive_masked_norm_cpu_float64 (__main__.TestInductorOpInfoCPU)": 522.8729858398438, + "test_comprehensive_masked_norm_cuda_float16 (__main__.TestInductorOpInfoCUDA)": 170.60566584269205, + "test_comprehensive_masked_norm_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 148.06033198038736, + "test_comprehensive_masked_norm_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 149.73183059692383, + "test_comprehensive_masked_prod_cpu_bool (__main__.TestInductorOpInfoCPU)": 106.41500091552734, + "test_comprehensive_masked_prod_cpu_float16 (__main__.TestInductorOpInfoCPU)": 109.46266428629558, + "test_comprehensive_masked_prod_cpu_float32 (__main__.TestInductorOpInfoCPU)": 108.69433339436848, + "test_comprehensive_masked_prod_cpu_float64 (__main__.TestInductorOpInfoCPU)": 111.31100209554036, + "test_comprehensive_masked_prod_cpu_int32 (__main__.TestInductorOpInfoCPU)": 103.59366607666016, + "test_comprehensive_masked_prod_cpu_int64 (__main__.TestInductorOpInfoCPU)": 98.89466603597005, + "test_comprehensive_masked_sum_cpu_bool (__main__.TestInductorOpInfoCPU)": 111.9769999186198, + "test_comprehensive_masked_sum_cpu_float16 (__main__.TestInductorOpInfoCPU)": 113.90066782633464, + "test_comprehensive_masked_sum_cpu_float32 (__main__.TestInductorOpInfoCPU)": 105.57700093587239, + "test_comprehensive_masked_sum_cpu_float64 (__main__.TestInductorOpInfoCPU)": 105.22766621907552, + "test_comprehensive_masked_sum_cpu_int32 (__main__.TestInductorOpInfoCPU)": 104.93033345540364, + "test_comprehensive_masked_sum_cpu_int64 (__main__.TestInductorOpInfoCPU)": 102.3076680501302, + "test_comprehensive_nn_functional_gaussian_nll_loss_cuda_float32 (__main__.TestDecompCUDA)": 78.71599960327148, + "test_comprehensive_nn_functional_gaussian_nll_loss_cuda_float64 (__main__.TestDecompCUDA)": 81.63083330790202, + "test_comprehensive_nn_functional_glu_cpu_float16 (__main__.TestInductorOpInfoCPU)": 83.04366556803386, + "test_comprehensive_nn_functional_glu_cpu_float32 (__main__.TestInductorOpInfoCPU)": 82.60833231608073, + "test_comprehensive_nn_functional_glu_cpu_float64 (__main__.TestInductorOpInfoCPU)": 82.17033386230469, + "test_comprehensive_nn_functional_grid_sample_cpu_float32 (__main__.TestDecompCPU)": 86.39266713460286, + "test_comprehensive_nn_functional_grid_sample_cpu_float64 (__main__.TestDecompCPU)": 87.72166442871094, + "test_comprehensive_nn_functional_grid_sample_cuda_float32 (__main__.TestDecompCUDA)": 179.57733662923178, + "test_comprehensive_nn_functional_grid_sample_cuda_float64 (__main__.TestDecompCUDA)": 199.76399993896484, + "test_comprehensive_nn_functional_interpolate_bicubic_cpu_uint8 (__main__.TestInductorOpInfoCPU)": 63.46200052897135, + "test_comprehensive_nn_functional_interpolate_bicubic_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 78.04750061035156, + "test_comprehensive_nn_functional_interpolate_bicubic_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 75.1818339029948, + "test_comprehensive_nn_functional_interpolate_trilinear_cuda_float32 (__main__.TestDecompCUDA)": 63.27016639709473, + "test_comprehensive_nn_functional_interpolate_trilinear_cuda_float64 (__main__.TestDecompCUDA)": 80.72016716003418, + "test_comprehensive_nn_functional_max_pool1d_cpu_float16 (__main__.TestInductorOpInfoCPU)": 188.21700032552084, + "test_comprehensive_nn_functional_max_pool1d_cpu_float32 (__main__.TestInductorOpInfoCPU)": 189.15300496419272, + "test_comprehensive_nn_functional_max_pool1d_cpu_float64 (__main__.TestInductorOpInfoCPU)": 191.53266398111978, + "test_comprehensive_nn_functional_max_pool2d_cpu_float16 (__main__.TestInductorOpInfoCPU)": 1001.8180135091146, + "test_comprehensive_nn_functional_max_pool2d_cpu_float32 (__main__.TestInductorOpInfoCPU)": 941.7516682942709, + "test_comprehensive_nn_functional_max_pool2d_cpu_float64 (__main__.TestInductorOpInfoCPU)": 957.2676798502604, + "test_comprehensive_nn_functional_max_pool2d_cpu_int32 (__main__.TestInductorOpInfoCPU)": 884.413330078125, + "test_comprehensive_nn_functional_max_pool2d_cpu_int64 (__main__.TestInductorOpInfoCPU)": 917.0406697591146, + "test_comprehensive_nn_functional_max_pool2d_cuda_float16 (__main__.TestInductorOpInfoCUDA)": 882.3416646321615, + "test_comprehensive_nn_functional_max_pool2d_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 891.5490112304688, + "test_comprehensive_nn_functional_max_pool2d_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 951.3874918619791, + "test_comprehensive_nn_functional_max_unpool2d_cpu_float16 (__main__.TestInductorOpInfoCPU)": 217.1499989827474, + "test_comprehensive_nn_functional_max_unpool2d_cpu_float32 (__main__.TestInductorOpInfoCPU)": 226.78500366210938, + "test_comprehensive_nn_functional_max_unpool2d_cpu_float64 (__main__.TestInductorOpInfoCPU)": 220.6536661783854, + "test_comprehensive_nn_functional_max_unpool2d_cuda_float16 (__main__.TestInductorOpInfoCUDA)": 62.534000396728516, + "test_comprehensive_nn_functional_max_unpool2d_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 115.53050104777019, + "test_comprehensive_nn_functional_max_unpool2d_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 111.29116821289062, + "test_comprehensive_nn_functional_max_unpool3d_cpu_float16 (__main__.TestInductorOpInfoCPU)": 147.32699584960938, + "test_comprehensive_nn_functional_max_unpool3d_cpu_float32 (__main__.TestInductorOpInfoCPU)": 146.26766459147134, + "test_comprehensive_nn_functional_max_unpool3d_cpu_float64 (__main__.TestInductorOpInfoCPU)": 149.46000162760416, + "test_comprehensive_nn_functional_pad_constant_cpu_float16 (__main__.TestInductorOpInfoCPU)": 80.33966827392578, + "test_comprehensive_nn_functional_pad_constant_cpu_float32 (__main__.TestInductorOpInfoCPU)": 77.02700297037761, + "test_comprehensive_nn_functional_pad_constant_cpu_float64 (__main__.TestInductorOpInfoCPU)": 78.06666564941406, + "test_comprehensive_nn_functional_pad_constant_cpu_int32 (__main__.TestInductorOpInfoCPU)": 66.0947225358751, + "test_comprehensive_nn_functional_pad_constant_cpu_int64 (__main__.TestInductorOpInfoCPU)": 79.26499938964844, + "test_comprehensive_nn_functional_poisson_nll_loss_cpu_float16 (__main__.TestInductorOpInfoCPU)": 141.70166524251303, + "test_comprehensive_nn_functional_poisson_nll_loss_cpu_float32 (__main__.TestInductorOpInfoCPU)": 142.16866811116537, + "test_comprehensive_nn_functional_poisson_nll_loss_cpu_float64 (__main__.TestInductorOpInfoCPU)": 128.4029998779297, + "test_comprehensive_nn_functional_poisson_nll_loss_cpu_int32 (__main__.TestInductorOpInfoCPU)": 124.0116678873698, + "test_comprehensive_nn_functional_poisson_nll_loss_cpu_int64 (__main__.TestInductorOpInfoCPU)": 126.01200103759766, + "test_comprehensive_nn_functional_unfold_cpu_bool (__main__.TestInductorOpInfoCPU)": 151.81866963704428, + "test_comprehensive_nn_functional_unfold_cpu_float16 (__main__.TestInductorOpInfoCPU)": 264.4179992675781, + "test_comprehensive_nn_functional_unfold_cpu_float32 (__main__.TestInductorOpInfoCPU)": 267.08266194661456, + "test_comprehensive_nn_functional_unfold_cpu_float64 (__main__.TestInductorOpInfoCPU)": 265.7936604817708, + "test_comprehensive_nn_functional_unfold_cuda_float16 (__main__.TestInductorOpInfoCUDA)": 85.77500025431316, + "test_comprehensive_nn_functional_unfold_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 78.86816660563152, + "test_comprehensive_nn_functional_unfold_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 79.76916758219402, + "test_comprehensive_ormqr_cuda_complex128 (__main__.TestDecompCUDA)": 95.18166605631511, + "test_comprehensive_ormqr_cuda_complex64 (__main__.TestDecompCUDA)": 91.77883275349934, + "test_comprehensive_ormqr_cuda_float32 (__main__.TestDecompCUDA)": 60.81966654459635, + "test_comprehensive_ormqr_cuda_float32 (__main__.TestInductorOpInfoCUDA)": 119.70216623942058, + "test_comprehensive_ormqr_cuda_float64 (__main__.TestInductorOpInfoCUDA)": 93.25683466593425, + "test_comprehensive_svd_cuda_complex128 (__main__.TestDecompCUDA)": 70.70699964250836, + "test_comprehensive_svd_lowrank_cuda_complex128 (__main__.TestDecompCUDA)": 61.070499420166016, + "test_cond_autograd_nested (__main__.TestControlFlow)": 79.72944344414606, + "test_constructor_autograd_SparseBSC_cuda (__main__.TestSparseAnyCUDA)": 117.21133550008138, + "test_constructor_autograd_SparseBSR_cuda (__main__.TestSparseAnyCUDA)": 95.2274996439616, + "test_constructor_autograd_SparseCSC_cuda (__main__.TestSparseAnyCUDA)": 77.67016474405925, + "test_constructor_autograd_SparseCSR_cuda (__main__.TestSparseAnyCUDA)": 69.53000020980835, + "test_conv1d_basic (__main__.TestXNNPACKConv1dTransformPass)": 198.49655405680338, + "test_conv1d_with_relu_fc (__main__.TestXNNPACKConv1dTransformPass)": 447.2755482991536, + "test_conv2d_unary_cpu_cpp_wrapper (__main__.TestCppWrapper)": 268.80133056640625, + "test_conv3d_cuda (__main__.AOTInductorTestABICompatibleGpu)": 135.77333323160806, + "test_conv_bn_fuse_dynamic_shapes_cpu (__main__.DynamicShapesCpuTests)": 60.003444247775604, + "test_conv_transpose_with_output_size_and_no_batch_dim_ConvTranspose3d_cuda (__main__.TestConvolutionNNDeviceTypeCUDA)": 114.27466656515996, + "test_correctness_AdamW_use_closure_False_cuda_float32 (__main__.CompiledOptimizerParityTestsCUDA)": 72.10799980163574, + "test_correctness_AdamW_use_closure_True_cuda_float32 (__main__.CompiledOptimizerParityTestsCUDA)": 123.51449966430664, + "test_correctness_Adam_use_closure_False_cuda_float32 (__main__.CompiledOptimizerParityTestsCUDA)": 67.44266573588054, + "test_correctness_Adam_use_closure_True_cuda_float32 (__main__.CompiledOptimizerParityTestsCUDA)": 124.48583348592122, + "test_correctness_NAdam_use_closure_True_cuda_float32 (__main__.CompiledOptimizerParityTestsCUDA)": 72.73799896240234, + "test_count_nonzero_all (__main__.TestBool)": 613.8192070855034, + "test_custom_module_lstm (__main__.TestQuantizedOps)": 387.75410630967883, + "test_ddp_uneven_inputs (__main__.TestDistBackendWithSpawn)": 186.20633189876875, + "test_dispatch_symbolic_meta_outplace_all_strides_nn_functional_gaussian_nll_loss_cuda_float32 (__main__.TestMetaCUDA)": 84.25950113932292, + "test_eig_check_magma_cuda_float32 (__main__.TestLinalgCUDA)": 152.92416445414224, + "test_fail_creation_ops.py (__main__.TestTyping)": 75.64839254106793, + "test_fn_fwgrad_bwgrad_cumprod_cuda_complex128 (__main__.TestFwdGradientsCUDA)": 70.79500071207683, + "test_fn_gradgrad_cumprod_cuda_complex128 (__main__.TestBwdGradientsCUDA)": 79.6708329518636, + "test_fn_gradgrad_map_nested_cpu_float64 (__main__.TestBwdGradientsCPU)": 77.31333414713542, + "test_fn_gradgrad_map_nested_cuda_float64 (__main__.TestBwdGradientsCUDA)": 60.696099281311035, + "test_fn_gradgrad_map_triple_nested_cpu_float64 (__main__.TestBwdGradientsCPU)": 461.1873270670573, + "test_fn_gradgrad_map_triple_nested_cuda_float64 (__main__.TestBwdGradientsCUDA)": 339.5521697998047, + "test_fuse_large_params_cpu (__main__.CpuTests)": 63.09900029500326, + "test_fuse_large_params_dynamic_shapes_cpu (__main__.DynamicShapesCodegenCpuTests)": 92.18155585394965, + "test_fuse_large_params_dynamic_shapes_cpu (__main__.DynamicShapesCpuTests)": 96.59077707926433, + "test_fuse_large_params_dynamic_shapes_cuda (__main__.DynamicShapesCodegenGPUTests)": 63.364166259765625, + "test_fuse_large_params_dynamic_shapes_cuda (__main__.DynamicShapesGPUTests)": 67.6820011138916, + "test_grad_nn_Transformer_cuda_float64 (__main__.TestModuleCUDA)": 80.87316767374675, + "test_gradgrad_nn_LSTM_eval_mode_cuda_float64 (__main__.TestModuleCUDA)": 99.19066619873047, + "test_gradgrad_nn_LSTM_train_mode_cuda_float64 (__main__.TestModuleCUDA)": 95.8260014851888, + "test_gradgrad_nn_TransformerDecoderLayer_cuda_float64 (__main__.TestModuleCUDA)": 184.47216796875, + "test_gradgrad_nn_TransformerEncoder_eval_mode_cuda_float64 (__main__.TestModuleCUDA)": 113.43166605631511, + "test_gradgrad_nn_TransformerEncoder_train_mode_cuda_float64 (__main__.TestModuleCUDA)": 125.3043327331543, + "test_grid_sampler_2d_cpu (__main__.CpuHalideTests)": 186.22100321451822, + "test_indexing (__main__.TestAutogradWithCompiledAutograd)": 70.04255633884006, + "test_indirect_device_assert (__main__.TritonCodeGenTests)": 189.80333455403647, + "test_inductor_no_recursionerror_on_for_loops_dynamic_shapes (__main__.DynamicShapesReproTests)": 72.33966615464952, + "test_inplace_gradgrad_cumprod_cuda_complex128 (__main__.TestBwdGradientsCUDA)": 80.61100006103516, + "test_inputs_overlapping_with_mutation_stress_dynamic_shapes (__main__.DynamicShapesAotAutogradFallbackTests)": 140.90377638075086, + "test_jit_cuda_archflags (__main__.TestCppExtensionJIT)": 111.77066548665364, + "test_linalg_solve_triangular_large_cuda_complex128 (__main__.TestLinalgCUDA)": 565.1104965209961, + "test_linalg_solve_triangular_large_cuda_complex64 (__main__.TestLinalgCUDA)": 71.39283434549968, + "test_linalg_solve_triangular_large_cuda_float64 (__main__.TestLinalgCUDA)": 65.84149837493896, + "test_linear (__main__.TestStaticQuantizedModule)": 147.759886847602, + "test_linear_binary_cpp_wrapper (__main__.TestCppWrapper)": 164.447998046875, + "test_linear_binary_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 167.55933125813803, + "test_linear_binary_folding_cpu (__main__.FreezingCpuTests)": 60.865777757432724, + "test_linear_packed_cpp_wrapper (__main__.TestCppWrapper)": 81.31299845377605, + "test_linear_packed_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 76.33000183105469, + "test_lobpcg_ortho_cuda_float64 (__main__.TestLinalgCUDA)": 76.25166829427083, + "test_lstm_cpu (__main__.TestMkldnnCPU)": 84.14966710408528, + "test_many_overlapping_inputs_does_not_explode_guards_dynamic_shapes (__main__.DynamicShapesReproTests)": 109.88477749294705, + "test_max_autotune_cutlass_backend_addmm_dynamic_False_max_autotune_gemm_backends_ATen,Triton,CUTLASS (__main__.TestCutlassBackend)": 84.41333262125652, + "test_max_pool2d_with_indices_backward4_dynamic_shapes_cpu (__main__.DynamicShapesCodegenCpuTests)": 64.33055538601346, + "test_max_pool2d_with_indices_backward4_dynamic_shapes_cpu (__main__.DynamicShapesCpuTests)": 63.2979990641276, + "test_mixed_mm_exhaustive_dtypes (__main__.TestPatternMatcher)": 67.55066744486491, + "test_proper_exit (__main__.TestDataLoader)": 224.79383595784506, + "test_proper_exit (__main__.TestDataLoaderPersistentWorkers)": 217.4080022176107, + "test_put_cuda_uint8 (__main__.TestTorchDeviceTypeCUDA)": 103.19804314888704, + "test_python_ref_executor__refs_special_zeta_executor_aten_cuda_float64 (__main__.TestCommonCUDA)": 60.95939979553223, + "test_qconv2d_add_cpu_cpp_wrapper (__main__.TestCppWrapper)": 62.51866658528646, + "test_qconv2d_add_cpu_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 63.23300043741862, + "test_qconv2d_add_relu_cpu_cpp_wrapper (__main__.TestCppWrapper)": 62.45500055948893, + "test_qconv2d_add_relu_cpu_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 63.74866739908854, + "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_False_is_dynamic_False (__main__.TestPatternMatcher)": 73.47400156656902, + "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_False_is_dynamic_False_cpp_wrapper (__main__.TestCppWrapper)": 111.55533345540364, + "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_False_is_dynamic_False_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 113.39233144124348, + "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_False_is_dynamic_True (__main__.TestPatternMatcher)": 71.40777672661676, + "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_False_is_dynamic_True_cpp_wrapper (__main__.TestCppWrapper)": 114.24800109863281, + "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_False_is_dynamic_True_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 113.41766611735027, + "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_True_is_dynamic_False (__main__.TestPatternMatcher)": 65.87688954671223, + "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_True_is_dynamic_False_cpp_wrapper (__main__.TestCppWrapper)": 107.47800191243489, + "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_True_is_dynamic_False_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 113.447998046875, + "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_True_is_dynamic_True (__main__.TestPatternMatcher)": 68.04311116536458, + "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_True_is_dynamic_True_cpp_wrapper (__main__.TestCppWrapper)": 115.14266459147136, + "test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_True_is_dynamic_True_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 115.07366434733073, + "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_False_is_dynamic_False (__main__.TestPatternMatcher)": 76.1270014444987, + "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_False_is_dynamic_False_cpp_wrapper (__main__.TestCppWrapper)": 113.39599863688152, + "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_False_is_dynamic_False_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 114.23433176676433, + "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_False_is_dynamic_True (__main__.TestPatternMatcher)": 71.46933237711589, + "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_False_is_dynamic_True_cpp_wrapper (__main__.TestCppWrapper)": 113.58066813151042, + "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_False_is_dynamic_True_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 110.88400014241536, + "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_True_is_dynamic_False (__main__.TestPatternMatcher)": 65.69477844238281, + "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_True_is_dynamic_False_cpp_wrapper (__main__.TestCppWrapper)": 107.49266560872395, + "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_True_is_dynamic_False_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 114.12266540527344, + "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_True_is_dynamic_True (__main__.TestPatternMatcher)": 75.25133514404297, + "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_True_is_dynamic_True_cpp_wrapper (__main__.TestCppWrapper)": 115.43599955240886, + "test_qlinear_add_int8_mixed_bf16_use_relu_True_is_qat_True_is_dynamic_True_dynamic_shapes_cpp_wrapper (__main__.DynamicShapesCppWrapperCpuTests)": 113.93733215332031, + "test_quick_core_backward__unsafe_masked_index_cpu_float64 (__main__.TestDecompCPU)": 319.4066670735677, + "test_quick_core_backward__unsafe_masked_index_cuda_float64 (__main__.TestDecompCUDA)": 617.1271667480469, + "test_quick_core_backward__unsafe_masked_index_put_accumulate_cpu_float64 (__main__.TestDecompCPU)": 555.4753214518229, + "test_quick_core_backward__unsafe_masked_index_put_accumulate_cuda_float64 (__main__.TestDecompCUDA)": 895.808827718099, + "test_quick_core_backward_nn_functional_max_unpool3d_grad_cpu_float64 (__main__.TestDecompCPU)": 71.04066721598308, + "test_quick_core_backward_nn_functional_max_unpool3d_grad_cuda_float64 (__main__.TestDecompCUDA)": 186.27233378092447, + "test_quick_core_backward_roll_cpu_float64 (__main__.TestDecompCPU)": 93.46799977620442, + "test_quick_core_backward_roll_cuda_float64 (__main__.TestDecompCUDA)": 146.96949895222983, + "test_quick_core_backward_select_scatter_cuda_float64 (__main__.TestDecompCUDA)": 101.87816492716472, + "test_quick_core_backward_split_cuda_float64 (__main__.TestDecompCUDA)": 60.058855874197825, + "test_quick_core_backward_split_with_sizes_copy_cpu_float64 (__main__.TestDecompCPU)": 74.16299947102864, + "test_quick_core_backward_split_with_sizes_copy_cuda_float64 (__main__.TestDecompCUDA)": 117.55266952514648, + "test_quick_core_backward_std_cuda_float64 (__main__.TestDecompCUDA)": 80.77400080362956, + "test_replicatepad_64bit_indexing_cuda_float16 (__main__.TestNNDeviceTypeCUDA)": 332.54766845703125, + "test_rosenbrock_sparse_with_lrsched_False_SGD_cuda_float64 (__main__.TestOptimRenewedCUDA)": 90.56116612752278, + "test_rosenbrock_sparse_with_lrsched_True_SGD_cuda_float64 (__main__.TestOptimRenewedCUDA)": 102.87433195114136, + "test_save_load_large_string_attribute (__main__.TestSaveLoad)": 114.60433197021484, + "test_shuffler_iterdatapipe (__main__.IntegrationTestDataLoaderDataPipe)": 163.62533145480685, + "test_slow_tasks (__main__.TestFunctionalAutogradBenchmark)": 107.40444437662761, + "test_sum_all_cpu_float64 (__main__.TestReductionsCPU)": 284.4367377569002, + "test_svd_lowrank_cuda_complex128 (__main__.TestLinalgCUDA)": 157.5870008468628, + "test_terminate_handler_on_crash (__main__.TestTorch)": 97.66699902216594, + "test_terminate_signal (__main__.ForkTest)": 130.4108906057146, + "test_terminate_signal (__main__.ParallelForkServerShouldWorkTest)": 130.55344214538732, + "test_terminate_signal (__main__.SpawnTest)": 133.65200106302896, + "test_train_parity_multi_group (__main__.TestFullyShard1DTrainingCore)": 117.12999943544855, + "test_transpose_copy (__main__.CPUReproTests)": 75.40855577256944, + "test_triton_bsr_scatter_mm_blocksize_64_cuda_bfloat16 (__main__.TestSparseCompressedTritonKernelsCUDA)": 63.13999875386556, + "test_triton_bsr_scatter_mm_blocksize_64_cuda_float32 (__main__.TestSparseCompressedTritonKernelsCUDA)": 66.69533348083496, + "test_triton_bsr_softmax_cuda_bfloat16 (__main__.TestSparseCompressedTritonKernelsCUDA)": 160.92549896240234, + "test_triton_bsr_softmax_cuda_float16 (__main__.TestSparseCompressedTritonKernelsCUDA)": 147.94483184814453, + "test_triton_bsr_softmax_cuda_float32 (__main__.TestSparseCompressedTritonKernelsCUDA)": 119.7986691792806, + "test_unary_ops (__main__.TestTEFuserDynamic)": 190.41244761149088, + "test_unary_ops (__main__.TestTEFuserStatic)": 165.1259978612264, + "test_upsample_bicubic2d_cpu (__main__.CpuHalideTests)": 96.73766581217448, + "test_variant_consistency_jit_nn_functional_max_pool2d_cpu_float32 (__main__.TestJitCPU)": 88.42266591389973, + "test_variant_consistency_jit_nn_functional_max_pool2d_cuda_float32 (__main__.TestJitCUDA)": 71.24600028991699, + "test_vec_bitwise (__main__.CPUReproTests)": 63.55355538262261, + "test_vec_compare_op_cpu_only (__main__.CPUReproTests)": 66.54970640294692, + "test_views1_dynamic_shapes_cpu (__main__.DynamicShapesCpuTests)": 61.20033391316732, + "test_vmapjvpvjp_linalg_lu_solve_cpu_float32 (__main__.TestOperatorsCPU)": 60.52709543137323, + "test_vmapjvpvjp_linalg_lu_solve_cuda_float32 (__main__.TestOperatorsCUDA)": 70.6786657969157, + "test_vmapjvpvjp_linalg_multi_dot_cuda_float32 (__main__.TestOperatorsCUDA)": 72.03637456893921, + "test_vmapjvpvjp_linalg_solve_triangular_cuda_float32 (__main__.TestOperatorsCUDA)": 68.06242861066546, + "test_vmapjvpvjp_linalg_svd_cuda_float32 (__main__.TestOperatorsCUDA)": 62.29850069681803, + "test_vmapjvpvjp_max_pool2d_with_indices_backward_cpu_float32 (__main__.TestOperatorsCPU)": 99.96300252278645, + "test_vmapjvpvjp_max_pool2d_with_indices_backward_cuda_float32 (__main__.TestOperatorsCUDA)": 73.61033248901367, + "test_vmapjvpvjp_nn_functional_max_pool2d_cpu_float32 (__main__.TestOperatorsCPU)": 69.7383321126302, + "test_vmapjvpvjp_nn_functional_max_pool2d_cuda_float32 (__main__.TestOperatorsCUDA)": 67.93116633097331, + "test_vmapjvpvjp_svd_cuda_float32 (__main__.TestOperatorsCUDA)": 67.69150098164876, + "test_vmapjvpvjp_unbind_cuda_float32 (__main__.TestOperatorsCUDA)": 75.51783498128255, + "test_vmapvjpvjp_meshgrid_list_of_tensors_cuda_float32 (__main__.TestOperatorsCUDA)": 71.56416575113933, + "test_vmapvjpvjp_meshgrid_variadic_tensors_cuda_float32 (__main__.TestOperatorsCUDA)": 71.6493345896403, + "test_vmapvjpvjp_nn_functional_bilinear_cuda_float32 (__main__.TestOperatorsCUDA)": 132.8029988606771 } \ No newline at end of file