# t/rank0.1790655703137742536.pt.trace.json.gz
kernels 519345 busy_ms 8248.5 window_ms 8327.6
eager kernel busy_ms 3231.8
graph replays by node count: n_nodes count mean_busy_us mean_span_us
    2717   142    30439.7    30593.8
     165   152      697.5      709.7
     132   144     2349.2     2355.9
     101   148     1569.1     1573.8
     137     8     2219.2     2225.8
top kernels: total_ms count mean_us name
    1894.01   20448    92.63 void exl3_gemm_kernel<4, false, 2, 16, 32, 256, 4, 3>(__half const*, unsigned short const*, void*, int, int, i
    1804.67    3312   544.89 void exl3_i8_gemm_kernel<4, 2, 128, 128, 128, 1, 4, 2, 8, false, false, true, false, true, true, true, 128, 2,
    1420.06   35840    39.62 void exl3_gemm_kernel<4, false, 2, 32, 32, 128, 4, 3>(__half const*, unsigned short const*, void*, int, int, i
     468.99   14208    33.01 fused_sigmoid_gating_delta_rule_update_kernel
     315.57    7954    39.67 void exl3_gemm_kernel<6, false, 1, 32, 32, 128, 4, 3>(__half const*, unsigned short const*, void*, int, int, i
     239.34     284   842.74 void exl3_gemm_kernel<6, false, 2, 32, 32, 128, 4, 3>(__half const*, unsigned short const*, void*, int, int, i
     209.04    2400    87.10 (anonymous namespace)::ni8_prefill_kernel((anonymous namespace)::Ni8Params)
     111.23     576   193.11 void arbi_serve_exl3_a_prep::a_prep_kernel<5, 32, true, true, arbi_serve_exl3_a_prep::SrcSiluMul<__half>, true
     103.56   41229     2.51 void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{la
      93.69    8445    11.09 void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_
      77.04   11829     6.51 void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, fl
      73.67     824    89.41 kernel_cutlass_kernel_split_d_tkvkernelscudaprefillturbo_attn_cute_bypassTurboAttnCuteBypass_object_at__tensor
      68.72     720    95.45 void arbi_serve_exl3_a_prep::a_prep_kernel<2, 32, true, true, arbi_serve_exl3_a_prep::SrcAddRMSNorm<__nv_bfloa
      61.56     432   142.50 chunk_gated_delta_rule_fwd_kernel_h_emit
      48.68     432   112.69 chunk_fwd_kernel_o
      47.88     569    84.16 Memcpy DtoH (Device -> Pinned)
      43.39    1629    26.64 _topc_mask_kernel
      41.91     565    74.18 void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous nam
      41.60   21660     1.92 void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::b
      40.00   24563     1.63 void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float
      37.51     864    43.42 void arbi_serve_exl3_a_prep::a_prep_kernel<2, 32, true, true, arbi_serve_exl3_a_prep::SrcHalf, true, 128>(arbi
      32.67     432    75.61 _kern_tiled
      32.13     432    74.36 recompute_w_u_fwd_kernel
      28.09    1629    17.25 _topc_partials_kernel
      26.56   22720     1.17 void at::native::vectorized_elementwise_kernel<4, at::native::float16_copy_kernel_cuda(at::TensorIteratorBase&
      25.44      29   877.37 ampere_fp16_s1688gemm_fp16_128x64_sliced1x2_ldg8_f2f_nn
      23.24    2416     9.62 fused_compress_store_kernel(__nv_bfloat16 const*, __nv_bfloat16 const*, unsigned char*, long const*, float con
      22.92     432    53.06 chunk_gated_delta_rule_fwd_kkt_solve_kernel
      21.73      25   869.19 void exl3_gemm_kernel<6, false, 2, 16, 32, 128, 4, 3>(__half const*, unsigned short const*, void*, int, int, i
      21.70    2320     9.35 (anonymous namespace)::tq_prefill_split_combine_kernel(float const*, float const*, __nv_bfloat16*, float*, int
      20.98    6720     3.12 void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_64x1_tn_align8>(cutlass_80_wmma_ten
      20.71    2368     8.74 (anonymous namespace)::ni8_vmax_kernel(unsigned char const*, int const*, int const*, int const*, int const*, i
      20.66    7104     2.91 _conv_window_commit_kernel
      19.38   11360     1.71 void at::native::vectorized_elementwise_kernel<4, at::native::sigmoid_kernel_cuda(at::TensorIteratorBase&)::{l
      19.29     516    37.38 void exl3_gemm_kernel<6, false, 1, 16, 32, 128, 4, 3>(__half const*, unsigned short const*, void*, int, int, i
      18.94    6816     2.78 _causal_conv1d_verify_kernel
      18.35     288    63.72 triton_per_fused__to_copy_add_mean_mul_pow_rsqrt_silu_view_18
      18.13     192    94.41 void exl3_i8_gemm_kernel_mb<8, 23, 48, 128, 128, 1, 8, 2, 8, false, false, true, false, false, false, true, 12
      18.02   17298     1.04 memcpy32_post
      17.71    4480     3.95 void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align8>(cutlass_80_wmma_te
      17.26    8406     2.05 void at::native::index_elementwise_kernel<128, 4, at::native::index_copy_kernel_impl<at::native::OpaqueType<2>
      16.48     512    32.19 void exl3_gemm_kernel<4, false, 2, 16, 32, 128, 4, 3>(__half const*, unsigned short const*, void*, int, int, i
      16.37     148   110.60 _residual_gumbel_partial_kernel
      15.64   12425     1.26 void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<c10::BFloat16>, std::array<char*
      15.09    6912     2.18 triton_red_fused__to_copy_add_exl3_gemm_group2_mean_mul_pow_rsqrt_1
      14.82    1064    13.93 _qwrite_draw_partials_kernel
      13.63   11829     1.15 void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_
      13.50   13632     0.99 triton_poi_fused__to_copy_exl3_add_rmsnorm_gemm2_8
      13.20    3000     4.40 void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_
      12.17    6816     1.78 triton_poi_fused__to_copy_cat_split_with_sizes_view_2
ring-ish kernels:
     209.04    2400    87.10 (anonymous namespace)::ni8_prefill_kernel((anonymous namespace)::Ni8Params)
      73.67     824    89.41 kernel_cutlass_kernel_split_d_tkvkernelscudaprefillturbo_attn_cute_bypassTurboAttnCuteBypass_object_at__tensor
      23.24    2416     9.62 fused_compress_store_kernel(__nv_bfloat16 const*, __nv_bfloat16 const*, unsigned char*, long const*, float con
      21.70    2320     9.35 (anonymous namespace)::tq_prefill_split_combine_kernel(float const*, float const*, __nv_bfloat16*, float*, int
       8.66    5995     1.44 void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<float>, std::array<char*, 1ul> >(int
       8.21     144    57.05 triton_poi_fused__to_copy_cat_gdn_prefill_conv_flat_index_select_mul_slice_split_with_sizes_squeeze_view_12
       7.64      96    79.58 triton_poi_fused__to_copy_cat_gdn_prefill_conv_flat_index_select_mul_slice_split_with_sizes_squeeze_view_16
       7.02    6816     1.03 triton_poi_fused_add_arange_gdn_replay_x_conv_save_mul_view_3
       6.88     144    47.80 triton_poi_fused__to_copy_gdn_prefill_conv_flat_index_select_mul_squeeze_view_2
       6.48     160    40.52 ampere_bf16_s1688gemm_bf16_128x128_ldg8_f2f_stages_32x1_nn
       5.85    6816     0.86 triton_poi_fused__to_copy_arange_gdn_verify_replay_save_split_split_with_sizes_view_4
       5.78    2235     2.59 void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<c10::BFloat16>, std::array<char*, 1u
       5.51      40   137.75 ampere_fp16_s1688gemm_fp16_128x128_ldg8_f2f_stages_32x1_nn
       5.46    6816     0.80 triton_poi_fused__to_copy_arange_gdn_verify_replay_save_split_split_with_sizes_view_5
       3.94      48    82.08 triton_poi_fused__to_copy_cat_gdn_prefill_conv_flat_index_select_mul_slice_split_with_sizes_squeeze_view_11
       3.88    2386     1.63 void at::native::_scatter_gather_elementwise_kernel<128, 8, at::native::_cuda_scatter_gather_internal_kernel<f
       3.42      48    71.30 triton_poi_fused__to_copy_gdn_prefill_conv_flat_index_select_mul_squeeze_view_1
       2.44    3066     0.80 void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, 
       2.36     432     5.46 chunk_local_cumsum_scalar_kernel
       1.82    2077     0.87 void at::native::vectorized_elementwise_kernel<2, at::native::FillFunctor<long>, std::array<char*, 1ul> >(int,
       1.75    1480     1.18 void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast<at::native::FillFunctor<c10::BF
       1.31    1182     1.11 void at::native::vectorized_elementwise_kernel<2, at::native::(anonymous namespace)::where_kernel_impl(at::Ten
       1.25     720     1.73 triton_per_fused__to_copy_add_arange_cat_copy_cos_dflash_varlen_attend_mean_mul_neg_pow_rsqrt_sin_slice_unsque
       1.11     720     1.54 triton_per_fused__to_copy_add_arange_as_strided_cat_cos_dflash_place_block_write_mean_mul_neg_pow_rsqrt_sin_sl
       0.73     288     2.55 triton_poi_fused__to_copy_add_arange_cat_index_copy_mul_slice_split_with_sizes_view_8
       0.67     192     3.49 triton_poi_fused__to_copy_add_arange_cat_index_copy_mul_slice_split_with_sizes_view_4
       0.57      30    19.14 ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_32x6_nn
       0.55     613     0.89 void (anonymous namespace)::elementwise_kernel_with_index<int, at::native::arange_cuda_out(c10::Scalar const&,
       0.54     326     1.64 void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::FillFunctor<float> 
       0.52      96     5.44 void at::native::index_elementwise_kernel<128, 4, at::native::index_fill_kernel_impl<at::native::OpaqueType<4>
       0.50     304     1.65 at::native::mbtopk::computeDigitCumSum(short*, unsigned int*, unsigned int)
       0.43     288     1.49 triton_poi_fused__to_copy_arange_copy_select_silu_slice_split_split_with_sizes_view_15
       0.42     288     1.44 triton_poi_fused__to_copy_cat_gdn_prefill_conv_flat_index_select_mul_slice_split_with_sizes_squeeze_view_2
       0.39     288     1.34 triton_poi_fused__to_copy_arange_copy_select_silu_slice_split_split_with_sizes_view_0
       0.37     192     1.90 triton_poi_fused__to_copy_arange_copy_select_silu_slice_split_split_with_sizes_view_17
       0.33     151     2.19 _stage_claim
       0.33      10    32.65 ampere_fp16_s16816gemm_fp16_64x64_ldg8_f2f_stages_64x5_nn
       0.28     192     1.48 triton_poi_fused__to_copy_arange_copy_select_silu_slice_split_split_with_sizes_view_16
       0.28     288     0.97 triton_poi_fused__to_copy_arange_copy_select_silu_slice_split_split_with_sizes_view_1
       0.27     288     0.93 triton_poi_fused__to_copy_add_arange_cat_copy_cos_dflash_varlen_attend_mean_mul_pow_rsqrt_sin_unsqueeze_view_9
       0.27     152     1.75 triton_poi_fused_add_clamp_copy__index_select_remainder_searchsorted_sub_view_1
       0.23      96     2.36 triton_poi_fused__to_copy_add_arange_cat_index_copy_mul_slice_split_with_sizes_view_3
       0.22     148     1.47 void at::native::_scatter_gather_elementwise_kernel<128, 8, at::native::_cuda_scatter_gather_internal_kernel<t
       0.21      96     2.23 triton_poi_fused__to_copy_gdn_prefill_conv_flat_index_select_mul_squeeze_view_3
       0.21     151     1.40 void at::native::(anonymous namespace)::searchsorted_cuda_kernel<long, long>(long*, long const*, long const*, 
       0.19     147     1.32 void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::FillFunctor<int> >(
       0.19     148     1.29 void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::where_kernel_impl(at::Ten
       0.19     144     1.32 triton_poi_fused__unsafe_view_add_clone_copy__expand_remainder_view_2
       0.19     151     1.24 _stage_tag
       0.17     152     1.13 void at::native::vectorized_elementwise_kernel<2, at::native::(anonymous namespace)::masked_fill_kernel(at::Te
       0.17     144     1.15 triton_per_fused_add_clamp_cumsum_0
       0.16     144     1.14 triton_poi_fused_add_clamp_copy_copy__remainder_sub_view_8
       0.14     144     0.99 triton_poi_fused__to_copy_add_arange_cat_copy_copy__cos_dflash_varlen_attend_mean_mul_pow_rsqrt_sin_unsqueeze_
       0.14      96     1.44 void at::native::index_elementwise_kernel<128, 4, at::native::index_fill_kernel_impl<at::native::OpaqueType<2>
       0.13     144     0.94 triton_poi_fused_add_clamp_copy__remainder_sub_view_19
       0.13      96     1.34 triton_poi_fused__to_copy_arange_copy_select_silu_slice_split_split_with_sizes_view_13
       0.12      96     1.30 triton_poi_fused__to_copy_arange_copy_select_silu_slice_split_split_with_sizes_view_14
       0.11     152     0.73 void at::native::mbtopk::fill<unsigned int, unsigned int>(unsigned int*, unsigned int, unsigned int)
       0.07      40     1.65 triton_per_fused__to_copy_add_arange_cat_clamp_copy_cos_dflash_varlen_attend_mean_mul_neg_pow_remainder_rsqrt_
       0.06      40     1.46 triton_per_fused__to_copy_add_arange_as_strided_cat_cos_dflash_place_block_write_expand_mean_mul_neg_pow_remai
       0.04      40     0.89 triton_poi_fused__to_copy_add_as_strided_dflash_place_block_write_expand_remainder_view_6
       0.03      32     1.04 triton_poi_fused__to_copy_add_arange_cat_clamp_copy_cos_dflash_varlen_attend_mean_mul_pow_remainder_rsqrt_sin_
       0.02      16     1.13 triton_poi_fused__to_copy_add_arange_as_strided_cat_cos_dflash_place_block_write_expand_mean_mul_pow_remainder
       0.02      16     1.04 triton_poi_fused__to_copy_add_arange_cat_clamp_copy_copy__cos_dflash_varlen_attend_mean_mul_pow_remainder_rsqr
       0.01       8     1.38 triton_poi_fused_add_clamp_cumsum_0
       0.01      11     0.87 void at::native::unrolled_elementwise_kernel<at::native::FillFunctor<long>, std::array<char*, 1ul>, 4, Trivial
       0.01       8     1.12 triton_poi_fused__to_copy_add_arange_as_strided_cat_clamp_copy__cos_dflash_place_block_write_expand_mean_mul_p
       0.01       6     0.95 void at::native::unrolled_elementwise_kernel<at::native::FillFunctor<int>, std::array<char*, 1ul>, 4, TrivialO
       0.00       2     0.99 void at::native::vectorized_elementwise_kernel<2, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, 
