CoreMLInspect-OpenELM-270M-Instruct-Compiled-Model-CPU-NE.txt

Analyzing model for compute unit [cpuAndNeuralEngine]...

Key: C=CPU, G=GPU, N=NeuralEngine
<Estimate of total operation cost>% <primary compute|supported compute>  <operation>
func main(input_ids) {
           var_5 = const()
           var_22 = const()
           inputs_embeds_axis_0 = const()
           inputs_embeds_batch_dims_0 = const()
           inputs_embeds_validate_indices_0 = const()
           model_transformer_token_embeddings_weight_to_fp16 = const()
           input_ids_to_int16_dtype_0 = const()
0.00% C|   cast_34 = ios17.cast(x: ["input_ids"], dtype: ["input_ids_to_int16_dtype_0"])
7.66% C|   inputs_embeds_cast_fp16_cast_int16 = ios17.gather(x: ["model_transformer_token_embeddings_weight_to_fp16"], indices: ["cast_34"], axis: ["inputs_embeds_axis_0"], batch_dims: ["inputs_embeds_batch_dims_0"], validate_indices: ["inputs_embeds_validate_indices_0"])
           var_10_promoted_to_fp16 = const()
0.04% N|C  var_94_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16"], x: ["inputs_embeds_cast_fp16_cast_int16"])
           var_95 = const()
0.02% N|C  var_96_cast_fp16 = ios16.reduce_mean(axes: ["var_95"], keep_dims: ["var_22"], x: ["var_94_cast_fp16"])
           var_97_to_fp16 = const()
0.00% N|C  var_98_cast_fp16 = ios17.add(y: ["var_97_to_fp16"], x: ["var_96_cast_fp16"])
           var_99_epsilon_0 = const()
0.00% N|C  var_99_cast_fp16 = ios17.rsqrt(x: ["var_98_cast_fp16"], epsilon: ["var_99_epsilon_0"])
0.04% N|C  var_100_cast_fp16 = ios17.mul(y: ["var_99_cast_fp16"], x: ["inputs_embeds_cast_fp16_cast_int16"])
           model_transformer_layers_0_attn_norm_weight_to_fp16 = const()
0.04% N|C  hidden_states_1_cast_fp16 = ios17.mul(y: ["model_transformer_layers_0_attn_norm_weight_to_fp16"], x: ["var_100_cast_fp16"])
           model_transformer_layers_0_attn_qkv_proj_weight_to_fp16 = const()
           linear_0_bias_0_to_fp16 = const()
0.38% N|C  linear_0_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_0_attn_qkv_proj_weight_to_fp16"], bias: ["linear_0_bias_0_to_fp16"], x: ["hidden_states_1_cast_fp16"])
           var_114 = const()
0.04% N|C  qkv_3_cast_fp16 = ios17.reshape(shape: ["var_114"], x: ["linear_0_cast_fp16"])
           var_116_perm_0 = const()
           var_117 = const()
           var_118_axis_0 = const()
0.08% N|C  transpose_95 = ios17.transpose(x: ["qkv_3_cast_fp16"], perm: ["var_116_perm_0"])
0.03% N|C  var_118_cast_fp16_0, var_118_cast_fp16_1, var_118_cast_fp16_2 = split(axis: ["var_118_axis_0"], x: ["transpose_95"], split_sizes: ["var_117"])
           var_10_promoted_to_fp16_1 = const()
0.03% N|C  var_124_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_1"], x: ["var_118_cast_fp16_0"])
           var_125 = const()
0.01% N|C  var_126_cast_fp16 = ios16.reduce_mean(axes: ["var_125"], x: ["var_124_cast_fp16"], keep_dims: ["var_22"])
           var_127_to_fp16 = const()
0.00% N|C  var_128_cast_fp16 = ios17.add(y: ["var_127_to_fp16"], x: ["var_126_cast_fp16"])
           var_129_epsilon_0 = const()
0.00% N|C  var_129_cast_fp16 = ios17.rsqrt(epsilon: ["var_129_epsilon_0"], x: ["var_128_cast_fp16"])
0.02% N|C  var_130_cast_fp16 = ios17.mul(x: ["var_118_cast_fp16_0"], y: ["var_129_cast_fp16"])
           model_transformer_layers_0_attn_q_norm_weight_to_fp16 = const()
0.02% N|C  query_1_cast_fp16 = ios17.mul(y: ["model_transformer_layers_0_attn_q_norm_weight_to_fp16"], x: ["var_130_cast_fp16"])
           var_10_promoted_to_fp16_2 = const()
0.01% N|C  var_135_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_2"], x: ["var_118_cast_fp16_1"])
           var_136 = const()
0.00% N|C  var_137_cast_fp16 = ios16.reduce_mean(axes: ["var_136"], keep_dims: ["var_22"], x: ["var_135_cast_fp16"])
           var_138_to_fp16 = const()
0.00% N|C  var_139_cast_fp16 = ios17.add(y: ["var_138_to_fp16"], x: ["var_137_cast_fp16"])
           var_140_epsilon_0 = const()
0.00% N|C  var_140_cast_fp16 = ios17.rsqrt(x: ["var_139_cast_fp16"], epsilon: ["var_140_epsilon_0"])
0.01% N|C  var_141_cast_fp16 = ios17.mul(y: ["var_140_cast_fp16"], x: ["var_118_cast_fp16_1"])
           model_transformer_layers_0_attn_k_norm_weight_to_fp16 = const()
0.01% N|C  key_1_cast_fp16 = ios17.mul(x: ["var_141_cast_fp16"], y: ["model_transformer_layers_0_attn_k_norm_weight_to_fp16"])
           var_156_to_fp16 = const()
0.02% N|C  var_158_cast_fp16 = ios17.mul(x: ["query_1_cast_fp16"], y: ["var_156_to_fp16"])
           var_159_split_sizes_0 = const()
           var_159_axis_0 = const()
0.02% N|C  var_159_cast_fp16_0, var_159_cast_fp16_1 = split(axis: ["var_159_axis_0"], x: ["query_1_cast_fp16"], split_sizes: ["var_159_split_sizes_0"])
           const_6_promoted_to_fp16 = const()
0.01% N|C  var_161_cast_fp16 = ios17.mul(x: ["var_159_cast_fp16_1"], y: ["const_6_promoted_to_fp16"])
           var_163_interleave_0 = const()
0.01% N|C  var_163_cast_fp16 = ios17.concat(axis: ["var_5"], interleave: ["var_163_interleave_0"], values: ["var_161_cast_fp16", "var_159_cast_fp16_0"])
           var_152_to_fp16 = const()
0.02% N|C  var_164_cast_fp16 = ios17.mul(x: ["var_163_cast_fp16"], y: ["var_152_to_fp16"])
0.06% N|C  query_float_1_cast_fp16 = ios17.add(y: ["var_164_cast_fp16"], x: ["var_158_cast_fp16"])
0.01% N|C  var_170_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["key_1_cast_fp16"])
           var_171_split_sizes_0 = const()
           var_171_axis_0 = const()
0.00% N|C  var_171_cast_fp16_0, var_171_cast_fp16_1 = split(split_sizes: ["var_171_split_sizes_0"], x: ["key_1_cast_fp16"], axis: ["var_171_axis_0"])
           const_7_promoted_to_fp16 = const()
0.00% N|C  var_173_cast_fp16 = ios17.mul(x: ["var_171_cast_fp16_1"], y: ["const_7_promoted_to_fp16"])
           var_175_interleave_0 = const()
0.00% N|C  var_175_cast_fp16 = ios17.concat(axis: ["var_5"], values: ["var_173_cast_fp16", "var_171_cast_fp16_0"], interleave: ["var_175_interleave_0"])
0.01% N|C  var_176_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_175_cast_fp16"])
0.01% N|C  key_float_1_cast_fp16 = ios17.add(y: ["var_176_cast_fp16"], x: ["var_170_cast_fp16"])
           reshape_0_shape_0 = const()
0.01% N|C  reshape_0_cast_fp16 = ios17.reshape(x: ["key_float_1_cast_fp16"], shape: ["reshape_0_shape_0"])
           tile_0_reps_0 = const()
0.02% N|C  tile_0_cast_fp16 = tile(x: ["reshape_0_cast_fp16"], reps: ["tile_0_reps_0"])
           reshape_1_shape_0 = const()
0.03% N|C  reshape_1_cast_fp16 = ios17.reshape(x: ["tile_0_cast_fp16"], shape: ["reshape_1_shape_0"])
           transpose_0_perm_0 = const()
           keys_3_shape_0 = const()
0.05% N|C  transpose_94 = ios17.transpose(perm: ["transpose_0_perm_0"], x: ["reshape_1_cast_fp16"])
0.03% N|C  keys_3_cast_fp16 = ios17.reshape(x: ["transpose_94"], shape: ["keys_3_shape_0"])
           reshape_2_shape_0 = const()
0.01% N|C  reshape_2_cast_fp16 = ios17.reshape(shape: ["reshape_2_shape_0"], x: ["var_118_cast_fp16_2"])
           tile_1_reps_0 = const()
0.02% N|C  tile_1_cast_fp16 = tile(reps: ["tile_1_reps_0"], x: ["reshape_2_cast_fp16"])
           reshape_3_shape_0 = const()
0.03% N|C  reshape_3_cast_fp16 = ios17.reshape(shape: ["reshape_3_shape_0"], x: ["tile_1_cast_fp16"])
           transpose_1_perm_0 = const()
           values_3_shape_0 = const()
0.05% N|C  transpose_93 = ios17.transpose(perm: ["transpose_1_perm_0"], x: ["reshape_3_cast_fp16"])
0.03% N|C  values_3_cast_fp16 = ios17.reshape(x: ["transpose_93"], shape: ["values_3_shape_0"])
           mul_0_y_0_to_fp16 = const()
0.02% N|C  mul_0_cast_fp16 = ios17.mul(x: ["query_float_1_cast_fp16"], y: ["mul_0_y_0_to_fp16"])
           matmul_0_transpose_y_0 = const()
           matmul_0_transpose_x_0 = const()
0.05% N|C  matmul_0_cast_fp16 = ios17.matmul(transpose_y: ["matmul_0_transpose_y_0"], transpose_x: ["matmul_0_transpose_x_0"], y: ["keys_3_cast_fp16"], x: ["mul_0_cast_fp16"])
           causal_mask_3_to_fp16 = const()
0.06% N|C  add_1_cast_fp16 = ios17.add(y: ["causal_mask_3_to_fp16"], x: ["matmul_0_cast_fp16"])
           softmax_0_axis_0 = const()
0.05% N|C  softmax_0_cast_fp16 = ios16.softmax(axis: ["softmax_0_axis_0"], x: ["add_1_cast_fp16"])
           attn_output_1_transpose_x_0 = const()
           attn_output_1_transpose_y_0 = const()
0.05% N|C  attn_output_1_cast_fp16 = ios17.matmul(transpose_x: ["attn_output_1_transpose_x_0"], x: ["softmax_0_cast_fp16"], y: ["values_3_cast_fp16"], transpose_y: ["attn_output_1_transpose_y_0"])
           var_192_perm_0 = const()
           var_194 = const()
0.05% N|C  transpose_92 = ios17.transpose(x: ["attn_output_1_cast_fp16"], perm: ["var_192_perm_0"])
0.03% N|C  input_3_cast_fp16 = ios17.reshape(x: ["transpose_92"], shape: ["var_194"])
           model_transformer_layers_0_attn_out_proj_weight_to_fp16 = const()
           linear_1_bias_0_to_fp16 = const()
0.14% N|C  linear_1_cast_fp16 = ios17.linear(bias: ["linear_1_bias_0_to_fp16"], weight: ["model_transformer_layers_0_attn_out_proj_weight_to_fp16"], x: ["input_3_cast_fp16"])
0.09% N|C  x_15_cast_fp16 = ios17.add(y: ["linear_1_cast_fp16"], x: ["inputs_embeds_cast_fp16_cast_int16"])
           var_10_promoted_to_fp16_3 = const()
0.04% N|C  var_201_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_3"], x: ["x_15_cast_fp16"])
           var_202 = const()
0.02% N|C  var_203_cast_fp16 = ios16.reduce_mean(axes: ["var_202"], x: ["var_201_cast_fp16"], keep_dims: ["var_22"])
           var_204_to_fp16 = const()
0.00% N|C  var_205_cast_fp16 = ios17.add(y: ["var_204_to_fp16"], x: ["var_203_cast_fp16"])
           var_206_epsilon_0 = const()
0.00% N|C  var_206_cast_fp16 = ios17.rsqrt(x: ["var_205_cast_fp16"], epsilon: ["var_206_epsilon_0"])
0.04% N|C  var_207_cast_fp16 = ios17.mul(x: ["x_15_cast_fp16"], y: ["var_206_cast_fp16"])
           model_transformer_layers_0_ffn_norm_weight_to_fp16 = const()
0.04% N|C  input_5_cast_fp16 = ios17.mul(x: ["var_207_cast_fp16"], y: ["model_transformer_layers_0_ffn_norm_weight_to_fp16"])
           model_transformer_layers_0_ffn_proj_1_weight_to_fp16 = const()
           linear_2_bias_0_to_fp16 = const()
0.38% N|C  linear_2_cast_fp16 = ios17.linear(x: ["input_5_cast_fp16"], bias: ["linear_2_bias_0_to_fp16"], weight: ["model_transformer_layers_0_ffn_proj_1_weight_to_fp16"])
           var_217_split_sizes_0 = const()
           var_217_axis_0 = const()
0.04% N|C  var_217_cast_fp16_0, var_217_cast_fp16_1 = split(split_sizes: ["var_217_split_sizes_0"], x: ["linear_2_cast_fp16"], axis: ["var_217_axis_0"])
0.03% N|C  var_219_cast_fp16 = ios16.silu(x: ["var_217_cast_fp16_0"])
0.05% N|C  input_9_cast_fp16 = ios17.mul(x: ["var_219_cast_fp16"], y: ["var_217_cast_fp16_1"])
           model_transformer_layers_0_ffn_proj_2_weight_to_fp16 = const()
0.14% N|C  linear_3_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_0_ffn_proj_2_weight_to_fp16"], x: ["input_9_cast_fp16"], bias: ["linear_1_bias_0_to_fp16"])
0.09% N|C  x_19_cast_fp16 = ios17.add(x: ["x_15_cast_fp16"], y: ["linear_3_cast_fp16"])
           var_10_promoted_to_fp16_4 = const()
0.04% N|C  var_230_cast_fp16 = ios17.pow(x: ["x_19_cast_fp16"], y: ["var_10_promoted_to_fp16_4"])
           var_231 = const()
0.02% N|C  var_232_cast_fp16 = ios16.reduce_mean(axes: ["var_231"], keep_dims: ["var_22"], x: ["var_230_cast_fp16"])
           var_233_to_fp16 = const()
0.00% N|C  var_234_cast_fp16 = ios17.add(y: ["var_233_to_fp16"], x: ["var_232_cast_fp16"])
           var_235_epsilon_0 = const()
0.00% N|C  var_235_cast_fp16 = ios17.rsqrt(x: ["var_234_cast_fp16"], epsilon: ["var_235_epsilon_0"])
0.04% N|C  var_236_cast_fp16 = ios17.mul(x: ["x_19_cast_fp16"], y: ["var_235_cast_fp16"])
           model_transformer_layers_1_attn_norm_weight_to_fp16 = const()
0.04% N|C  hidden_states_7_cast_fp16 = ios17.mul(y: ["model_transformer_layers_1_attn_norm_weight_to_fp16"], x: ["var_236_cast_fp16"])
           model_transformer_layers_1_attn_qkv_proj_weight_to_fp16 = const()
0.38% N|C  linear_4_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_1_attn_qkv_proj_weight_to_fp16"], bias: ["linear_0_bias_0_to_fp16"], x: ["hidden_states_7_cast_fp16"])
           var_250 = const()
0.04% N|C  qkv_7_cast_fp16 = ios17.reshape(shape: ["var_250"], x: ["linear_4_cast_fp16"])
           var_252_perm_0 = const()
           var_253 = const()
           var_254_axis_0 = const()
0.08% N|C  transpose_91 = ios17.transpose(perm: ["var_252_perm_0"], x: ["qkv_7_cast_fp16"])
0.03% N|C  var_254_cast_fp16_0, var_254_cast_fp16_1, var_254_cast_fp16_2 = split(x: ["transpose_91"], axis: ["var_254_axis_0"], split_sizes: ["var_253"])
           var_10_promoted_to_fp16_5 = const()
0.03% N|C  var_260_cast_fp16 = ios17.pow(x: ["var_254_cast_fp16_0"], y: ["var_10_promoted_to_fp16_5"])
           var_261 = const()
0.01% N|C  var_262_cast_fp16 = ios16.reduce_mean(x: ["var_260_cast_fp16"], axes: ["var_261"], keep_dims: ["var_22"])
           var_263_to_fp16 = const()
0.00% N|C  var_264_cast_fp16 = ios17.add(y: ["var_263_to_fp16"], x: ["var_262_cast_fp16"])
           var_265_epsilon_0 = const()
0.00% N|C  var_265_cast_fp16 = ios17.rsqrt(epsilon: ["var_265_epsilon_0"], x: ["var_264_cast_fp16"])
0.02% N|C  var_266_cast_fp16 = ios17.mul(x: ["var_254_cast_fp16_0"], y: ["var_265_cast_fp16"])
           model_transformer_layers_1_attn_q_norm_weight_to_fp16 = const()
0.02% N|C  query_3_cast_fp16 = ios17.mul(y: ["model_transformer_layers_1_attn_q_norm_weight_to_fp16"], x: ["var_266_cast_fp16"])
           var_10_promoted_to_fp16_6 = const()
0.01% N|C  var_271_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_6"], x: ["var_254_cast_fp16_1"])
           var_272 = const()
0.00% N|C  var_273_cast_fp16 = ios16.reduce_mean(axes: ["var_272"], keep_dims: ["var_22"], x: ["var_271_cast_fp16"])
           var_274_to_fp16 = const()
0.00% N|C  var_275_cast_fp16 = ios17.add(x: ["var_273_cast_fp16"], y: ["var_274_to_fp16"])
           var_276_epsilon_0 = const()
0.00% N|C  var_276_cast_fp16 = ios17.rsqrt(x: ["var_275_cast_fp16"], epsilon: ["var_276_epsilon_0"])
0.01% N|C  var_277_cast_fp16 = ios17.mul(y: ["var_276_cast_fp16"], x: ["var_254_cast_fp16_1"])
           model_transformer_layers_1_attn_k_norm_weight_to_fp16 = const()
0.01% N|C  key_3_cast_fp16 = ios17.mul(y: ["model_transformer_layers_1_attn_k_norm_weight_to_fp16"], x: ["var_277_cast_fp16"])
0.02% N|C  var_294_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["query_3_cast_fp16"])
           var_295_split_sizes_0 = const()
           var_295_axis_0 = const()
0.02% N|C  var_295_cast_fp16_0, var_295_cast_fp16_1 = split(split_sizes: ["var_295_split_sizes_0"], axis: ["var_295_axis_0"], x: ["query_3_cast_fp16"])
           const_13_promoted_to_fp16 = const()
0.01% N|C  var_297_cast_fp16 = ios17.mul(x: ["var_295_cast_fp16_1"], y: ["const_13_promoted_to_fp16"])
           var_299_interleave_0 = const()
0.01% N|C  var_299_cast_fp16 = ios17.concat(axis: ["var_5"], interleave: ["var_299_interleave_0"], values: ["var_297_cast_fp16", "var_295_cast_fp16_0"])
0.02% N|C  var_300_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_299_cast_fp16"])
0.06% N|C  query_float_3_cast_fp16 = ios17.add(y: ["var_300_cast_fp16"], x: ["var_294_cast_fp16"])
0.01% N|C  var_306_cast_fp16 = ios17.mul(x: ["key_3_cast_fp16"], y: ["var_156_to_fp16"])
           var_307_split_sizes_0 = const()
           var_307_axis_0 = const()
0.00% N|C  var_307_cast_fp16_0, var_307_cast_fp16_1 = split(split_sizes: ["var_307_split_sizes_0"], axis: ["var_307_axis_0"], x: ["key_3_cast_fp16"])
           const_14_promoted_to_fp16 = const()
0.00% N|C  var_309_cast_fp16 = ios17.mul(y: ["const_14_promoted_to_fp16"], x: ["var_307_cast_fp16_1"])
           var_311_interleave_0 = const()
0.00% N|C  var_311_cast_fp16 = ios17.concat(values: ["var_309_cast_fp16", "var_307_cast_fp16_0"], axis: ["var_5"], interleave: ["var_311_interleave_0"])
0.01% N|C  var_312_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_311_cast_fp16"])
0.01% N|C  key_float_3_cast_fp16 = ios17.add(y: ["var_312_cast_fp16"], x: ["var_306_cast_fp16"])
           reshape_4_shape_0 = const()
0.01% N|C  reshape_4_cast_fp16 = ios17.reshape(x: ["key_float_3_cast_fp16"], shape: ["reshape_4_shape_0"])
           tile_2_reps_0 = const()
0.02% N|C  tile_2_cast_fp16 = tile(x: ["reshape_4_cast_fp16"], reps: ["tile_2_reps_0"])
           reshape_5_shape_0 = const()
0.03% N|C  reshape_5_cast_fp16 = ios17.reshape(shape: ["reshape_5_shape_0"], x: ["tile_2_cast_fp16"])
           transpose_2_perm_0 = const()
           keys_7_shape_0 = const()
0.05% N|C  transpose_90 = ios17.transpose(perm: ["transpose_2_perm_0"], x: ["reshape_5_cast_fp16"])
0.03% N|C  keys_7_cast_fp16 = ios17.reshape(shape: ["keys_7_shape_0"], x: ["transpose_90"])
           reshape_6_shape_0 = const()
0.01% N|C  reshape_6_cast_fp16 = ios17.reshape(shape: ["reshape_6_shape_0"], x: ["var_254_cast_fp16_2"])
           tile_3_reps_0 = const()
0.02% N|C  tile_3_cast_fp16 = tile(reps: ["tile_3_reps_0"], x: ["reshape_6_cast_fp16"])
           reshape_7_shape_0 = const()
0.03% N|C  reshape_7_cast_fp16 = ios17.reshape(x: ["tile_3_cast_fp16"], shape: ["reshape_7_shape_0"])
           transpose_3_perm_0 = const()
           values_7_shape_0 = const()
0.05% N|C  transpose_89 = ios17.transpose(x: ["reshape_7_cast_fp16"], perm: ["transpose_3_perm_0"])
0.03% N|C  values_7_cast_fp16 = ios17.reshape(shape: ["values_7_shape_0"], x: ["transpose_89"])
           mul_1_y_0_to_fp16 = const()
0.02% N|C  mul_1_cast_fp16 = ios17.mul(y: ["mul_1_y_0_to_fp16"], x: ["query_float_3_cast_fp16"])
           matmul_1_transpose_y_0 = const()
           matmul_1_transpose_x_0 = const()
0.05% N|C  matmul_1_cast_fp16 = ios17.matmul(x: ["mul_1_cast_fp16"], transpose_x: ["matmul_1_transpose_x_0"], y: ["keys_7_cast_fp16"], transpose_y: ["matmul_1_transpose_y_0"])
0.06% N|C  add_3_cast_fp16 = ios17.add(x: ["matmul_1_cast_fp16"], y: ["causal_mask_3_to_fp16"])
           softmax_1_axis_0 = const()
0.05% N|C  softmax_1_cast_fp16 = ios16.softmax(axis: ["softmax_1_axis_0"], x: ["add_3_cast_fp16"])
           attn_output_5_transpose_x_0 = const()
           attn_output_5_transpose_y_0 = const()
0.05% N|C  attn_output_5_cast_fp16 = ios17.matmul(x: ["softmax_1_cast_fp16"], y: ["values_7_cast_fp16"], transpose_y: ["attn_output_5_transpose_y_0"], transpose_x: ["attn_output_5_transpose_x_0"])
           var_328_perm_0 = const()
           var_330 = const()
0.05% N|C  transpose_88 = ios17.transpose(perm: ["var_328_perm_0"], x: ["attn_output_5_cast_fp16"])
0.03% N|C  input_11_cast_fp16 = ios17.reshape(shape: ["var_330"], x: ["transpose_88"])
           model_transformer_layers_1_attn_out_proj_weight_to_fp16 = const()
0.14% N|C  linear_5_cast_fp16 = ios17.linear(bias: ["linear_1_bias_0_to_fp16"], x: ["input_11_cast_fp16"], weight: ["model_transformer_layers_1_attn_out_proj_weight_to_fp16"])
0.09% N|C  x_35_cast_fp16 = ios17.add(y: ["linear_5_cast_fp16"], x: ["x_19_cast_fp16"])
           var_10_promoted_to_fp16_7 = const()
0.04% N|C  var_337_cast_fp16 = ios17.pow(x: ["x_35_cast_fp16"], y: ["var_10_promoted_to_fp16_7"])
           var_338 = const()
0.02% N|C  var_339_cast_fp16 = ios16.reduce_mean(x: ["var_337_cast_fp16"], keep_dims: ["var_22"], axes: ["var_338"])
           var_340_to_fp16 = const()
0.00% N|C  var_341_cast_fp16 = ios17.add(x: ["var_339_cast_fp16"], y: ["var_340_to_fp16"])
           var_342_epsilon_0 = const()
0.00% N|C  var_342_cast_fp16 = ios17.rsqrt(x: ["var_341_cast_fp16"], epsilon: ["var_342_epsilon_0"])
0.04% N|C  var_343_cast_fp16 = ios17.mul(y: ["var_342_cast_fp16"], x: ["x_35_cast_fp16"])
           model_transformer_layers_1_ffn_norm_weight_to_fp16 = const()
0.04% N|C  input_13_cast_fp16 = ios17.mul(y: ["model_transformer_layers_1_ffn_norm_weight_to_fp16"], x: ["var_343_cast_fp16"])
           model_transformer_layers_1_ffn_proj_1_weight_to_fp16 = const()
           linear_6_bias_0_to_fp16 = const()
0.38% N|C  linear_6_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_1_ffn_proj_1_weight_to_fp16"], x: ["input_13_cast_fp16"], bias: ["linear_6_bias_0_to_fp16"])
           var_353_split_sizes_0 = const()
           var_353_axis_0 = const()
0.05% N|C  var_353_cast_fp16_0, var_353_cast_fp16_1 = split(x: ["linear_6_cast_fp16"], split_sizes: ["var_353_split_sizes_0"], axis: ["var_353_axis_0"])
0.03% N|C  var_355_cast_fp16 = ios16.silu(x: ["var_353_cast_fp16_0"])
0.06% N|C  input_17_cast_fp16 = ios17.mul(y: ["var_353_cast_fp16_1"], x: ["var_355_cast_fp16"])
           model_transformer_layers_1_ffn_proj_2_weight_to_fp16 = const()
0.24% N|C  linear_7_cast_fp16 = ios17.linear(bias: ["linear_1_bias_0_to_fp16"], x: ["input_17_cast_fp16"], weight: ["model_transformer_layers_1_ffn_proj_2_weight_to_fp16"])
0.09% N|C  x_39_cast_fp16 = ios17.add(y: ["linear_7_cast_fp16"], x: ["x_35_cast_fp16"])
           var_10_promoted_to_fp16_8 = const()
0.04% N|C  var_366_cast_fp16 = ios17.pow(x: ["x_39_cast_fp16"], y: ["var_10_promoted_to_fp16_8"])
           var_367 = const()
0.02% N|C  var_368_cast_fp16 = ios16.reduce_mean(x: ["var_366_cast_fp16"], keep_dims: ["var_22"], axes: ["var_367"])
           var_369_to_fp16 = const()
0.00% N|C  var_370_cast_fp16 = ios17.add(y: ["var_369_to_fp16"], x: ["var_368_cast_fp16"])
           var_371_epsilon_0 = const()
0.00% N|C  var_371_cast_fp16 = ios17.rsqrt(epsilon: ["var_371_epsilon_0"], x: ["var_370_cast_fp16"])
0.04% N|C  var_372_cast_fp16 = ios17.mul(x: ["x_39_cast_fp16"], y: ["var_371_cast_fp16"])
           model_transformer_layers_2_attn_norm_weight_to_fp16 = const()
0.04% N|C  hidden_states_13_cast_fp16 = ios17.mul(x: ["var_372_cast_fp16"], y: ["model_transformer_layers_2_attn_norm_weight_to_fp16"])
           model_transformer_layers_2_attn_qkv_proj_weight_to_fp16 = const()
0.38% N|C  linear_8_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_2_attn_qkv_proj_weight_to_fp16"], x: ["hidden_states_13_cast_fp16"], bias: ["linear_0_bias_0_to_fp16"])
           var_386 = const()
0.04% N|C  qkv_11_cast_fp16 = ios17.reshape(x: ["linear_8_cast_fp16"], shape: ["var_386"])
           var_388_perm_0 = const()
           var_389 = const()
           var_390_axis_0 = const()
0.08% N|C  transpose_87 = ios17.transpose(perm: ["var_388_perm_0"], x: ["qkv_11_cast_fp16"])
0.03% N|C  var_390_cast_fp16_0, var_390_cast_fp16_1, var_390_cast_fp16_2 = split(axis: ["var_390_axis_0"], x: ["transpose_87"], split_sizes: ["var_389"])
           var_10_promoted_to_fp16_9 = const()
0.03% N|C  var_396_cast_fp16 = ios17.pow(x: ["var_390_cast_fp16_0"], y: ["var_10_promoted_to_fp16_9"])
           var_397 = const()
0.01% N|C  var_398_cast_fp16 = ios16.reduce_mean(axes: ["var_397"], x: ["var_396_cast_fp16"], keep_dims: ["var_22"])
           var_399_to_fp16 = const()
0.00% N|C  var_400_cast_fp16 = ios17.add(y: ["var_399_to_fp16"], x: ["var_398_cast_fp16"])
           var_401_epsilon_0 = const()
0.00% N|C  var_401_cast_fp16 = ios17.rsqrt(x: ["var_400_cast_fp16"], epsilon: ["var_401_epsilon_0"])
0.02% N|C  var_402_cast_fp16 = ios17.mul(y: ["var_401_cast_fp16"], x: ["var_390_cast_fp16_0"])
           model_transformer_layers_2_attn_q_norm_weight_to_fp16 = const()
0.02% N|C  query_5_cast_fp16 = ios17.mul(x: ["var_402_cast_fp16"], y: ["model_transformer_layers_2_attn_q_norm_weight_to_fp16"])
           var_10_promoted_to_fp16_10 = const()
0.01% N|C  var_407_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_10"], x: ["var_390_cast_fp16_1"])
           var_408 = const()
0.00% N|C  var_409_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_407_cast_fp16"], axes: ["var_408"])
           var_410_to_fp16 = const()
0.00% N|C  var_411_cast_fp16 = ios17.add(y: ["var_410_to_fp16"], x: ["var_409_cast_fp16"])
           var_412_epsilon_0 = const()
0.00% N|C  var_412_cast_fp16 = ios17.rsqrt(x: ["var_411_cast_fp16"], epsilon: ["var_412_epsilon_0"])
0.01% N|C  var_413_cast_fp16 = ios17.mul(x: ["var_390_cast_fp16_1"], y: ["var_412_cast_fp16"])
           model_transformer_layers_2_attn_k_norm_weight_to_fp16 = const()
0.01% N|C  key_5_cast_fp16 = ios17.mul(y: ["model_transformer_layers_2_attn_k_norm_weight_to_fp16"], x: ["var_413_cast_fp16"])
0.02% N|C  var_430_cast_fp16 = ios17.mul(x: ["query_5_cast_fp16"], y: ["var_156_to_fp16"])
           var_431_split_sizes_0 = const()
           var_431_axis_0 = const()
0.02% N|C  var_431_cast_fp16_0, var_431_cast_fp16_1 = split(axis: ["var_431_axis_0"], split_sizes: ["var_431_split_sizes_0"], x: ["query_5_cast_fp16"])
           const_20_promoted_to_fp16 = const()
0.01% N|C  var_433_cast_fp16 = ios17.mul(y: ["const_20_promoted_to_fp16"], x: ["var_431_cast_fp16_1"])
           var_435_interleave_0 = const()
0.01% N|C  var_435_cast_fp16 = ios17.concat(axis: ["var_5"], interleave: ["var_435_interleave_0"], values: ["var_433_cast_fp16", "var_431_cast_fp16_0"])
0.02% N|C  var_436_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_435_cast_fp16"])
0.06% N|C  query_float_5_cast_fp16 = ios17.add(y: ["var_436_cast_fp16"], x: ["var_430_cast_fp16"])
0.01% N|C  var_442_cast_fp16 = ios17.mul(x: ["key_5_cast_fp16"], y: ["var_156_to_fp16"])
           var_443_split_sizes_0 = const()
           var_443_axis_0 = const()
0.00% N|C  var_443_cast_fp16_0, var_443_cast_fp16_1 = split(split_sizes: ["var_443_split_sizes_0"], x: ["key_5_cast_fp16"], axis: ["var_443_axis_0"])
           const_21_promoted_to_fp16 = const()
0.00% N|C  var_445_cast_fp16 = ios17.mul(x: ["var_443_cast_fp16_1"], y: ["const_21_promoted_to_fp16"])
           var_447_interleave_0 = const()
0.00% N|C  var_447_cast_fp16 = ios17.concat(values: ["var_445_cast_fp16", "var_443_cast_fp16_0"], interleave: ["var_447_interleave_0"], axis: ["var_5"])
0.01% N|C  var_448_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_447_cast_fp16"])
0.01% N|C  key_float_5_cast_fp16 = ios17.add(y: ["var_448_cast_fp16"], x: ["var_442_cast_fp16"])
           reshape_8_shape_0 = const()
0.01% N|C  reshape_8_cast_fp16 = ios17.reshape(x: ["key_float_5_cast_fp16"], shape: ["reshape_8_shape_0"])
           tile_4_reps_0 = const()
0.02% N|C  tile_4_cast_fp16 = tile(x: ["reshape_8_cast_fp16"], reps: ["tile_4_reps_0"])
           reshape_9_shape_0 = const()
0.03% N|C  reshape_9_cast_fp16 = ios17.reshape(x: ["tile_4_cast_fp16"], shape: ["reshape_9_shape_0"])
           transpose_4_perm_0 = const()
           keys_11_shape_0 = const()
0.05% N|C  transpose_86 = ios17.transpose(perm: ["transpose_4_perm_0"], x: ["reshape_9_cast_fp16"])
0.03% N|C  keys_11_cast_fp16 = ios17.reshape(x: ["transpose_86"], shape: ["keys_11_shape_0"])
           reshape_10_shape_0 = const()
0.01% N|C  reshape_10_cast_fp16 = ios17.reshape(x: ["var_390_cast_fp16_2"], shape: ["reshape_10_shape_0"])
           tile_5_reps_0 = const()
0.02% N|C  tile_5_cast_fp16 = tile(reps: ["tile_5_reps_0"], x: ["reshape_10_cast_fp16"])
           reshape_11_shape_0 = const()
0.03% N|C  reshape_11_cast_fp16 = ios17.reshape(x: ["tile_5_cast_fp16"], shape: ["reshape_11_shape_0"])
           transpose_5_perm_0 = const()
           values_11_shape_0 = const()
0.05% N|C  transpose_85 = ios17.transpose(x: ["reshape_11_cast_fp16"], perm: ["transpose_5_perm_0"])
0.03% N|C  values_11_cast_fp16 = ios17.reshape(shape: ["values_11_shape_0"], x: ["transpose_85"])
           mul_2_y_0_to_fp16 = const()
0.02% N|C  mul_2_cast_fp16 = ios17.mul(y: ["mul_2_y_0_to_fp16"], x: ["query_float_5_cast_fp16"])
           matmul_2_transpose_y_0 = const()
           matmul_2_transpose_x_0 = const()
0.05% N|C  matmul_2_cast_fp16 = ios17.matmul(transpose_y: ["matmul_2_transpose_y_0"], transpose_x: ["matmul_2_transpose_x_0"], y: ["keys_11_cast_fp16"], x: ["mul_2_cast_fp16"])
0.06% N|C  add_5_cast_fp16 = ios17.add(x: ["matmul_2_cast_fp16"], y: ["causal_mask_3_to_fp16"])
           softmax_2_axis_0 = const()
0.05% N|C  softmax_2_cast_fp16 = ios16.softmax(axis: ["softmax_2_axis_0"], x: ["add_5_cast_fp16"])
           attn_output_9_transpose_x_0 = const()
           attn_output_9_transpose_y_0 = const()
0.05% N|C  attn_output_9_cast_fp16 = ios17.matmul(transpose_y: ["attn_output_9_transpose_y_0"], x: ["softmax_2_cast_fp16"], transpose_x: ["attn_output_9_transpose_x_0"], y: ["values_11_cast_fp16"])
           var_464_perm_0 = const()
           var_466 = const()
0.05% N|C  transpose_84 = ios17.transpose(perm: ["var_464_perm_0"], x: ["attn_output_9_cast_fp16"])
0.03% N|C  input_19_cast_fp16 = ios17.reshape(shape: ["var_466"], x: ["transpose_84"])
           model_transformer_layers_2_attn_out_proj_weight_to_fp16 = const()
0.14% N|C  linear_9_cast_fp16 = ios17.linear(x: ["input_19_cast_fp16"], weight: ["model_transformer_layers_2_attn_out_proj_weight_to_fp16"], bias: ["linear_1_bias_0_to_fp16"])
0.09% N|C  x_55_cast_fp16 = ios17.add(x: ["x_39_cast_fp16"], y: ["linear_9_cast_fp16"])
           var_10_promoted_to_fp16_11 = const()
0.04% N|C  var_473_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_11"], x: ["x_55_cast_fp16"])
           var_474 = const()
0.02% N|C  var_475_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], axes: ["var_474"], x: ["var_473_cast_fp16"])
           var_476_to_fp16 = const()
0.00% N|C  var_477_cast_fp16 = ios17.add(x: ["var_475_cast_fp16"], y: ["var_476_to_fp16"])
           var_478_epsilon_0 = const()
0.00% N|C  var_478_cast_fp16 = ios17.rsqrt(x: ["var_477_cast_fp16"], epsilon: ["var_478_epsilon_0"])
0.04% N|C  var_479_cast_fp16 = ios17.mul(y: ["var_478_cast_fp16"], x: ["x_55_cast_fp16"])
           model_transformer_layers_2_ffn_norm_weight_to_fp16 = const()
0.04% N|C  input_21_cast_fp16 = ios17.mul(x: ["var_479_cast_fp16"], y: ["model_transformer_layers_2_ffn_norm_weight_to_fp16"])
           model_transformer_layers_2_ffn_proj_1_weight_to_fp16 = const()
           linear_10_bias_0_to_fp16 = const()
0.15% N|C  linear_10_cast_fp16 = ios17.linear(bias: ["linear_10_bias_0_to_fp16"], x: ["input_21_cast_fp16"], weight: ["model_transformer_layers_2_ffn_proj_1_weight_to_fp16"])
           var_489_split_sizes_0 = const()
           var_489_axis_0 = const()
0.06% N|C  var_489_cast_fp16_0, var_489_cast_fp16_1 = split(axis: ["var_489_axis_0"], x: ["linear_10_cast_fp16"], split_sizes: ["var_489_split_sizes_0"])
0.04% N|C  var_491_cast_fp16 = ios16.silu(x: ["var_489_cast_fp16_0"])
0.08% N|C  input_25_cast_fp16 = ios17.mul(y: ["var_489_cast_fp16_1"], x: ["var_491_cast_fp16"])
           model_transformer_layers_2_ffn_proj_2_weight_to_fp16 = const()
0.38% N|C  linear_11_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_2_ffn_proj_2_weight_to_fp16"], bias: ["linear_1_bias_0_to_fp16"], x: ["input_25_cast_fp16"])
0.09% N|C  x_59_cast_fp16 = ios17.add(x: ["x_55_cast_fp16"], y: ["linear_11_cast_fp16"])
           var_10_promoted_to_fp16_12 = const()
0.04% N|C  var_502_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_12"], x: ["x_59_cast_fp16"])
           var_503 = const()
0.02% N|C  var_504_cast_fp16 = ios16.reduce_mean(axes: ["var_503"], keep_dims: ["var_22"], x: ["var_502_cast_fp16"])
           var_505_to_fp16 = const()
0.00% N|C  var_506_cast_fp16 = ios17.add(y: ["var_505_to_fp16"], x: ["var_504_cast_fp16"])
           var_507_epsilon_0 = const()
0.00% N|C  var_507_cast_fp16 = ios17.rsqrt(x: ["var_506_cast_fp16"], epsilon: ["var_507_epsilon_0"])
0.04% N|C  var_508_cast_fp16 = ios17.mul(y: ["var_507_cast_fp16"], x: ["x_59_cast_fp16"])
           model_transformer_layers_3_attn_norm_weight_to_fp16 = const()
0.04% N|C  hidden_states_19_cast_fp16 = ios17.mul(y: ["model_transformer_layers_3_attn_norm_weight_to_fp16"], x: ["var_508_cast_fp16"])
           model_transformer_layers_3_attn_qkv_proj_weight_to_fp16 = const()
0.38% N|C  linear_12_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_3_attn_qkv_proj_weight_to_fp16"], x: ["hidden_states_19_cast_fp16"], bias: ["linear_0_bias_0_to_fp16"])
           var_522 = const()
0.04% N|C  qkv_15_cast_fp16 = ios17.reshape(shape: ["var_522"], x: ["linear_12_cast_fp16"])
           var_524_perm_0 = const()
           var_525 = const()
           var_526_axis_0 = const()
0.08% N|C  transpose_83 = ios17.transpose(x: ["qkv_15_cast_fp16"], perm: ["var_524_perm_0"])
0.03% N|C  var_526_cast_fp16_0, var_526_cast_fp16_1, var_526_cast_fp16_2 = split(x: ["transpose_83"], axis: ["var_526_axis_0"], split_sizes: ["var_525"])
           var_10_promoted_to_fp16_13 = const()
0.03% N|C  var_532_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_13"], x: ["var_526_cast_fp16_0"])
           var_533 = const()
0.01% N|C  var_534_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_532_cast_fp16"], axes: ["var_533"])
           var_535_to_fp16 = const()
0.00% N|C  var_536_cast_fp16 = ios17.add(y: ["var_535_to_fp16"], x: ["var_534_cast_fp16"])
           var_537_epsilon_0 = const()
0.00% N|C  var_537_cast_fp16 = ios17.rsqrt(x: ["var_536_cast_fp16"], epsilon: ["var_537_epsilon_0"])
0.02% N|C  var_538_cast_fp16 = ios17.mul(x: ["var_526_cast_fp16_0"], y: ["var_537_cast_fp16"])
           model_transformer_layers_3_attn_q_norm_weight_to_fp16 = const()
0.02% N|C  query_7_cast_fp16 = ios17.mul(y: ["model_transformer_layers_3_attn_q_norm_weight_to_fp16"], x: ["var_538_cast_fp16"])
           var_10_promoted_to_fp16_14 = const()
0.01% N|C  var_543_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_14"], x: ["var_526_cast_fp16_1"])
           var_544 = const()
0.00% N|C  var_545_cast_fp16 = ios16.reduce_mean(x: ["var_543_cast_fp16"], keep_dims: ["var_22"], axes: ["var_544"])
           var_546_to_fp16 = const()
0.00% N|C  var_547_cast_fp16 = ios17.add(x: ["var_545_cast_fp16"], y: ["var_546_to_fp16"])
           var_548_epsilon_0 = const()
0.00% N|C  var_548_cast_fp16 = ios17.rsqrt(x: ["var_547_cast_fp16"], epsilon: ["var_548_epsilon_0"])
0.01% N|C  var_549_cast_fp16 = ios17.mul(y: ["var_548_cast_fp16"], x: ["var_526_cast_fp16_1"])
           model_transformer_layers_3_attn_k_norm_weight_to_fp16 = const()
0.01% N|C  key_7_cast_fp16 = ios17.mul(y: ["model_transformer_layers_3_attn_k_norm_weight_to_fp16"], x: ["var_549_cast_fp16"])
0.02% N|C  var_566_cast_fp16 = ios17.mul(x: ["query_7_cast_fp16"], y: ["var_156_to_fp16"])
           var_567_split_sizes_0 = const()
           var_567_axis_0 = const()
0.02% N|C  var_567_cast_fp16_0, var_567_cast_fp16_1 = split(x: ["query_7_cast_fp16"], axis: ["var_567_axis_0"], split_sizes: ["var_567_split_sizes_0"])
           const_27_promoted_to_fp16 = const()
0.01% N|C  var_569_cast_fp16 = ios17.mul(x: ["var_567_cast_fp16_1"], y: ["const_27_promoted_to_fp16"])
           var_571_interleave_0 = const()
0.01% N|C  var_571_cast_fp16 = ios17.concat(values: ["var_569_cast_fp16", "var_567_cast_fp16_0"], axis: ["var_5"], interleave: ["var_571_interleave_0"])
0.02% N|C  var_572_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_571_cast_fp16"])
0.06% N|C  query_float_7_cast_fp16 = ios17.add(y: ["var_572_cast_fp16"], x: ["var_566_cast_fp16"])
0.01% N|C  var_578_cast_fp16 = ios17.mul(x: ["key_7_cast_fp16"], y: ["var_156_to_fp16"])
           var_579_split_sizes_0 = const()
           var_579_axis_0 = const()
0.00% N|C  var_579_cast_fp16_0, var_579_cast_fp16_1 = split(axis: ["var_579_axis_0"], x: ["key_7_cast_fp16"], split_sizes: ["var_579_split_sizes_0"])
           const_28_promoted_to_fp16 = const()
0.00% N|C  var_581_cast_fp16 = ios17.mul(y: ["const_28_promoted_to_fp16"], x: ["var_579_cast_fp16_1"])
           var_583_interleave_0 = const()
0.00% N|C  var_583_cast_fp16 = ios17.concat(values: ["var_581_cast_fp16", "var_579_cast_fp16_0"], axis: ["var_5"], interleave: ["var_583_interleave_0"])
0.01% N|C  var_584_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_583_cast_fp16"])
0.01% N|C  key_float_7_cast_fp16 = ios17.add(x: ["var_578_cast_fp16"], y: ["var_584_cast_fp16"])
           reshape_12_shape_0 = const()
0.01% N|C  reshape_12_cast_fp16 = ios17.reshape(x: ["key_float_7_cast_fp16"], shape: ["reshape_12_shape_0"])
           tile_6_reps_0 = const()
0.02% N|C  tile_6_cast_fp16 = tile(x: ["reshape_12_cast_fp16"], reps: ["tile_6_reps_0"])
           reshape_13_shape_0 = const()
0.03% N|C  reshape_13_cast_fp16 = ios17.reshape(x: ["tile_6_cast_fp16"], shape: ["reshape_13_shape_0"])
           transpose_6_perm_0 = const()
           keys_15_shape_0 = const()
0.05% N|C  transpose_82 = ios17.transpose(x: ["reshape_13_cast_fp16"], perm: ["transpose_6_perm_0"])
0.03% N|C  keys_15_cast_fp16 = ios17.reshape(x: ["transpose_82"], shape: ["keys_15_shape_0"])
           reshape_14_shape_0 = const()
0.01% N|C  reshape_14_cast_fp16 = ios17.reshape(shape: ["reshape_14_shape_0"], x: ["var_526_cast_fp16_2"])
           tile_7_reps_0 = const()
0.02% N|C  tile_7_cast_fp16 = tile(reps: ["tile_7_reps_0"], x: ["reshape_14_cast_fp16"])
           reshape_15_shape_0 = const()
0.03% N|C  reshape_15_cast_fp16 = ios17.reshape(shape: ["reshape_15_shape_0"], x: ["tile_7_cast_fp16"])
           transpose_7_perm_0 = const()
           values_15_shape_0 = const()
0.05% N|C  transpose_81 = ios17.transpose(perm: ["transpose_7_perm_0"], x: ["reshape_15_cast_fp16"])
0.03% N|C  values_15_cast_fp16 = ios17.reshape(x: ["transpose_81"], shape: ["values_15_shape_0"])
           mul_3_y_0_to_fp16 = const()
0.02% N|C  mul_3_cast_fp16 = ios17.mul(y: ["mul_3_y_0_to_fp16"], x: ["query_float_7_cast_fp16"])
           matmul_3_transpose_y_0 = const()
           matmul_3_transpose_x_0 = const()
0.05% N|C  matmul_3_cast_fp16 = ios17.matmul(y: ["keys_15_cast_fp16"], transpose_x: ["matmul_3_transpose_x_0"], transpose_y: ["matmul_3_transpose_y_0"], x: ["mul_3_cast_fp16"])
0.06% N|C  add_7_cast_fp16 = ios17.add(x: ["matmul_3_cast_fp16"], y: ["causal_mask_3_to_fp16"])
           softmax_3_axis_0 = const()
0.05% N|C  softmax_3_cast_fp16 = ios16.softmax(axis: ["softmax_3_axis_0"], x: ["add_7_cast_fp16"])
           attn_output_13_transpose_x_0 = const()
           attn_output_13_transpose_y_0 = const()
0.05% N|C  attn_output_13_cast_fp16 = ios17.matmul(y: ["values_15_cast_fp16"], x: ["softmax_3_cast_fp16"], transpose_y: ["attn_output_13_transpose_y_0"], transpose_x: ["attn_output_13_transpose_x_0"])
           var_600_perm_0 = const()
           var_602 = const()
0.05% N|C  transpose_80 = ios17.transpose(x: ["attn_output_13_cast_fp16"], perm: ["var_600_perm_0"])
0.03% N|C  input_27_cast_fp16 = ios17.reshape(x: ["transpose_80"], shape: ["var_602"])
           model_transformer_layers_3_attn_out_proj_weight_to_fp16 = const()
0.14% N|C  linear_13_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_3_attn_out_proj_weight_to_fp16"], x: ["input_27_cast_fp16"], bias: ["linear_1_bias_0_to_fp16"])
0.09% N|C  x_75_cast_fp16 = ios17.add(x: ["x_59_cast_fp16"], y: ["linear_13_cast_fp16"])
           var_10_promoted_to_fp16_15 = const()
0.04% N|C  var_609_cast_fp16 = ios17.pow(x: ["x_75_cast_fp16"], y: ["var_10_promoted_to_fp16_15"])
           var_610 = const()
0.02% N|C  var_611_cast_fp16 = ios16.reduce_mean(x: ["var_609_cast_fp16"], keep_dims: ["var_22"], axes: ["var_610"])
           var_612_to_fp16 = const()
0.00% N|C  var_613_cast_fp16 = ios17.add(x: ["var_611_cast_fp16"], y: ["var_612_to_fp16"])
           var_614_epsilon_0 = const()
0.00% N|C  var_614_cast_fp16 = ios17.rsqrt(epsilon: ["var_614_epsilon_0"], x: ["var_613_cast_fp16"])
0.04% N|C  var_615_cast_fp16 = ios17.mul(x: ["x_75_cast_fp16"], y: ["var_614_cast_fp16"])
           model_transformer_layers_3_ffn_norm_weight_to_fp16 = const()
0.04% N|C  input_29_cast_fp16 = ios17.mul(x: ["var_615_cast_fp16"], y: ["model_transformer_layers_3_ffn_norm_weight_to_fp16"])
           model_transformer_layers_3_ffn_proj_1_weight_to_fp16 = const()
           linear_14_bias_0_to_fp16 = const()
0.15% N|C  linear_14_cast_fp16 = ios17.linear(bias: ["linear_14_bias_0_to_fp16"], weight: ["model_transformer_layers_3_ffn_proj_1_weight_to_fp16"], x: ["input_29_cast_fp16"])
           var_625_split_sizes_0 = const()
           var_625_axis_0 = const()
0.08% N|C  var_625_cast_fp16_0, var_625_cast_fp16_1 = split(x: ["linear_14_cast_fp16"], axis: ["var_625_axis_0"], split_sizes: ["var_625_split_sizes_0"])
0.05% N|C  var_627_cast_fp16 = ios16.silu(x: ["var_625_cast_fp16_0"])
0.09% N|C  input_33_cast_fp16 = ios17.mul(x: ["var_627_cast_fp16"], y: ["var_625_cast_fp16_1"])
           model_transformer_layers_3_ffn_proj_2_weight_to_fp16 = const()
0.54% N|C  linear_15_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_3_ffn_proj_2_weight_to_fp16"], bias: ["linear_1_bias_0_to_fp16"], x: ["input_33_cast_fp16"])
0.09% N|C  x_79_cast_fp16 = ios17.add(y: ["linear_15_cast_fp16"], x: ["x_75_cast_fp16"])
           var_10_promoted_to_fp16_16 = const()
0.04% N|C  var_638_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_16"], x: ["x_79_cast_fp16"])
           var_639 = const()
0.02% N|C  var_640_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_638_cast_fp16"], axes: ["var_639"])
           var_641_to_fp16 = const()
0.00% N|C  var_642_cast_fp16 = ios17.add(y: ["var_641_to_fp16"], x: ["var_640_cast_fp16"])
           var_643_epsilon_0 = const()
0.00% N|C  var_643_cast_fp16 = ios17.rsqrt(x: ["var_642_cast_fp16"], epsilon: ["var_643_epsilon_0"])
0.04% N|C  var_644_cast_fp16 = ios17.mul(y: ["var_643_cast_fp16"], x: ["x_79_cast_fp16"])
           model_transformer_layers_4_attn_norm_weight_to_fp16 = const()
0.04% N|C  hidden_states_25_cast_fp16 = ios17.mul(x: ["var_644_cast_fp16"], y: ["model_transformer_layers_4_attn_norm_weight_to_fp16"])
           model_transformer_layers_4_attn_qkv_proj_weight_to_fp16 = const()
0.38% N|C  linear_16_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_4_attn_qkv_proj_weight_to_fp16"], x: ["hidden_states_25_cast_fp16"], bias: ["linear_0_bias_0_to_fp16"])
           var_658 = const()
0.04% N|C  qkv_19_cast_fp16 = ios17.reshape(x: ["linear_16_cast_fp16"], shape: ["var_658"])
           var_660_perm_0 = const()
           var_661 = const()
           var_662_axis_0 = const()
0.08% N|C  transpose_79 = ios17.transpose(perm: ["var_660_perm_0"], x: ["qkv_19_cast_fp16"])
0.03% N|C  var_662_cast_fp16_0, var_662_cast_fp16_1, var_662_cast_fp16_2 = split(axis: ["var_662_axis_0"], x: ["transpose_79"], split_sizes: ["var_661"])
           var_10_promoted_to_fp16_17 = const()
0.03% N|C  var_668_cast_fp16 = ios17.pow(x: ["var_662_cast_fp16_0"], y: ["var_10_promoted_to_fp16_17"])
           var_669 = const()
0.01% N|C  var_670_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], axes: ["var_669"], x: ["var_668_cast_fp16"])
           var_671_to_fp16 = const()
0.00% N|C  var_672_cast_fp16 = ios17.add(y: ["var_671_to_fp16"], x: ["var_670_cast_fp16"])
           var_673_epsilon_0 = const()
0.00% N|C  var_673_cast_fp16 = ios17.rsqrt(x: ["var_672_cast_fp16"], epsilon: ["var_673_epsilon_0"])
0.02% N|C  var_674_cast_fp16 = ios17.mul(x: ["var_662_cast_fp16_0"], y: ["var_673_cast_fp16"])
           model_transformer_layers_4_attn_q_norm_weight_to_fp16 = const()
0.02% N|C  query_9_cast_fp16 = ios17.mul(y: ["model_transformer_layers_4_attn_q_norm_weight_to_fp16"], x: ["var_674_cast_fp16"])
           var_10_promoted_to_fp16_18 = const()
0.01% N|C  var_679_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_18"], x: ["var_662_cast_fp16_1"])
           var_680 = const()
0.00% N|C  var_681_cast_fp16 = ios16.reduce_mean(x: ["var_679_cast_fp16"], axes: ["var_680"], keep_dims: ["var_22"])
           var_682_to_fp16 = const()
0.00% N|C  var_683_cast_fp16 = ios17.add(y: ["var_682_to_fp16"], x: ["var_681_cast_fp16"])
           var_684_epsilon_0 = const()
0.00% N|C  var_684_cast_fp16 = ios17.rsqrt(x: ["var_683_cast_fp16"], epsilon: ["var_684_epsilon_0"])
0.01% N|C  var_685_cast_fp16 = ios17.mul(x: ["var_662_cast_fp16_1"], y: ["var_684_cast_fp16"])
           model_transformer_layers_4_attn_k_norm_weight_to_fp16 = const()
0.01% N|C  key_9_cast_fp16 = ios17.mul(x: ["var_685_cast_fp16"], y: ["model_transformer_layers_4_attn_k_norm_weight_to_fp16"])
0.02% N|C  var_702_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["query_9_cast_fp16"])
           var_703_split_sizes_0 = const()
           var_703_axis_0 = const()
0.02% N|C  var_703_cast_fp16_0, var_703_cast_fp16_1 = split(split_sizes: ["var_703_split_sizes_0"], x: ["query_9_cast_fp16"], axis: ["var_703_axis_0"])
           const_34_promoted_to_fp16 = const()
0.01% N|C  var_705_cast_fp16 = ios17.mul(y: ["const_34_promoted_to_fp16"], x: ["var_703_cast_fp16_1"])
           var_707_interleave_0 = const()
0.01% N|C  var_707_cast_fp16 = ios17.concat(values: ["var_705_cast_fp16", "var_703_cast_fp16_0"], interleave: ["var_707_interleave_0"], axis: ["var_5"])
0.02% N|C  var_708_cast_fp16 = ios17.mul(x: ["var_707_cast_fp16"], y: ["var_152_to_fp16"])
0.06% N|C  query_float_9_cast_fp16 = ios17.add(y: ["var_708_cast_fp16"], x: ["var_702_cast_fp16"])
0.01% N|C  var_714_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["key_9_cast_fp16"])
           var_715_split_sizes_0 = const()
           var_715_axis_0 = const()
0.00% N|C  var_715_cast_fp16_0, var_715_cast_fp16_1 = split(axis: ["var_715_axis_0"], x: ["key_9_cast_fp16"], split_sizes: ["var_715_split_sizes_0"])
           const_35_promoted_to_fp16 = const()
0.00% N|C  var_717_cast_fp16 = ios17.mul(x: ["var_715_cast_fp16_1"], y: ["const_35_promoted_to_fp16"])
           var_719_interleave_0 = const()
0.00% N|C  var_719_cast_fp16 = ios17.concat(values: ["var_717_cast_fp16", "var_715_cast_fp16_0"], axis: ["var_5"], interleave: ["var_719_interleave_0"])
0.01% N|C  var_720_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_719_cast_fp16"])
0.01% N|C  key_float_9_cast_fp16 = ios17.add(x: ["var_714_cast_fp16"], y: ["var_720_cast_fp16"])
           reshape_16_shape_0 = const()
0.01% N|C  reshape_16_cast_fp16 = ios17.reshape(shape: ["reshape_16_shape_0"], x: ["key_float_9_cast_fp16"])
           tile_8_reps_0 = const()
0.02% N|C  tile_8_cast_fp16 = tile(x: ["reshape_16_cast_fp16"], reps: ["tile_8_reps_0"])
           reshape_17_shape_0 = const()
0.03% N|C  reshape_17_cast_fp16 = ios17.reshape(x: ["tile_8_cast_fp16"], shape: ["reshape_17_shape_0"])
           transpose_8_perm_0 = const()
           keys_19_shape_0 = const()
0.05% N|C  transpose_78 = ios17.transpose(perm: ["transpose_8_perm_0"], x: ["reshape_17_cast_fp16"])
0.03% N|C  keys_19_cast_fp16 = ios17.reshape(x: ["transpose_78"], shape: ["keys_19_shape_0"])
           reshape_18_shape_0 = const()
0.01% N|C  reshape_18_cast_fp16 = ios17.reshape(shape: ["reshape_18_shape_0"], x: ["var_662_cast_fp16_2"])
           tile_9_reps_0 = const()
0.02% N|C  tile_9_cast_fp16 = tile(x: ["reshape_18_cast_fp16"], reps: ["tile_9_reps_0"])
           reshape_19_shape_0 = const()
0.03% N|C  reshape_19_cast_fp16 = ios17.reshape(shape: ["reshape_19_shape_0"], x: ["tile_9_cast_fp16"])
           transpose_9_perm_0 = const()
           values_19_shape_0 = const()
0.05% N|C  transpose_77 = ios17.transpose(x: ["reshape_19_cast_fp16"], perm: ["transpose_9_perm_0"])
0.03% N|C  values_19_cast_fp16 = ios17.reshape(shape: ["values_19_shape_0"], x: ["transpose_77"])
           mul_4_y_0_to_fp16 = const()
0.02% N|C  mul_4_cast_fp16 = ios17.mul(y: ["mul_4_y_0_to_fp16"], x: ["query_float_9_cast_fp16"])
           matmul_4_transpose_y_0 = const()
           matmul_4_transpose_x_0 = const()
0.05% N|C  matmul_4_cast_fp16 = ios17.matmul(x: ["mul_4_cast_fp16"], transpose_y: ["matmul_4_transpose_y_0"], transpose_x: ["matmul_4_transpose_x_0"], y: ["keys_19_cast_fp16"])
0.06% N|C  add_9_cast_fp16 = ios17.add(x: ["matmul_4_cast_fp16"], y: ["causal_mask_3_to_fp16"])
           softmax_4_axis_0 = const()
0.05% N|C  softmax_4_cast_fp16 = ios16.softmax(axis: ["softmax_4_axis_0"], x: ["add_9_cast_fp16"])
           attn_output_17_transpose_x_0 = const()
           attn_output_17_transpose_y_0 = const()
0.05% N|C  attn_output_17_cast_fp16 = ios17.matmul(x: ["softmax_4_cast_fp16"], y: ["values_19_cast_fp16"], transpose_x: ["attn_output_17_transpose_x_0"], transpose_y: ["attn_output_17_transpose_y_0"])
           var_736_perm_0 = const()
           var_738 = const()
0.05% N|C  transpose_76 = ios17.transpose(x: ["attn_output_17_cast_fp16"], perm: ["var_736_perm_0"])
0.03% N|C  input_35_cast_fp16 = ios17.reshape(shape: ["var_738"], x: ["transpose_76"])
           model_transformer_layers_4_attn_out_proj_weight_to_fp16 = const()
0.14% N|C  linear_17_cast_fp16 = ios17.linear(x: ["input_35_cast_fp16"], bias: ["linear_1_bias_0_to_fp16"], weight: ["model_transformer_layers_4_attn_out_proj_weight_to_fp16"])
0.09% N|C  x_95_cast_fp16 = ios17.add(y: ["linear_17_cast_fp16"], x: ["x_79_cast_fp16"])
           var_10_promoted_to_fp16_19 = const()
0.04% N|C  var_745_cast_fp16 = ios17.pow(x: ["x_95_cast_fp16"], y: ["var_10_promoted_to_fp16_19"])
           var_746 = const()
0.02% N|C  var_747_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_745_cast_fp16"], axes: ["var_746"])
           var_748_to_fp16 = const()
0.00% N|C  var_749_cast_fp16 = ios17.add(x: ["var_747_cast_fp16"], y: ["var_748_to_fp16"])
           var_750_epsilon_0 = const()
0.00% N|C  var_750_cast_fp16 = ios17.rsqrt(x: ["var_749_cast_fp16"], epsilon: ["var_750_epsilon_0"])
0.04% N|C  var_751_cast_fp16 = ios17.mul(y: ["var_750_cast_fp16"], x: ["x_95_cast_fp16"])
           model_transformer_layers_4_ffn_norm_weight_to_fp16 = const()
0.04% N|C  input_37_cast_fp16 = ios17.mul(y: ["model_transformer_layers_4_ffn_norm_weight_to_fp16"], x: ["var_751_cast_fp16"])
           model_transformer_layers_4_ffn_proj_1_weight_to_fp16 = const()
           linear_18_bias_0_to_fp16 = const()
0.15% N|C  linear_18_cast_fp16 = ios17.linear(bias: ["linear_18_bias_0_to_fp16"], x: ["input_37_cast_fp16"], weight: ["model_transformer_layers_4_ffn_proj_1_weight_to_fp16"])
           var_761_split_sizes_0 = const()
           var_761_axis_0 = const()
0.09% N|C  var_761_cast_fp16_0, var_761_cast_fp16_1 = split(axis: ["var_761_axis_0"], x: ["linear_18_cast_fp16"], split_sizes: ["var_761_split_sizes_0"])
0.06% N|C  var_763_cast_fp16 = ios16.silu(x: ["var_761_cast_fp16_0"])
0.11% N|C  input_41_cast_fp16 = ios17.mul(x: ["var_763_cast_fp16"], y: ["var_761_cast_fp16_1"])
           model_transformer_layers_4_ffn_proj_2_weight_to_fp16 = const()
0.74% N|C  linear_19_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_4_ffn_proj_2_weight_to_fp16"], bias: ["linear_1_bias_0_to_fp16"], x: ["input_41_cast_fp16"])
0.09% N|C  x_99_cast_fp16 = ios17.add(y: ["linear_19_cast_fp16"], x: ["x_95_cast_fp16"])
           var_10_promoted_to_fp16_20 = const()
0.04% N|C  var_774_cast_fp16 = ios17.pow(x: ["x_99_cast_fp16"], y: ["var_10_promoted_to_fp16_20"])
           var_775 = const()
0.02% N|C  var_776_cast_fp16 = ios16.reduce_mean(x: ["var_774_cast_fp16"], keep_dims: ["var_22"], axes: ["var_775"])
           var_777_to_fp16 = const()
0.00% N|C  var_778_cast_fp16 = ios17.add(y: ["var_777_to_fp16"], x: ["var_776_cast_fp16"])
           var_779_epsilon_0 = const()
0.00% N|C  var_779_cast_fp16 = ios17.rsqrt(x: ["var_778_cast_fp16"], epsilon: ["var_779_epsilon_0"])
0.04% N|C  var_780_cast_fp16 = ios17.mul(x: ["x_99_cast_fp16"], y: ["var_779_cast_fp16"])
           model_transformer_layers_5_attn_norm_weight_to_fp16 = const()
0.04% N|C  hidden_states_31_cast_fp16 = ios17.mul(x: ["var_780_cast_fp16"], y: ["model_transformer_layers_5_attn_norm_weight_to_fp16"])
           model_transformer_layers_5_attn_qkv_proj_weight_to_fp16 = const()
0.38% N|C  linear_20_cast_fp16 = ios17.linear(bias: ["linear_2_bias_0_to_fp16"], x: ["hidden_states_31_cast_fp16"], weight: ["model_transformer_layers_5_attn_qkv_proj_weight_to_fp16"])
           var_794 = const()
0.05% N|C  qkv_23_cast_fp16 = ios17.reshape(shape: ["var_794"], x: ["linear_20_cast_fp16"])
           var_796_perm_0 = const()
           var_797 = const()
           var_798_axis_0 = const()
0.11% N|C  transpose_75 = ios17.transpose(x: ["qkv_23_cast_fp16"], perm: ["var_796_perm_0"])
0.04% N|C  var_798_cast_fp16_0, var_798_cast_fp16_1, var_798_cast_fp16_2 = split(split_sizes: ["var_797"], axis: ["var_798_axis_0"], x: ["transpose_75"])
           var_10_promoted_to_fp16_21 = const()
0.03% N|C  var_804_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_21"], x: ["var_798_cast_fp16_0"])
           var_805 = const()
0.02% N|C  var_806_cast_fp16 = ios16.reduce_mean(x: ["var_804_cast_fp16"], keep_dims: ["var_22"], axes: ["var_805"])
           var_807_to_fp16 = const()
0.00% N|C  var_808_cast_fp16 = ios17.add(y: ["var_807_to_fp16"], x: ["var_806_cast_fp16"])
           var_809_epsilon_0 = const()
0.00% N|C  var_809_cast_fp16 = ios17.rsqrt(x: ["var_808_cast_fp16"], epsilon: ["var_809_epsilon_0"])
0.03% N|C  var_810_cast_fp16 = ios17.mul(y: ["var_809_cast_fp16"], x: ["var_798_cast_fp16_0"])
           model_transformer_layers_5_attn_q_norm_weight_to_fp16 = const()
0.03% N|C  query_11_cast_fp16 = ios17.mul(y: ["model_transformer_layers_5_attn_q_norm_weight_to_fp16"], x: ["var_810_cast_fp16"])
           var_10_promoted_to_fp16_22 = const()
0.01% N|C  var_815_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_22"], x: ["var_798_cast_fp16_1"])
           var_816 = const()
0.00% N|C  var_817_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], axes: ["var_816"], x: ["var_815_cast_fp16"])
           var_818_to_fp16 = const()
0.00% N|C  var_819_cast_fp16 = ios17.add(y: ["var_818_to_fp16"], x: ["var_817_cast_fp16"])
           var_820_epsilon_0 = const()
0.00% N|C  var_820_cast_fp16 = ios17.rsqrt(x: ["var_819_cast_fp16"], epsilon: ["var_820_epsilon_0"])
0.01% N|C  var_821_cast_fp16 = ios17.mul(x: ["var_798_cast_fp16_1"], y: ["var_820_cast_fp16"])
           model_transformer_layers_5_attn_k_norm_weight_to_fp16 = const()
0.01% N|C  key_11_cast_fp16 = ios17.mul(y: ["model_transformer_layers_5_attn_k_norm_weight_to_fp16"], x: ["var_821_cast_fp16"])
0.03% N|C  var_838_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["query_11_cast_fp16"])
           var_839_split_sizes_0 = const()
           var_839_axis_0 = const()
0.03% N|C  var_839_cast_fp16_0, var_839_cast_fp16_1 = split(x: ["query_11_cast_fp16"], split_sizes: ["var_839_split_sizes_0"], axis: ["var_839_axis_0"])
           const_41_promoted_to_fp16 = const()
0.02% N|C  var_841_cast_fp16 = ios17.mul(x: ["var_839_cast_fp16_1"], y: ["const_41_promoted_to_fp16"])
           var_843_interleave_0 = const()
0.01% N|C  var_843_cast_fp16 = ios17.concat(interleave: ["var_843_interleave_0"], axis: ["var_5"], values: ["var_841_cast_fp16", "var_839_cast_fp16_0"])
0.03% N|C  var_844_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_843_cast_fp16"])
0.08% N|C  query_float_11_cast_fp16 = ios17.add(x: ["var_838_cast_fp16"], y: ["var_844_cast_fp16"])
0.01% N|C  var_850_cast_fp16 = ios17.mul(x: ["key_11_cast_fp16"], y: ["var_156_to_fp16"])
           var_851_split_sizes_0 = const()
           var_851_axis_0 = const()
0.01% N|C  var_851_cast_fp16_0, var_851_cast_fp16_1 = split(x: ["key_11_cast_fp16"], axis: ["var_851_axis_0"], split_sizes: ["var_851_split_sizes_0"])
           const_42_promoted_to_fp16 = const()
0.00% N|C  var_853_cast_fp16 = ios17.mul(y: ["const_42_promoted_to_fp16"], x: ["var_851_cast_fp16_1"])
           var_855_interleave_0 = const()
0.00% N|C  var_855_cast_fp16 = ios17.concat(interleave: ["var_855_interleave_0"], values: ["var_853_cast_fp16", "var_851_cast_fp16_0"], axis: ["var_5"])
0.01% N|C  var_856_cast_fp16 = ios17.mul(x: ["var_855_cast_fp16"], y: ["var_152_to_fp16"])
0.02% N|C  key_float_11_cast_fp16 = ios17.add(x: ["var_850_cast_fp16"], y: ["var_856_cast_fp16"])
           reshape_20_shape_0 = const()
0.01% N|C  reshape_20_cast_fp16 = ios17.reshape(shape: ["reshape_20_shape_0"], x: ["key_float_11_cast_fp16"])
           tile_10_reps_0 = const()
0.02% N|C  tile_10_cast_fp16 = tile(x: ["reshape_20_cast_fp16"], reps: ["tile_10_reps_0"])
           reshape_21_shape_0 = const()
0.03% N|C  reshape_21_cast_fp16 = ios17.reshape(x: ["tile_10_cast_fp16"], shape: ["reshape_21_shape_0"])
           transpose_10_perm_0 = const()
           keys_23_shape_0 = const()
0.07% N|C  transpose_74 = ios17.transpose(x: ["reshape_21_cast_fp16"], perm: ["transpose_10_perm_0"])
0.03% N|C  keys_23_cast_fp16 = ios17.reshape(shape: ["keys_23_shape_0"], x: ["transpose_74"])
           reshape_22_shape_0 = const()
0.01% N|C  reshape_22_cast_fp16 = ios17.reshape(x: ["var_798_cast_fp16_2"], shape: ["reshape_22_shape_0"])
           tile_11_reps_0 = const()
0.02% N|C  tile_11_cast_fp16 = tile(x: ["reshape_22_cast_fp16"], reps: ["tile_11_reps_0"])
           reshape_23_shape_0 = const()
0.03% N|C  reshape_23_cast_fp16 = ios17.reshape(x: ["tile_11_cast_fp16"], shape: ["reshape_23_shape_0"])
           transpose_11_perm_0 = const()
           values_23_shape_0 = const()
0.07% N|C  transpose_73 = ios17.transpose(x: ["reshape_23_cast_fp16"], perm: ["transpose_11_perm_0"])
0.03% N|C  values_23_cast_fp16 = ios17.reshape(shape: ["values_23_shape_0"], x: ["transpose_73"])
           mul_5_y_0_to_fp16 = const()
0.03% N|C  mul_5_cast_fp16 = ios17.mul(y: ["mul_5_y_0_to_fp16"], x: ["query_float_11_cast_fp16"])
           matmul_5_transpose_y_0 = const()
           matmul_5_transpose_x_0 = const()
0.07% N|C  matmul_5_cast_fp16 = ios17.matmul(transpose_x: ["matmul_5_transpose_x_0"], x: ["mul_5_cast_fp16"], transpose_y: ["matmul_5_transpose_y_0"], y: ["keys_23_cast_fp16"])
0.08% N|C  add_11_cast_fp16 = ios17.add(x: ["matmul_5_cast_fp16"], y: ["causal_mask_3_to_fp16"])
           softmax_5_axis_0 = const()
0.07% N|C  softmax_5_cast_fp16 = ios16.softmax(x: ["add_11_cast_fp16"], axis: ["softmax_5_axis_0"])
           attn_output_21_transpose_x_0 = const()
           attn_output_21_transpose_y_0 = const()
0.07% N|C  attn_output_21_cast_fp16 = ios17.matmul(transpose_x: ["attn_output_21_transpose_x_0"], y: ["values_23_cast_fp16"], x: ["softmax_5_cast_fp16"], transpose_y: ["attn_output_21_transpose_y_0"])
           var_872_perm_0 = const()
           var_874 = const()
0.07% N|C  transpose_72 = ios17.transpose(x: ["attn_output_21_cast_fp16"], perm: ["var_872_perm_0"])
0.03% N|C  input_43_cast_fp16 = ios17.reshape(shape: ["var_874"], x: ["transpose_72"])
           model_transformer_layers_5_attn_out_proj_weight_to_fp16 = const()
0.24% N|C  linear_21_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_5_attn_out_proj_weight_to_fp16"], x: ["input_43_cast_fp16"], bias: ["linear_1_bias_0_to_fp16"])
0.09% N|C  x_115_cast_fp16 = ios17.add(x: ["x_99_cast_fp16"], y: ["linear_21_cast_fp16"])
           var_10_promoted_to_fp16_23 = const()
0.04% N|C  var_881_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_23"], x: ["x_115_cast_fp16"])
           var_882 = const()
0.02% N|C  var_883_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], axes: ["var_882"], x: ["var_881_cast_fp16"])
           var_884_to_fp16 = const()
0.00% N|C  var_885_cast_fp16 = ios17.add(x: ["var_883_cast_fp16"], y: ["var_884_to_fp16"])
           var_886_epsilon_0 = const()
0.00% N|C  var_886_cast_fp16 = ios17.rsqrt(epsilon: ["var_886_epsilon_0"], x: ["var_885_cast_fp16"])
0.04% N|C  var_887_cast_fp16 = ios17.mul(y: ["var_886_cast_fp16"], x: ["x_115_cast_fp16"])
           model_transformer_layers_5_ffn_norm_weight_to_fp16 = const()
0.04% N|C  input_45_cast_fp16 = ios17.mul(y: ["model_transformer_layers_5_ffn_norm_weight_to_fp16"], x: ["var_887_cast_fp16"])
           model_transformer_layers_5_ffn_proj_1_weight_to_fp16 = const()
           linear_22_bias_0_to_fp16 = const()
0.15% N|C  linear_22_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_5_ffn_proj_1_weight_to_fp16"], bias: ["linear_22_bias_0_to_fp16"], x: ["input_45_cast_fp16"])
           var_897_split_sizes_0 = const()
           var_897_axis_0 = const()
0.10% N|C  var_897_cast_fp16_0, var_897_cast_fp16_1 = split(axis: ["var_897_axis_0"], split_sizes: ["var_897_split_sizes_0"], x: ["linear_22_cast_fp16"])
0.07% N|C  var_899_cast_fp16 = ios16.silu(x: ["var_897_cast_fp16_0"])
0.12% N|C  input_49_cast_fp16 = ios17.mul(x: ["var_899_cast_fp16"], y: ["var_897_cast_fp16_1"])
           model_transformer_layers_5_ffn_proj_2_weight_to_fp16 = const()
0.96% N|C  linear_23_cast_fp16 = ios17.linear(bias: ["linear_1_bias_0_to_fp16"], weight: ["model_transformer_layers_5_ffn_proj_2_weight_to_fp16"], x: ["input_49_cast_fp16"])
0.09% N|C  x_119_cast_fp16 = ios17.add(y: ["linear_23_cast_fp16"], x: ["x_115_cast_fp16"])
           var_10_promoted_to_fp16_24 = const()
0.04% N|C  var_910_cast_fp16 = ios17.pow(x: ["x_119_cast_fp16"], y: ["var_10_promoted_to_fp16_24"])
           var_911 = const()
0.02% N|C  var_912_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_910_cast_fp16"], axes: ["var_911"])
           var_913_to_fp16 = const()
0.00% N|C  var_914_cast_fp16 = ios17.add(y: ["var_913_to_fp16"], x: ["var_912_cast_fp16"])
           var_915_epsilon_0 = const()
0.00% N|C  var_915_cast_fp16 = ios17.rsqrt(x: ["var_914_cast_fp16"], epsilon: ["var_915_epsilon_0"])
0.04% N|C  var_916_cast_fp16 = ios17.mul(x: ["x_119_cast_fp16"], y: ["var_915_cast_fp16"])
           model_transformer_layers_6_attn_norm_weight_to_fp16 = const()
0.04% N|C  hidden_states_37_cast_fp16 = ios17.mul(y: ["model_transformer_layers_6_attn_norm_weight_to_fp16"], x: ["var_916_cast_fp16"])
           model_transformer_layers_6_attn_qkv_proj_weight_to_fp16 = const()
0.38% N|C  linear_24_cast_fp16 = ios17.linear(bias: ["linear_2_bias_0_to_fp16"], weight: ["model_transformer_layers_6_attn_qkv_proj_weight_to_fp16"], x: ["hidden_states_37_cast_fp16"])
           var_930 = const()
0.05% N|C  qkv_27_cast_fp16 = ios17.reshape(x: ["linear_24_cast_fp16"], shape: ["var_930"])
           var_932_perm_0 = const()
           var_933 = const()
           var_934_axis_0 = const()
0.11% N|C  transpose_71 = ios17.transpose(x: ["qkv_27_cast_fp16"], perm: ["var_932_perm_0"])
0.04% N|C  var_934_cast_fp16_0, var_934_cast_fp16_1, var_934_cast_fp16_2 = split(x: ["transpose_71"], split_sizes: ["var_933"], axis: ["var_934_axis_0"])
           var_10_promoted_to_fp16_25 = const()
0.03% N|C  var_940_cast_fp16 = ios17.pow(x: ["var_934_cast_fp16_0"], y: ["var_10_promoted_to_fp16_25"])
           var_941 = const()
0.02% N|C  var_942_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_940_cast_fp16"], axes: ["var_941"])
           var_943_to_fp16 = const()
0.00% N|C  var_944_cast_fp16 = ios17.add(y: ["var_943_to_fp16"], x: ["var_942_cast_fp16"])
           var_945_epsilon_0 = const()
0.00% N|C  var_945_cast_fp16 = ios17.rsqrt(x: ["var_944_cast_fp16"], epsilon: ["var_945_epsilon_0"])
0.03% N|C  var_946_cast_fp16 = ios17.mul(x: ["var_934_cast_fp16_0"], y: ["var_945_cast_fp16"])
           model_transformer_layers_6_attn_q_norm_weight_to_fp16 = const()
0.03% N|C  query_13_cast_fp16 = ios17.mul(y: ["model_transformer_layers_6_attn_q_norm_weight_to_fp16"], x: ["var_946_cast_fp16"])
           var_10_promoted_to_fp16_26 = const()
0.01% N|C  var_951_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_26"], x: ["var_934_cast_fp16_1"])
           var_952 = const()
0.00% N|C  var_953_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_951_cast_fp16"], axes: ["var_952"])
           var_954_to_fp16 = const()
0.00% N|C  var_955_cast_fp16 = ios17.add(x: ["var_953_cast_fp16"], y: ["var_954_to_fp16"])
           var_956_epsilon_0 = const()
0.00% N|C  var_956_cast_fp16 = ios17.rsqrt(x: ["var_955_cast_fp16"], epsilon: ["var_956_epsilon_0"])
0.01% N|C  var_957_cast_fp16 = ios17.mul(x: ["var_934_cast_fp16_1"], y: ["var_956_cast_fp16"])
           model_transformer_layers_6_attn_k_norm_weight_to_fp16 = const()
0.01% N|C  key_13_cast_fp16 = ios17.mul(y: ["model_transformer_layers_6_attn_k_norm_weight_to_fp16"], x: ["var_957_cast_fp16"])
0.03% N|C  var_974_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["query_13_cast_fp16"])
           var_975_split_sizes_0 = const()
           var_975_axis_0 = const()
0.03% N|C  var_975_cast_fp16_0, var_975_cast_fp16_1 = split(x: ["query_13_cast_fp16"], split_sizes: ["var_975_split_sizes_0"], axis: ["var_975_axis_0"])
           const_48_promoted_to_fp16 = const()
0.02% N|C  var_977_cast_fp16 = ios17.mul(y: ["const_48_promoted_to_fp16"], x: ["var_975_cast_fp16_1"])
           var_979_interleave_0 = const()
0.01% N|C  var_979_cast_fp16 = ios17.concat(interleave: ["var_979_interleave_0"], axis: ["var_5"], values: ["var_977_cast_fp16", "var_975_cast_fp16_0"])
0.03% N|C  var_980_cast_fp16 = ios17.mul(x: ["var_979_cast_fp16"], y: ["var_152_to_fp16"])
0.08% N|C  query_float_13_cast_fp16 = ios17.add(y: ["var_980_cast_fp16"], x: ["var_974_cast_fp16"])
0.01% N|C  var_986_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["key_13_cast_fp16"])
           var_987_split_sizes_0 = const()
           var_987_axis_0 = const()
0.01% N|C  var_987_cast_fp16_0, var_987_cast_fp16_1 = split(axis: ["var_987_axis_0"], split_sizes: ["var_987_split_sizes_0"], x: ["key_13_cast_fp16"])
           const_49_promoted_to_fp16 = const()
0.00% N|C  var_989_cast_fp16 = ios17.mul(x: ["var_987_cast_fp16_1"], y: ["const_49_promoted_to_fp16"])
           var_991_interleave_0 = const()
0.00% N|C  var_991_cast_fp16 = ios17.concat(interleave: ["var_991_interleave_0"], values: ["var_989_cast_fp16", "var_987_cast_fp16_0"], axis: ["var_5"])
0.01% N|C  var_992_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_991_cast_fp16"])
0.02% N|C  key_float_13_cast_fp16 = ios17.add(x: ["var_986_cast_fp16"], y: ["var_992_cast_fp16"])
           reshape_24_shape_0 = const()
0.01% N|C  reshape_24_cast_fp16 = ios17.reshape(x: ["key_float_13_cast_fp16"], shape: ["reshape_24_shape_0"])
           tile_12_reps_0 = const()
0.02% N|C  tile_12_cast_fp16 = tile(reps: ["tile_12_reps_0"], x: ["reshape_24_cast_fp16"])
           reshape_25_shape_0 = const()
0.03% N|C  reshape_25_cast_fp16 = ios17.reshape(x: ["tile_12_cast_fp16"], shape: ["reshape_25_shape_0"])
           transpose_12_perm_0 = const()
           keys_27_shape_0 = const()
0.07% N|C  transpose_70 = ios17.transpose(x: ["reshape_25_cast_fp16"], perm: ["transpose_12_perm_0"])
0.03% N|C  keys_27_cast_fp16 = ios17.reshape(shape: ["keys_27_shape_0"], x: ["transpose_70"])
           reshape_26_shape_0 = const()
0.01% N|C  reshape_26_cast_fp16 = ios17.reshape(shape: ["reshape_26_shape_0"], x: ["var_934_cast_fp16_2"])
           tile_13_reps_0 = const()
0.02% N|C  tile_13_cast_fp16 = tile(x: ["reshape_26_cast_fp16"], reps: ["tile_13_reps_0"])
           reshape_27_shape_0 = const()
0.03% N|C  reshape_27_cast_fp16 = ios17.reshape(shape: ["reshape_27_shape_0"], x: ["tile_13_cast_fp16"])
           transpose_13_perm_0 = const()
           values_27_shape_0 = const()
0.07% N|C  transpose_69 = ios17.transpose(perm: ["transpose_13_perm_0"], x: ["reshape_27_cast_fp16"])
0.03% N|C  values_27_cast_fp16 = ios17.reshape(x: ["transpose_69"], shape: ["values_27_shape_0"])
           mul_6_y_0_to_fp16 = const()
0.03% N|C  mul_6_cast_fp16 = ios17.mul(x: ["query_float_13_cast_fp16"], y: ["mul_6_y_0_to_fp16"])
           matmul_6_transpose_y_0 = const()
           matmul_6_transpose_x_0 = const()
0.07% N|C  matmul_6_cast_fp16 = ios17.matmul(transpose_y: ["matmul_6_transpose_y_0"], transpose_x: ["matmul_6_transpose_x_0"], y: ["keys_27_cast_fp16"], x: ["mul_6_cast_fp16"])
0.08% N|C  add_13_cast_fp16 = ios17.add(x: ["matmul_6_cast_fp16"], y: ["causal_mask_3_to_fp16"])
           softmax_6_axis_0 = const()
0.07% N|C  softmax_6_cast_fp16 = ios16.softmax(axis: ["softmax_6_axis_0"], x: ["add_13_cast_fp16"])
           attn_output_25_transpose_x_0 = const()
           attn_output_25_transpose_y_0 = const()
0.07% N|C  attn_output_25_cast_fp16 = ios17.matmul(transpose_y: ["attn_output_25_transpose_y_0"], transpose_x: ["attn_output_25_transpose_x_0"], x: ["softmax_6_cast_fp16"], y: ["values_27_cast_fp16"])
           var_1008_perm_0 = const()
           var_1010 = const()
0.07% N|C  transpose_68 = ios17.transpose(perm: ["var_1008_perm_0"], x: ["attn_output_25_cast_fp16"])
0.03% N|C  input_51_cast_fp16 = ios17.reshape(x: ["transpose_68"], shape: ["var_1010"])
           model_transformer_layers_6_attn_out_proj_weight_to_fp16 = const()
0.24% N|C  linear_25_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_6_attn_out_proj_weight_to_fp16"], x: ["input_51_cast_fp16"], bias: ["linear_1_bias_0_to_fp16"])
0.09% N|C  x_135_cast_fp16 = ios17.add(x: ["x_119_cast_fp16"], y: ["linear_25_cast_fp16"])
           var_10_promoted_to_fp16_27 = const()
0.04% N|C  var_1017_cast_fp16 = ios17.pow(x: ["x_135_cast_fp16"], y: ["var_10_promoted_to_fp16_27"])
           var_1018 = const()
0.02% N|C  var_1019_cast_fp16 = ios16.reduce_mean(x: ["var_1017_cast_fp16"], keep_dims: ["var_22"], axes: ["var_1018"])
           var_1020_to_fp16 = const()
0.00% N|C  var_1021_cast_fp16 = ios17.add(y: ["var_1020_to_fp16"], x: ["var_1019_cast_fp16"])
           var_1022_epsilon_0 = const()
0.00% N|C  var_1022_cast_fp16 = ios17.rsqrt(x: ["var_1021_cast_fp16"], epsilon: ["var_1022_epsilon_0"])
0.04% N|C  var_1023_cast_fp16 = ios17.mul(y: ["var_1022_cast_fp16"], x: ["x_135_cast_fp16"])
           model_transformer_layers_6_ffn_norm_weight_to_fp16 = const()
0.04% N|C  input_53_cast_fp16 = ios17.mul(x: ["var_1023_cast_fp16"], y: ["model_transformer_layers_6_ffn_norm_weight_to_fp16"])
           model_transformer_layers_6_ffn_proj_1_weight_to_fp16 = const()
           linear_26_bias_0_to_fp16 = const()
0.15% N|C  linear_26_cast_fp16 = ios17.linear(x: ["input_53_cast_fp16"], weight: ["model_transformer_layers_6_ffn_proj_1_weight_to_fp16"], bias: ["linear_26_bias_0_to_fp16"])
           var_1033_split_sizes_0 = const()
           var_1033_axis_0 = const()
0.13% N|C  var_1033_cast_fp16_0, var_1033_cast_fp16_1 = split(axis: ["var_1033_axis_0"], x: ["linear_26_cast_fp16"], split_sizes: ["var_1033_split_sizes_0"])
0.09% N|C  var_1035_cast_fp16 = ios16.silu(x: ["var_1033_cast_fp16_0"])
0.15% N|C  input_57_cast_fp16 = ios17.mul(x: ["var_1035_cast_fp16"], y: ["var_1033_cast_fp16_1"])
           model_transformer_layers_6_ffn_proj_2_weight_to_fp16 = const()
1.50% N|C  linear_27_cast_fp16 = ios17.linear(bias: ["linear_1_bias_0_to_fp16"], x: ["input_57_cast_fp16"], weight: ["model_transformer_layers_6_ffn_proj_2_weight_to_fp16"])
0.09% N|C  x_139_cast_fp16 = ios17.add(y: ["linear_27_cast_fp16"], x: ["x_135_cast_fp16"])
           var_10_promoted_to_fp16_28 = const()
0.04% N|C  var_1046_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_28"], x: ["x_139_cast_fp16"])
           var_1047 = const()
0.02% N|C  var_1048_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_1046_cast_fp16"], axes: ["var_1047"])
           var_1049_to_fp16 = const()
0.00% N|C  var_1050_cast_fp16 = ios17.add(x: ["var_1048_cast_fp16"], y: ["var_1049_to_fp16"])
           var_1051_epsilon_0 = const()
0.00% N|C  var_1051_cast_fp16 = ios17.rsqrt(x: ["var_1050_cast_fp16"], epsilon: ["var_1051_epsilon_0"])
0.04% N|C  var_1052_cast_fp16 = ios17.mul(x: ["x_139_cast_fp16"], y: ["var_1051_cast_fp16"])
           model_transformer_layers_7_attn_norm_weight_to_fp16 = const()
0.04% N|C  hidden_states_43_cast_fp16 = ios17.mul(y: ["model_transformer_layers_7_attn_norm_weight_to_fp16"], x: ["var_1052_cast_fp16"])
           model_transformer_layers_7_attn_qkv_proj_weight_to_fp16 = const()
0.38% N|C  linear_28_cast_fp16 = ios17.linear(bias: ["linear_2_bias_0_to_fp16"], weight: ["model_transformer_layers_7_attn_qkv_proj_weight_to_fp16"], x: ["hidden_states_43_cast_fp16"])
           var_1066 = const()
0.05% N|C  qkv_31_cast_fp16 = ios17.reshape(x: ["linear_28_cast_fp16"], shape: ["var_1066"])
           var_1068_perm_0 = const()
           var_1069 = const()
           var_1070_axis_0 = const()
0.11% N|C  transpose_67 = ios17.transpose(perm: ["var_1068_perm_0"], x: ["qkv_31_cast_fp16"])
0.04% N|C  var_1070_cast_fp16_0, var_1070_cast_fp16_1, var_1070_cast_fp16_2 = split(axis: ["var_1070_axis_0"], x: ["transpose_67"], split_sizes: ["var_1069"])
           var_10_promoted_to_fp16_29 = const()
0.03% N|C  var_1076_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_29"], x: ["var_1070_cast_fp16_0"])
           var_1077 = const()
0.02% N|C  var_1078_cast_fp16 = ios16.reduce_mean(x: ["var_1076_cast_fp16"], keep_dims: ["var_22"], axes: ["var_1077"])
           var_1079_to_fp16 = const()
0.00% N|C  var_1080_cast_fp16 = ios17.add(y: ["var_1079_to_fp16"], x: ["var_1078_cast_fp16"])
           var_1081_epsilon_0 = const()
0.00% N|C  var_1081_cast_fp16 = ios17.rsqrt(epsilon: ["var_1081_epsilon_0"], x: ["var_1080_cast_fp16"])
0.03% N|C  var_1082_cast_fp16 = ios17.mul(y: ["var_1081_cast_fp16"], x: ["var_1070_cast_fp16_0"])
           model_transformer_layers_7_attn_q_norm_weight_to_fp16 = const()
0.03% N|C  query_15_cast_fp16 = ios17.mul(y: ["model_transformer_layers_7_attn_q_norm_weight_to_fp16"], x: ["var_1082_cast_fp16"])
           var_10_promoted_to_fp16_30 = const()
0.01% N|C  var_1087_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_30"], x: ["var_1070_cast_fp16_1"])
           var_1088 = const()
0.00% N|C  var_1089_cast_fp16 = ios16.reduce_mean(x: ["var_1087_cast_fp16"], keep_dims: ["var_22"], axes: ["var_1088"])
           var_1090_to_fp16 = const()
0.00% N|C  var_1091_cast_fp16 = ios17.add(y: ["var_1090_to_fp16"], x: ["var_1089_cast_fp16"])
           var_1092_epsilon_0 = const()
0.00% N|C  var_1092_cast_fp16 = ios17.rsqrt(x: ["var_1091_cast_fp16"], epsilon: ["var_1092_epsilon_0"])
0.01% N|C  var_1093_cast_fp16 = ios17.mul(x: ["var_1070_cast_fp16_1"], y: ["var_1092_cast_fp16"])
           model_transformer_layers_7_attn_k_norm_weight_to_fp16 = const()
0.01% N|C  key_15_cast_fp16 = ios17.mul(y: ["model_transformer_layers_7_attn_k_norm_weight_to_fp16"], x: ["var_1093_cast_fp16"])
0.03% N|C  var_1110_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["query_15_cast_fp16"])
           var_1111_split_sizes_0 = const()
           var_1111_axis_0 = const()
0.03% N|C  var_1111_cast_fp16_0, var_1111_cast_fp16_1 = split(x: ["query_15_cast_fp16"], axis: ["var_1111_axis_0"], split_sizes: ["var_1111_split_sizes_0"])
           const_55_promoted_to_fp16 = const()
0.02% N|C  var_1113_cast_fp16 = ios17.mul(y: ["const_55_promoted_to_fp16"], x: ["var_1111_cast_fp16_1"])
           var_1115_interleave_0 = const()
0.01% N|C  var_1115_cast_fp16 = ios17.concat(axis: ["var_5"], values: ["var_1113_cast_fp16", "var_1111_cast_fp16_0"], interleave: ["var_1115_interleave_0"])
0.03% N|C  var_1116_cast_fp16 = ios17.mul(x: ["var_1115_cast_fp16"], y: ["var_152_to_fp16"])
0.08% N|C  query_float_15_cast_fp16 = ios17.add(x: ["var_1110_cast_fp16"], y: ["var_1116_cast_fp16"])
0.01% N|C  var_1122_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["key_15_cast_fp16"])
           var_1123_split_sizes_0 = const()
           var_1123_axis_0 = const()
0.01% N|C  var_1123_cast_fp16_0, var_1123_cast_fp16_1 = split(x: ["key_15_cast_fp16"], axis: ["var_1123_axis_0"], split_sizes: ["var_1123_split_sizes_0"])
           const_56_promoted_to_fp16 = const()
0.00% N|C  var_1125_cast_fp16 = ios17.mul(y: ["const_56_promoted_to_fp16"], x: ["var_1123_cast_fp16_1"])
           var_1127_interleave_0 = const()
0.00% N|C  var_1127_cast_fp16 = ios17.concat(values: ["var_1125_cast_fp16", "var_1123_cast_fp16_0"], axis: ["var_5"], interleave: ["var_1127_interleave_0"])
0.01% N|C  var_1128_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_1127_cast_fp16"])
0.02% N|C  key_float_15_cast_fp16 = ios17.add(x: ["var_1122_cast_fp16"], y: ["var_1128_cast_fp16"])
           reshape_28_shape_0 = const()
0.01% N|C  reshape_28_cast_fp16 = ios17.reshape(x: ["key_float_15_cast_fp16"], shape: ["reshape_28_shape_0"])
           tile_14_reps_0 = const()
0.02% N|C  tile_14_cast_fp16 = tile(x: ["reshape_28_cast_fp16"], reps: ["tile_14_reps_0"])
           reshape_29_shape_0 = const()
0.03% N|C  reshape_29_cast_fp16 = ios17.reshape(shape: ["reshape_29_shape_0"], x: ["tile_14_cast_fp16"])
           transpose_14_perm_0 = const()
           keys_31_shape_0 = const()
0.07% N|C  transpose_66 = ios17.transpose(perm: ["transpose_14_perm_0"], x: ["reshape_29_cast_fp16"])
0.03% N|C  keys_31_cast_fp16 = ios17.reshape(shape: ["keys_31_shape_0"], x: ["transpose_66"])
           reshape_30_shape_0 = const()
0.01% N|C  reshape_30_cast_fp16 = ios17.reshape(shape: ["reshape_30_shape_0"], x: ["var_1070_cast_fp16_2"])
           tile_15_reps_0 = const()
0.02% N|C  tile_15_cast_fp16 = tile(x: ["reshape_30_cast_fp16"], reps: ["tile_15_reps_0"])
           reshape_31_shape_0 = const()
0.03% N|C  reshape_31_cast_fp16 = ios17.reshape(shape: ["reshape_31_shape_0"], x: ["tile_15_cast_fp16"])
           transpose_15_perm_0 = const()
           values_31_shape_0 = const()
0.07% N|C  transpose_65 = ios17.transpose(x: ["reshape_31_cast_fp16"], perm: ["transpose_15_perm_0"])
0.03% N|C  values_31_cast_fp16 = ios17.reshape(shape: ["values_31_shape_0"], x: ["transpose_65"])
           mul_7_y_0_to_fp16 = const()
0.03% N|C  mul_7_cast_fp16 = ios17.mul(x: ["query_float_15_cast_fp16"], y: ["mul_7_y_0_to_fp16"])
           matmul_7_transpose_y_0 = const()
           matmul_7_transpose_x_0 = const()
0.07% N|C  matmul_7_cast_fp16 = ios17.matmul(x: ["mul_7_cast_fp16"], transpose_x: ["matmul_7_transpose_x_0"], transpose_y: ["matmul_7_transpose_y_0"], y: ["keys_31_cast_fp16"])
0.08% N|C  add_15_cast_fp16 = ios17.add(x: ["matmul_7_cast_fp16"], y: ["causal_mask_3_to_fp16"])
           softmax_7_axis_0 = const()
0.07% N|C  softmax_7_cast_fp16 = ios16.softmax(x: ["add_15_cast_fp16"], axis: ["softmax_7_axis_0"])
           attn_output_29_transpose_x_0 = const()
           attn_output_29_transpose_y_0 = const()
0.07% N|C  attn_output_29_cast_fp16 = ios17.matmul(x: ["softmax_7_cast_fp16"], y: ["values_31_cast_fp16"], transpose_y: ["attn_output_29_transpose_y_0"], transpose_x: ["attn_output_29_transpose_x_0"])
           var_1144_perm_0 = const()
           var_1146 = const()
0.07% N|C  transpose_64 = ios17.transpose(perm: ["var_1144_perm_0"], x: ["attn_output_29_cast_fp16"])
0.03% N|C  input_59_cast_fp16 = ios17.reshape(x: ["transpose_64"], shape: ["var_1146"])
           model_transformer_layers_7_attn_out_proj_weight_to_fp16 = const()
0.24% N|C  linear_29_cast_fp16 = ios17.linear(x: ["input_59_cast_fp16"], bias: ["linear_1_bias_0_to_fp16"], weight: ["model_transformer_layers_7_attn_out_proj_weight_to_fp16"])
0.09% N|C  x_155_cast_fp16 = ios17.add(x: ["x_139_cast_fp16"], y: ["linear_29_cast_fp16"])
           var_10_promoted_to_fp16_31 = const()
0.04% N|C  var_1153_cast_fp16 = ios17.pow(x: ["x_155_cast_fp16"], y: ["var_10_promoted_to_fp16_31"])
           var_1154 = const()
0.02% N|C  var_1155_cast_fp16 = ios16.reduce_mean(x: ["var_1153_cast_fp16"], axes: ["var_1154"], keep_dims: ["var_22"])
           var_1156_to_fp16 = const()
0.00% N|C  var_1157_cast_fp16 = ios17.add(y: ["var_1156_to_fp16"], x: ["var_1155_cast_fp16"])
           var_1158_epsilon_0 = const()
0.00% N|C  var_1158_cast_fp16 = ios17.rsqrt(x: ["var_1157_cast_fp16"], epsilon: ["var_1158_epsilon_0"])
0.04% N|C  var_1159_cast_fp16 = ios17.mul(y: ["var_1158_cast_fp16"], x: ["x_155_cast_fp16"])
           model_transformer_layers_7_ffn_norm_weight_to_fp16 = const()
0.04% N|C  input_61_cast_fp16 = ios17.mul(x: ["var_1159_cast_fp16"], y: ["model_transformer_layers_7_ffn_norm_weight_to_fp16"])
           model_transformer_layers_7_ffn_proj_1_weight_to_fp16 = const()
           linear_30_bias_0_to_fp16 = const()
0.15% N|C  linear_30_cast_fp16 = ios17.linear(bias: ["linear_30_bias_0_to_fp16"], weight: ["model_transformer_layers_7_ffn_proj_1_weight_to_fp16"], x: ["input_61_cast_fp16"])
           var_1169_split_sizes_0 = const()
           var_1169_axis_0 = const()
0.14% N|C  var_1169_cast_fp16_0, var_1169_cast_fp16_1 = split(axis: ["var_1169_axis_0"], split_sizes: ["var_1169_split_sizes_0"], x: ["linear_30_cast_fp16"])
0.09% N|C  var_1171_cast_fp16 = ios16.silu(x: ["var_1169_cast_fp16_0"])
0.17% N|C  input_65_cast_fp16 = ios17.mul(y: ["var_1169_cast_fp16_1"], x: ["var_1171_cast_fp16"])
           model_transformer_layers_7_ffn_proj_2_weight_to_fp16 = const()
1.82% N|C  linear_31_cast_fp16 = ios17.linear(bias: ["linear_1_bias_0_to_fp16"], weight: ["model_transformer_layers_7_ffn_proj_2_weight_to_fp16"], x: ["input_65_cast_fp16"])
0.09% N|C  x_159_cast_fp16 = ios17.add(y: ["linear_31_cast_fp16"], x: ["x_155_cast_fp16"])
           var_10_promoted_to_fp16_32 = const()
0.04% N|C  var_1182_cast_fp16 = ios17.pow(x: ["x_159_cast_fp16"], y: ["var_10_promoted_to_fp16_32"])
           var_1183 = const()
0.02% N|C  var_1184_cast_fp16 = ios16.reduce_mean(axes: ["var_1183"], keep_dims: ["var_22"], x: ["var_1182_cast_fp16"])
           var_1185_to_fp16 = const()
0.00% N|C  var_1186_cast_fp16 = ios17.add(y: ["var_1185_to_fp16"], x: ["var_1184_cast_fp16"])
           var_1187_epsilon_0 = const()
0.00% N|C  var_1187_cast_fp16 = ios17.rsqrt(x: ["var_1186_cast_fp16"], epsilon: ["var_1187_epsilon_0"])
0.04% N|C  var_1188_cast_fp16 = ios17.mul(y: ["var_1187_cast_fp16"], x: ["x_159_cast_fp16"])
           model_transformer_layers_8_attn_norm_weight_to_fp16 = const()
0.04% N|C  hidden_states_49_cast_fp16 = ios17.mul(y: ["model_transformer_layers_8_attn_norm_weight_to_fp16"], x: ["var_1188_cast_fp16"])
           model_transformer_layers_8_attn_qkv_proj_weight_to_fp16 = const()
0.38% N|C  linear_32_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_8_attn_qkv_proj_weight_to_fp16"], bias: ["linear_2_bias_0_to_fp16"], x: ["hidden_states_49_cast_fp16"])
           var_1202 = const()
0.05% N|C  qkv_35_cast_fp16 = ios17.reshape(x: ["linear_32_cast_fp16"], shape: ["var_1202"])
           var_1204_perm_0 = const()
           var_1205 = const()
           var_1206_axis_0 = const()
0.11% N|C  transpose_63 = ios17.transpose(x: ["qkv_35_cast_fp16"], perm: ["var_1204_perm_0"])
0.04% N|C  var_1206_cast_fp16_0, var_1206_cast_fp16_1, var_1206_cast_fp16_2 = split(axis: ["var_1206_axis_0"], x: ["transpose_63"], split_sizes: ["var_1205"])
           var_10_promoted_to_fp16_33 = const()
0.03% N|C  var_1212_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_33"], x: ["var_1206_cast_fp16_0"])
           var_1213 = const()
0.02% N|C  var_1214_cast_fp16 = ios16.reduce_mean(axes: ["var_1213"], x: ["var_1212_cast_fp16"], keep_dims: ["var_22"])
           var_1215_to_fp16 = const()
0.00% N|C  var_1216_cast_fp16 = ios17.add(y: ["var_1215_to_fp16"], x: ["var_1214_cast_fp16"])
           var_1217_epsilon_0 = const()
0.00% N|C  var_1217_cast_fp16 = ios17.rsqrt(x: ["var_1216_cast_fp16"], epsilon: ["var_1217_epsilon_0"])
0.03% N|C  var_1218_cast_fp16 = ios17.mul(x: ["var_1206_cast_fp16_0"], y: ["var_1217_cast_fp16"])
           model_transformer_layers_8_attn_q_norm_weight_to_fp16 = const()
0.03% N|C  query_17_cast_fp16 = ios17.mul(y: ["model_transformer_layers_8_attn_q_norm_weight_to_fp16"], x: ["var_1218_cast_fp16"])
           var_10_promoted_to_fp16_34 = const()
0.01% N|C  var_1223_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_34"], x: ["var_1206_cast_fp16_1"])
           var_1224 = const()
0.00% N|C  var_1225_cast_fp16 = ios16.reduce_mean(x: ["var_1223_cast_fp16"], axes: ["var_1224"], keep_dims: ["var_22"])
           var_1226_to_fp16 = const()
0.00% N|C  var_1227_cast_fp16 = ios17.add(x: ["var_1225_cast_fp16"], y: ["var_1226_to_fp16"])
           var_1228_epsilon_0 = const()
0.00% N|C  var_1228_cast_fp16 = ios17.rsqrt(x: ["var_1227_cast_fp16"], epsilon: ["var_1228_epsilon_0"])
0.01% N|C  var_1229_cast_fp16 = ios17.mul(y: ["var_1228_cast_fp16"], x: ["var_1206_cast_fp16_1"])
           model_transformer_layers_8_attn_k_norm_weight_to_fp16 = const()
0.01% N|C  key_17_cast_fp16 = ios17.mul(y: ["model_transformer_layers_8_attn_k_norm_weight_to_fp16"], x: ["var_1229_cast_fp16"])
0.03% N|C  var_1246_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["query_17_cast_fp16"])
           var_1247_split_sizes_0 = const()
           var_1247_axis_0 = const()
0.03% N|C  var_1247_cast_fp16_0, var_1247_cast_fp16_1 = split(x: ["query_17_cast_fp16"], split_sizes: ["var_1247_split_sizes_0"], axis: ["var_1247_axis_0"])
           const_62_promoted_to_fp16 = const()
0.02% N|C  var_1249_cast_fp16 = ios17.mul(y: ["const_62_promoted_to_fp16"], x: ["var_1247_cast_fp16_1"])
           var_1251_interleave_0 = const()
0.01% N|C  var_1251_cast_fp16 = ios17.concat(values: ["var_1249_cast_fp16", "var_1247_cast_fp16_0"], interleave: ["var_1251_interleave_0"], axis: ["var_5"])
0.03% N|C  var_1252_cast_fp16 = ios17.mul(x: ["var_1251_cast_fp16"], y: ["var_152_to_fp16"])
0.08% N|C  query_float_17_cast_fp16 = ios17.add(y: ["var_1252_cast_fp16"], x: ["var_1246_cast_fp16"])
0.01% N|C  var_1258_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["key_17_cast_fp16"])
           var_1259_split_sizes_0 = const()
           var_1259_axis_0 = const()
0.01% N|C  var_1259_cast_fp16_0, var_1259_cast_fp16_1 = split(split_sizes: ["var_1259_split_sizes_0"], axis: ["var_1259_axis_0"], x: ["key_17_cast_fp16"])
           const_63_promoted_to_fp16 = const()
0.00% N|C  var_1261_cast_fp16 = ios17.mul(y: ["const_63_promoted_to_fp16"], x: ["var_1259_cast_fp16_1"])
           var_1263_interleave_0 = const()
0.00% N|C  var_1263_cast_fp16 = ios17.concat(interleave: ["var_1263_interleave_0"], values: ["var_1261_cast_fp16", "var_1259_cast_fp16_0"], axis: ["var_5"])
0.01% N|C  var_1264_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_1263_cast_fp16"])
0.02% N|C  key_float_17_cast_fp16 = ios17.add(x: ["var_1258_cast_fp16"], y: ["var_1264_cast_fp16"])
           reshape_32_shape_0 = const()
0.01% N|C  reshape_32_cast_fp16 = ios17.reshape(x: ["key_float_17_cast_fp16"], shape: ["reshape_32_shape_0"])
           tile_16_reps_0 = const()
0.02% N|C  tile_16_cast_fp16 = tile(x: ["reshape_32_cast_fp16"], reps: ["tile_16_reps_0"])
           reshape_33_shape_0 = const()
0.03% N|C  reshape_33_cast_fp16 = ios17.reshape(x: ["tile_16_cast_fp16"], shape: ["reshape_33_shape_0"])
           transpose_16_perm_0 = const()
           keys_35_shape_0 = const()
0.07% N|C  transpose_62 = ios17.transpose(x: ["reshape_33_cast_fp16"], perm: ["transpose_16_perm_0"])
0.03% N|C  keys_35_cast_fp16 = ios17.reshape(x: ["transpose_62"], shape: ["keys_35_shape_0"])
           reshape_34_shape_0 = const()
0.01% N|C  reshape_34_cast_fp16 = ios17.reshape(x: ["var_1206_cast_fp16_2"], shape: ["reshape_34_shape_0"])
           tile_17_reps_0 = const()
0.02% N|C  tile_17_cast_fp16 = tile(reps: ["tile_17_reps_0"], x: ["reshape_34_cast_fp16"])
           reshape_35_shape_0 = const()
0.03% N|C  reshape_35_cast_fp16 = ios17.reshape(x: ["tile_17_cast_fp16"], shape: ["reshape_35_shape_0"])
           transpose_17_perm_0 = const()
           values_35_shape_0 = const()
0.07% N|C  transpose_61 = ios17.transpose(perm: ["transpose_17_perm_0"], x: ["reshape_35_cast_fp16"])
0.03% N|C  values_35_cast_fp16 = ios17.reshape(shape: ["values_35_shape_0"], x: ["transpose_61"])
           mul_8_y_0_to_fp16 = const()
0.03% N|C  mul_8_cast_fp16 = ios17.mul(y: ["mul_8_y_0_to_fp16"], x: ["query_float_17_cast_fp16"])
           matmul_8_transpose_y_0 = const()
           matmul_8_transpose_x_0 = const()
0.07% N|C  matmul_8_cast_fp16 = ios17.matmul(y: ["keys_35_cast_fp16"], x: ["mul_8_cast_fp16"], transpose_x: ["matmul_8_transpose_x_0"], transpose_y: ["matmul_8_transpose_y_0"])
0.08% N|C  add_17_cast_fp16 = ios17.add(x: ["matmul_8_cast_fp16"], y: ["causal_mask_3_to_fp16"])
           softmax_8_axis_0 = const()
0.07% N|C  softmax_8_cast_fp16 = ios16.softmax(x: ["add_17_cast_fp16"], axis: ["softmax_8_axis_0"])
           attn_output_33_transpose_x_0 = const()
           attn_output_33_transpose_y_0 = const()
0.07% N|C  attn_output_33_cast_fp16 = ios17.matmul(transpose_y: ["attn_output_33_transpose_y_0"], y: ["values_35_cast_fp16"], transpose_x: ["attn_output_33_transpose_x_0"], x: ["softmax_8_cast_fp16"])
           var_1280_perm_0 = const()
           var_1282 = const()
0.07% N|C  transpose_60 = ios17.transpose(x: ["attn_output_33_cast_fp16"], perm: ["var_1280_perm_0"])
0.03% N|C  input_67_cast_fp16 = ios17.reshape(x: ["transpose_60"], shape: ["var_1282"])
           model_transformer_layers_8_attn_out_proj_weight_to_fp16 = const()
0.24% N|C  linear_33_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_8_attn_out_proj_weight_to_fp16"], x: ["input_67_cast_fp16"], bias: ["linear_1_bias_0_to_fp16"])
0.09% N|C  x_175_cast_fp16 = ios17.add(y: ["linear_33_cast_fp16"], x: ["x_159_cast_fp16"])
           var_10_promoted_to_fp16_35 = const()
0.04% N|C  var_1289_cast_fp16 = ios17.pow(x: ["x_175_cast_fp16"], y: ["var_10_promoted_to_fp16_35"])
           var_1290 = const()
0.02% N|C  var_1291_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], axes: ["var_1290"], x: ["var_1289_cast_fp16"])
           var_1292_to_fp16 = const()
0.00% N|C  var_1293_cast_fp16 = ios17.add(y: ["var_1292_to_fp16"], x: ["var_1291_cast_fp16"])
           var_1294_epsilon_0 = const()
0.00% N|C  var_1294_cast_fp16 = ios17.rsqrt(x: ["var_1293_cast_fp16"], epsilon: ["var_1294_epsilon_0"])
0.04% N|C  var_1295_cast_fp16 = ios17.mul(y: ["var_1294_cast_fp16"], x: ["x_175_cast_fp16"])
           model_transformer_layers_8_ffn_norm_weight_to_fp16 = const()
0.04% N|C  input_69_cast_fp16 = ios17.mul(y: ["model_transformer_layers_8_ffn_norm_weight_to_fp16"], x: ["var_1295_cast_fp16"])
           model_transformer_layers_8_ffn_proj_1_weight_to_fp16 = const()
           linear_34_bias_0_to_fp16 = const()
0.15% N|C  linear_34_cast_fp16 = ios17.linear(bias: ["linear_34_bias_0_to_fp16"], weight: ["model_transformer_layers_8_ffn_proj_1_weight_to_fp16"], x: ["input_69_cast_fp16"])
           var_1305_split_sizes_0 = const()
           var_1305_axis_0 = const()
0.15% N|C  var_1305_cast_fp16_0, var_1305_cast_fp16_1 = split(split_sizes: ["var_1305_split_sizes_0"], axis: ["var_1305_axis_0"], x: ["linear_34_cast_fp16"])
0.10% N|C  var_1307_cast_fp16 = ios16.silu(x: ["var_1305_cast_fp16_0"])
0.18% N|C  input_73_cast_fp16 = ios17.mul(x: ["var_1307_cast_fp16"], y: ["var_1305_cast_fp16_1"])
           model_transformer_layers_8_ffn_proj_2_weight_to_fp16 = const()
2.16% N|C  linear_35_cast_fp16 = ios17.linear(bias: ["linear_1_bias_0_to_fp16"], x: ["input_73_cast_fp16"], weight: ["model_transformer_layers_8_ffn_proj_2_weight_to_fp16"])
0.09% N|C  x_179_cast_fp16 = ios17.add(y: ["linear_35_cast_fp16"], x: ["x_175_cast_fp16"])
           var_10_promoted_to_fp16_36 = const()
0.04% N|C  var_1318_cast_fp16 = ios17.pow(x: ["x_179_cast_fp16"], y: ["var_10_promoted_to_fp16_36"])
           var_1319 = const()
0.02% N|C  var_1320_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_1318_cast_fp16"], axes: ["var_1319"])
           var_1321_to_fp16 = const()
0.00% N|C  var_1322_cast_fp16 = ios17.add(x: ["var_1320_cast_fp16"], y: ["var_1321_to_fp16"])
           var_1323_epsilon_0 = const()
0.00% N|C  var_1323_cast_fp16 = ios17.rsqrt(x: ["var_1322_cast_fp16"], epsilon: ["var_1323_epsilon_0"])
0.04% N|C  var_1324_cast_fp16 = ios17.mul(x: ["x_179_cast_fp16"], y: ["var_1323_cast_fp16"])
           model_transformer_layers_9_attn_norm_weight_to_fp16 = const()
0.04% N|C  hidden_states_55_cast_fp16 = ios17.mul(y: ["model_transformer_layers_9_attn_norm_weight_to_fp16"], x: ["var_1324_cast_fp16"])
           model_transformer_layers_9_attn_qkv_proj_weight_to_fp16 = const()
0.38% N|C  linear_36_cast_fp16 = ios17.linear(bias: ["linear_2_bias_0_to_fp16"], x: ["hidden_states_55_cast_fp16"], weight: ["model_transformer_layers_9_attn_qkv_proj_weight_to_fp16"])
           var_1338 = const()
0.05% N|C  qkv_39_cast_fp16 = ios17.reshape(x: ["linear_36_cast_fp16"], shape: ["var_1338"])
           var_1340_perm_0 = const()
           var_1341 = const()
           var_1342_axis_0 = const()
0.11% N|C  transpose_59 = ios17.transpose(perm: ["var_1340_perm_0"], x: ["qkv_39_cast_fp16"])
0.04% N|C  var_1342_cast_fp16_0, var_1342_cast_fp16_1, var_1342_cast_fp16_2 = split(axis: ["var_1342_axis_0"], split_sizes: ["var_1341"], x: ["transpose_59"])
           var_10_promoted_to_fp16_37 = const()
0.03% N|C  var_1348_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_37"], x: ["var_1342_cast_fp16_0"])
           var_1349 = const()
0.02% N|C  var_1350_cast_fp16 = ios16.reduce_mean(x: ["var_1348_cast_fp16"], keep_dims: ["var_22"], axes: ["var_1349"])
           var_1351_to_fp16 = const()
0.00% N|C  var_1352_cast_fp16 = ios17.add(y: ["var_1351_to_fp16"], x: ["var_1350_cast_fp16"])
           var_1353_epsilon_0 = const()
0.00% N|C  var_1353_cast_fp16 = ios17.rsqrt(x: ["var_1352_cast_fp16"], epsilon: ["var_1353_epsilon_0"])
0.03% N|C  var_1354_cast_fp16 = ios17.mul(x: ["var_1342_cast_fp16_0"], y: ["var_1353_cast_fp16"])
           model_transformer_layers_9_attn_q_norm_weight_to_fp16 = const()
0.03% N|C  query_19_cast_fp16 = ios17.mul(x: ["var_1354_cast_fp16"], y: ["model_transformer_layers_9_attn_q_norm_weight_to_fp16"])
           var_10_promoted_to_fp16_38 = const()
0.01% N|C  var_1359_cast_fp16 = ios17.pow(x: ["var_1342_cast_fp16_1"], y: ["var_10_promoted_to_fp16_38"])
           var_1360 = const()
0.00% N|C  var_1361_cast_fp16 = ios16.reduce_mean(x: ["var_1359_cast_fp16"], keep_dims: ["var_22"], axes: ["var_1360"])
           var_1362_to_fp16 = const()
0.00% N|C  var_1363_cast_fp16 = ios17.add(x: ["var_1361_cast_fp16"], y: ["var_1362_to_fp16"])
           var_1364_epsilon_0 = const()
0.00% N|C  var_1364_cast_fp16 = ios17.rsqrt(epsilon: ["var_1364_epsilon_0"], x: ["var_1363_cast_fp16"])
0.01% N|C  var_1365_cast_fp16 = ios17.mul(x: ["var_1342_cast_fp16_1"], y: ["var_1364_cast_fp16"])
           model_transformer_layers_9_attn_k_norm_weight_to_fp16 = const()
0.01% N|C  key_19_cast_fp16 = ios17.mul(x: ["var_1365_cast_fp16"], y: ["model_transformer_layers_9_attn_k_norm_weight_to_fp16"])
0.03% N|C  var_1382_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["query_19_cast_fp16"])
           var_1383_split_sizes_0 = const()
           var_1383_axis_0 = const()
0.03% N|C  var_1383_cast_fp16_0, var_1383_cast_fp16_1 = split(split_sizes: ["var_1383_split_sizes_0"], x: ["query_19_cast_fp16"], axis: ["var_1383_axis_0"])
           const_69_promoted_to_fp16 = const()
0.02% N|C  var_1385_cast_fp16 = ios17.mul(y: ["const_69_promoted_to_fp16"], x: ["var_1383_cast_fp16_1"])
           var_1387_interleave_0 = const()
0.01% N|C  var_1387_cast_fp16 = ios17.concat(interleave: ["var_1387_interleave_0"], axis: ["var_5"], values: ["var_1385_cast_fp16", "var_1383_cast_fp16_0"])
0.03% N|C  var_1388_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_1387_cast_fp16"])
0.08% N|C  query_float_19_cast_fp16 = ios17.add(y: ["var_1388_cast_fp16"], x: ["var_1382_cast_fp16"])
0.01% N|C  var_1394_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["key_19_cast_fp16"])
           var_1395_split_sizes_0 = const()
           var_1395_axis_0 = const()
0.01% N|C  var_1395_cast_fp16_0, var_1395_cast_fp16_1 = split(x: ["key_19_cast_fp16"], split_sizes: ["var_1395_split_sizes_0"], axis: ["var_1395_axis_0"])
           const_70_promoted_to_fp16 = const()
0.00% N|C  var_1397_cast_fp16 = ios17.mul(y: ["const_70_promoted_to_fp16"], x: ["var_1395_cast_fp16_1"])
           var_1399_interleave_0 = const()
0.00% N|C  var_1399_cast_fp16 = ios17.concat(values: ["var_1397_cast_fp16", "var_1395_cast_fp16_0"], axis: ["var_5"], interleave: ["var_1399_interleave_0"])
0.01% N|C  var_1400_cast_fp16 = ios17.mul(x: ["var_1399_cast_fp16"], y: ["var_152_to_fp16"])
0.02% N|C  key_float_19_cast_fp16 = ios17.add(y: ["var_1400_cast_fp16"], x: ["var_1394_cast_fp16"])
           reshape_36_shape_0 = const()
0.01% N|C  reshape_36_cast_fp16 = ios17.reshape(x: ["key_float_19_cast_fp16"], shape: ["reshape_36_shape_0"])
           tile_18_reps_0 = const()
0.02% N|C  tile_18_cast_fp16 = tile(x: ["reshape_36_cast_fp16"], reps: ["tile_18_reps_0"])
           reshape_37_shape_0 = const()
0.03% N|C  reshape_37_cast_fp16 = ios17.reshape(x: ["tile_18_cast_fp16"], shape: ["reshape_37_shape_0"])
           transpose_18_perm_0 = const()
           keys_39_shape_0 = const()
0.07% N|C  transpose_58 = ios17.transpose(x: ["reshape_37_cast_fp16"], perm: ["transpose_18_perm_0"])
0.03% N|C  keys_39_cast_fp16 = ios17.reshape(shape: ["keys_39_shape_0"], x: ["transpose_58"])
           reshape_38_shape_0 = const()
0.01% N|C  reshape_38_cast_fp16 = ios17.reshape(shape: ["reshape_38_shape_0"], x: ["var_1342_cast_fp16_2"])
           tile_19_reps_0 = const()
0.02% N|C  tile_19_cast_fp16 = tile(x: ["reshape_38_cast_fp16"], reps: ["tile_19_reps_0"])
           reshape_39_shape_0 = const()
0.03% N|C  reshape_39_cast_fp16 = ios17.reshape(x: ["tile_19_cast_fp16"], shape: ["reshape_39_shape_0"])
           transpose_19_perm_0 = const()
           values_39_shape_0 = const()
0.07% N|C  transpose_57 = ios17.transpose(x: ["reshape_39_cast_fp16"], perm: ["transpose_19_perm_0"])
0.03% N|C  values_39_cast_fp16 = ios17.reshape(x: ["transpose_57"], shape: ["values_39_shape_0"])
           mul_9_y_0_to_fp16 = const()
0.03% N|C  mul_9_cast_fp16 = ios17.mul(x: ["query_float_19_cast_fp16"], y: ["mul_9_y_0_to_fp16"])
           matmul_9_transpose_y_0 = const()
           matmul_9_transpose_x_0 = const()
0.07% N|C  matmul_9_cast_fp16 = ios17.matmul(x: ["mul_9_cast_fp16"], transpose_x: ["matmul_9_transpose_x_0"], y: ["keys_39_cast_fp16"], transpose_y: ["matmul_9_transpose_y_0"])
0.08% N|C  add_19_cast_fp16 = ios17.add(y: ["causal_mask_3_to_fp16"], x: ["matmul_9_cast_fp16"])
           softmax_9_axis_0 = const()
0.07% N|C  softmax_9_cast_fp16 = ios16.softmax(x: ["add_19_cast_fp16"], axis: ["softmax_9_axis_0"])
           attn_output_37_transpose_x_0 = const()
           attn_output_37_transpose_y_0 = const()
0.07% N|C  attn_output_37_cast_fp16 = ios17.matmul(y: ["values_39_cast_fp16"], x: ["softmax_9_cast_fp16"], transpose_y: ["attn_output_37_transpose_y_0"], transpose_x: ["attn_output_37_transpose_x_0"])
           var_1416_perm_0 = const()
           var_1418 = const()
0.07% N|C  transpose_56 = ios17.transpose(x: ["attn_output_37_cast_fp16"], perm: ["var_1416_perm_0"])
0.03% N|C  input_75_cast_fp16 = ios17.reshape(shape: ["var_1418"], x: ["transpose_56"])
           model_transformer_layers_9_attn_out_proj_weight_to_fp16 = const()
0.24% N|C  linear_37_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_9_attn_out_proj_weight_to_fp16"], x: ["input_75_cast_fp16"], bias: ["linear_1_bias_0_to_fp16"])
0.09% N|C  x_195_cast_fp16 = ios17.add(x: ["x_179_cast_fp16"], y: ["linear_37_cast_fp16"])
           var_10_promoted_to_fp16_39 = const()
0.04% N|C  var_1425_cast_fp16 = ios17.pow(x: ["x_195_cast_fp16"], y: ["var_10_promoted_to_fp16_39"])
           var_1426 = const()
0.02% N|C  var_1427_cast_fp16 = ios16.reduce_mean(axes: ["var_1426"], x: ["var_1425_cast_fp16"], keep_dims: ["var_22"])
           var_1428_to_fp16 = const()
0.00% N|C  var_1429_cast_fp16 = ios17.add(y: ["var_1428_to_fp16"], x: ["var_1427_cast_fp16"])
           var_1430_epsilon_0 = const()
0.00% N|C  var_1430_cast_fp16 = ios17.rsqrt(epsilon: ["var_1430_epsilon_0"], x: ["var_1429_cast_fp16"])
0.04% N|C  var_1431_cast_fp16 = ios17.mul(y: ["var_1430_cast_fp16"], x: ["x_195_cast_fp16"])
           model_transformer_layers_9_ffn_norm_weight_to_fp16 = const()
0.04% N|C  input_77_cast_fp16 = ios17.mul(y: ["model_transformer_layers_9_ffn_norm_weight_to_fp16"], x: ["var_1431_cast_fp16"])
           model_transformer_layers_9_ffn_proj_1_weight_to_fp16 = const()
           linear_38_bias_0_to_fp16 = const()
0.15% N|C  linear_38_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_9_ffn_proj_1_weight_to_fp16"], bias: ["linear_38_bias_0_to_fp16"], x: ["input_77_cast_fp16"])
           var_1441_split_sizes_0 = const()
           var_1441_axis_0 = const()
0.17% N|C  var_1441_cast_fp16_0, var_1441_cast_fp16_1 = split(x: ["linear_38_cast_fp16"], axis: ["var_1441_axis_0"], split_sizes: ["var_1441_split_sizes_0"])
0.11% N|C  var_1443_cast_fp16 = ios16.silu(x: ["var_1441_cast_fp16_0"])
0.20% N|C  input_81_cast_fp16 = ios17.mul(y: ["var_1441_cast_fp16_1"], x: ["var_1443_cast_fp16"])
           model_transformer_layers_9_ffn_proj_2_weight_to_fp16 = const()
2.54% N|C  linear_39_cast_fp16 = ios17.linear(x: ["input_81_cast_fp16"], bias: ["linear_1_bias_0_to_fp16"], weight: ["model_transformer_layers_9_ffn_proj_2_weight_to_fp16"])
0.09% N|C  x_199_cast_fp16 = ios17.add(y: ["linear_39_cast_fp16"], x: ["x_195_cast_fp16"])
           var_10_promoted_to_fp16_40 = const()
0.04% N|C  var_1454_cast_fp16 = ios17.pow(x: ["x_199_cast_fp16"], y: ["var_10_promoted_to_fp16_40"])
           var_1455 = const()
0.02% N|C  var_1456_cast_fp16 = ios16.reduce_mean(x: ["var_1454_cast_fp16"], axes: ["var_1455"], keep_dims: ["var_22"])
           var_1457_to_fp16 = const()
0.00% N|C  var_1458_cast_fp16 = ios17.add(x: ["var_1456_cast_fp16"], y: ["var_1457_to_fp16"])
           var_1459_epsilon_0 = const()
0.00% N|C  var_1459_cast_fp16 = ios17.rsqrt(x: ["var_1458_cast_fp16"], epsilon: ["var_1459_epsilon_0"])
0.04% N|C  var_1460_cast_fp16 = ios17.mul(x: ["x_199_cast_fp16"], y: ["var_1459_cast_fp16"])
           model_transformer_layers_10_attn_norm_weight_to_fp16 = const()
0.04% N|C  hidden_states_61_cast_fp16 = ios17.mul(y: ["model_transformer_layers_10_attn_norm_weight_to_fp16"], x: ["var_1460_cast_fp16"])
           model_transformer_layers_10_attn_qkv_proj_weight_to_fp16 = const()
0.38% N|C  linear_40_cast_fp16 = ios17.linear(bias: ["linear_2_bias_0_to_fp16"], x: ["hidden_states_61_cast_fp16"], weight: ["model_transformer_layers_10_attn_qkv_proj_weight_to_fp16"])
           var_1474 = const()
0.05% N|C  qkv_43_cast_fp16 = ios17.reshape(shape: ["var_1474"], x: ["linear_40_cast_fp16"])
           var_1476_perm_0 = const()
           var_1477 = const()
           var_1478_axis_0 = const()
0.11% N|C  transpose_55 = ios17.transpose(x: ["qkv_43_cast_fp16"], perm: ["var_1476_perm_0"])
0.04% N|C  var_1478_cast_fp16_0, var_1478_cast_fp16_1, var_1478_cast_fp16_2 = split(x: ["transpose_55"], axis: ["var_1478_axis_0"], split_sizes: ["var_1477"])
           var_10_promoted_to_fp16_41 = const()
0.03% N|C  var_1484_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_41"], x: ["var_1478_cast_fp16_0"])
           var_1485 = const()
0.02% N|C  var_1486_cast_fp16 = ios16.reduce_mean(axes: ["var_1485"], x: ["var_1484_cast_fp16"], keep_dims: ["var_22"])
           var_1487_to_fp16 = const()
0.00% N|C  var_1488_cast_fp16 = ios17.add(y: ["var_1487_to_fp16"], x: ["var_1486_cast_fp16"])
           var_1489_epsilon_0 = const()
0.00% N|C  var_1489_cast_fp16 = ios17.rsqrt(x: ["var_1488_cast_fp16"], epsilon: ["var_1489_epsilon_0"])
0.03% N|C  var_1490_cast_fp16 = ios17.mul(y: ["var_1489_cast_fp16"], x: ["var_1478_cast_fp16_0"])
           model_transformer_layers_10_attn_q_norm_weight_to_fp16 = const()
0.03% N|C  query_21_cast_fp16 = ios17.mul(y: ["model_transformer_layers_10_attn_q_norm_weight_to_fp16"], x: ["var_1490_cast_fp16"])
           var_10_promoted_to_fp16_42 = const()
0.01% N|C  var_1495_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_42"], x: ["var_1478_cast_fp16_1"])
           var_1496 = const()
0.00% N|C  var_1497_cast_fp16 = ios16.reduce_mean(axes: ["var_1496"], keep_dims: ["var_22"], x: ["var_1495_cast_fp16"])
           var_1498_to_fp16 = const()
0.00% N|C  var_1499_cast_fp16 = ios17.add(x: ["var_1497_cast_fp16"], y: ["var_1498_to_fp16"])
           var_1500_epsilon_0 = const()
0.00% N|C  var_1500_cast_fp16 = ios17.rsqrt(epsilon: ["var_1500_epsilon_0"], x: ["var_1499_cast_fp16"])
0.01% N|C  var_1501_cast_fp16 = ios17.mul(y: ["var_1500_cast_fp16"], x: ["var_1478_cast_fp16_1"])
           model_transformer_layers_10_attn_k_norm_weight_to_fp16 = const()
0.01% N|C  key_21_cast_fp16 = ios17.mul(y: ["model_transformer_layers_10_attn_k_norm_weight_to_fp16"], x: ["var_1501_cast_fp16"])
0.03% N|C  var_1518_cast_fp16 = ios17.mul(x: ["query_21_cast_fp16"], y: ["var_156_to_fp16"])
           var_1519_split_sizes_0 = const()
           var_1519_axis_0 = const()
0.03% N|C  var_1519_cast_fp16_0, var_1519_cast_fp16_1 = split(x: ["query_21_cast_fp16"], axis: ["var_1519_axis_0"], split_sizes: ["var_1519_split_sizes_0"])
           const_76_promoted_to_fp16 = const()
0.02% N|C  var_1521_cast_fp16 = ios17.mul(x: ["var_1519_cast_fp16_1"], y: ["const_76_promoted_to_fp16"])
           var_1523_interleave_0 = const()
0.01% N|C  var_1523_cast_fp16 = ios17.concat(axis: ["var_5"], values: ["var_1521_cast_fp16", "var_1519_cast_fp16_0"], interleave: ["var_1523_interleave_0"])
0.03% N|C  var_1524_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_1523_cast_fp16"])
0.08% N|C  query_float_21_cast_fp16 = ios17.add(x: ["var_1518_cast_fp16"], y: ["var_1524_cast_fp16"])
0.01% N|C  var_1530_cast_fp16 = ios17.mul(x: ["key_21_cast_fp16"], y: ["var_156_to_fp16"])
           var_1531_split_sizes_0 = const()
           var_1531_axis_0 = const()
0.01% N|C  var_1531_cast_fp16_0, var_1531_cast_fp16_1 = split(axis: ["var_1531_axis_0"], split_sizes: ["var_1531_split_sizes_0"], x: ["key_21_cast_fp16"])
           const_77_promoted_to_fp16 = const()
0.00% N|C  var_1533_cast_fp16 = ios17.mul(y: ["const_77_promoted_to_fp16"], x: ["var_1531_cast_fp16_1"])
           var_1535_interleave_0 = const()
0.00% N|C  var_1535_cast_fp16 = ios17.concat(values: ["var_1533_cast_fp16", "var_1531_cast_fp16_0"], interleave: ["var_1535_interleave_0"], axis: ["var_5"])
0.01% N|C  var_1536_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_1535_cast_fp16"])
0.02% N|C  key_float_21_cast_fp16 = ios17.add(x: ["var_1530_cast_fp16"], y: ["var_1536_cast_fp16"])
           reshape_40_shape_0 = const()
0.01% N|C  reshape_40_cast_fp16 = ios17.reshape(x: ["key_float_21_cast_fp16"], shape: ["reshape_40_shape_0"])
           tile_20_reps_0 = const()
0.02% N|C  tile_20_cast_fp16 = tile(reps: ["tile_20_reps_0"], x: ["reshape_40_cast_fp16"])
           reshape_41_shape_0 = const()
0.03% N|C  reshape_41_cast_fp16 = ios17.reshape(shape: ["reshape_41_shape_0"], x: ["tile_20_cast_fp16"])
           transpose_20_perm_0 = const()
           keys_43_shape_0 = const()
0.07% N|C  transpose_54 = ios17.transpose(x: ["reshape_41_cast_fp16"], perm: ["transpose_20_perm_0"])
0.03% N|C  keys_43_cast_fp16 = ios17.reshape(x: ["transpose_54"], shape: ["keys_43_shape_0"])
           reshape_42_shape_0 = const()
0.01% N|C  reshape_42_cast_fp16 = ios17.reshape(shape: ["reshape_42_shape_0"], x: ["var_1478_cast_fp16_2"])
           tile_21_reps_0 = const()
0.02% N|C  tile_21_cast_fp16 = tile(reps: ["tile_21_reps_0"], x: ["reshape_42_cast_fp16"])
           reshape_43_shape_0 = const()
0.03% N|C  reshape_43_cast_fp16 = ios17.reshape(shape: ["reshape_43_shape_0"], x: ["tile_21_cast_fp16"])
           transpose_21_perm_0 = const()
           values_43_shape_0 = const()
0.07% N|C  transpose_53 = ios17.transpose(perm: ["transpose_21_perm_0"], x: ["reshape_43_cast_fp16"])
0.03% N|C  values_43_cast_fp16 = ios17.reshape(x: ["transpose_53"], shape: ["values_43_shape_0"])
           mul_10_y_0_to_fp16 = const()
0.03% N|C  mul_10_cast_fp16 = ios17.mul(y: ["mul_10_y_0_to_fp16"], x: ["query_float_21_cast_fp16"])
           matmul_10_transpose_y_0 = const()
           matmul_10_transpose_x_0 = const()
0.07% N|C  matmul_10_cast_fp16 = ios17.matmul(transpose_y: ["matmul_10_transpose_y_0"], x: ["mul_10_cast_fp16"], y: ["keys_43_cast_fp16"], transpose_x: ["matmul_10_transpose_x_0"])
0.08% N|C  add_21_cast_fp16 = ios17.add(y: ["causal_mask_3_to_fp16"], x: ["matmul_10_cast_fp16"])
           softmax_10_axis_0 = const()
0.07% N|C  softmax_10_cast_fp16 = ios16.softmax(x: ["add_21_cast_fp16"], axis: ["softmax_10_axis_0"])
           attn_output_41_transpose_x_0 = const()
           attn_output_41_transpose_y_0 = const()
0.07% N|C  attn_output_41_cast_fp16 = ios17.matmul(transpose_y: ["attn_output_41_transpose_y_0"], x: ["softmax_10_cast_fp16"], transpose_x: ["attn_output_41_transpose_x_0"], y: ["values_43_cast_fp16"])
           var_1552_perm_0 = const()
           var_1554 = const()
0.07% N|C  transpose_52 = ios17.transpose(perm: ["var_1552_perm_0"], x: ["attn_output_41_cast_fp16"])
0.03% N|C  input_83_cast_fp16 = ios17.reshape(x: ["transpose_52"], shape: ["var_1554"])
           model_transformer_layers_10_attn_out_proj_weight_to_fp16 = const()
0.24% N|C  linear_41_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_10_attn_out_proj_weight_to_fp16"], x: ["input_83_cast_fp16"], bias: ["linear_1_bias_0_to_fp16"])
0.09% N|C  x_215_cast_fp16 = ios17.add(y: ["linear_41_cast_fp16"], x: ["x_199_cast_fp16"])
           var_10_promoted_to_fp16_43 = const()
0.04% N|C  var_1561_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_43"], x: ["x_215_cast_fp16"])
           var_1562 = const()
0.02% N|C  var_1563_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_1561_cast_fp16"], axes: ["var_1562"])
           var_1564_to_fp16 = const()
0.00% N|C  var_1565_cast_fp16 = ios17.add(x: ["var_1563_cast_fp16"], y: ["var_1564_to_fp16"])
           var_1566_epsilon_0 = const()
0.00% N|C  var_1566_cast_fp16 = ios17.rsqrt(epsilon: ["var_1566_epsilon_0"], x: ["var_1565_cast_fp16"])
0.04% N|C  var_1567_cast_fp16 = ios17.mul(y: ["var_1566_cast_fp16"], x: ["x_215_cast_fp16"])
           model_transformer_layers_10_ffn_norm_weight_to_fp16 = const()
0.04% N|C  input_85_cast_fp16 = ios17.mul(y: ["model_transformer_layers_10_ffn_norm_weight_to_fp16"], x: ["var_1567_cast_fp16"])
           model_transformer_layers_10_ffn_proj_1_weight_to_fp16 = const()
           linear_42_bias_0_to_fp16 = const()
0.15% N|C  linear_42_cast_fp16 = ios17.linear(bias: ["linear_42_bias_0_to_fp16"], x: ["input_85_cast_fp16"], weight: ["model_transformer_layers_10_ffn_proj_1_weight_to_fp16"])
           var_1577_split_sizes_0 = const()
           var_1577_axis_0 = const()
0.18% N|C  var_1577_cast_fp16_0, var_1577_cast_fp16_1 = split(split_sizes: ["var_1577_split_sizes_0"], x: ["linear_42_cast_fp16"], axis: ["var_1577_axis_0"])
0.12% N|C  var_1579_cast_fp16 = ios16.silu(x: ["var_1577_cast_fp16_0"])
0.21% N|C  input_89_cast_fp16 = ios17.mul(x: ["var_1579_cast_fp16"], y: ["var_1577_cast_fp16_1"])
           model_transformer_layers_10_ffn_proj_2_weight_to_fp16 = const()
2.95% N|C  linear_43_cast_fp16 = ios17.linear(bias: ["linear_1_bias_0_to_fp16"], weight: ["model_transformer_layers_10_ffn_proj_2_weight_to_fp16"], x: ["input_89_cast_fp16"])
0.09% N|C  x_219_cast_fp16 = ios17.add(x: ["x_215_cast_fp16"], y: ["linear_43_cast_fp16"])
           var_10_promoted_to_fp16_44 = const()
0.04% N|C  var_1590_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_44"], x: ["x_219_cast_fp16"])
           var_1591 = const()
0.02% N|C  var_1592_cast_fp16 = ios16.reduce_mean(axes: ["var_1591"], keep_dims: ["var_22"], x: ["var_1590_cast_fp16"])
           var_1593_to_fp16 = const()
0.00% N|C  var_1594_cast_fp16 = ios17.add(x: ["var_1592_cast_fp16"], y: ["var_1593_to_fp16"])
           var_1595_epsilon_0 = const()
0.00% N|C  var_1595_cast_fp16 = ios17.rsqrt(epsilon: ["var_1595_epsilon_0"], x: ["var_1594_cast_fp16"])
0.04% N|C  var_1596_cast_fp16 = ios17.mul(x: ["x_219_cast_fp16"], y: ["var_1595_cast_fp16"])
           model_transformer_layers_11_attn_norm_weight_to_fp16 = const()
0.04% N|C  hidden_states_67_cast_fp16 = ios17.mul(x: ["var_1596_cast_fp16"], y: ["model_transformer_layers_11_attn_norm_weight_to_fp16"])
           model_transformer_layers_11_attn_qkv_proj_weight_to_fp16 = const()
0.38% N|C  linear_44_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_11_attn_qkv_proj_weight_to_fp16"], x: ["hidden_states_67_cast_fp16"], bias: ["linear_2_bias_0_to_fp16"])
           var_1610 = const()
0.05% N|C  qkv_47_cast_fp16 = ios17.reshape(x: ["linear_44_cast_fp16"], shape: ["var_1610"])
           var_1612_perm_0 = const()
           var_1613 = const()
           var_1614_axis_0 = const()
0.11% N|C  transpose_51 = ios17.transpose(perm: ["var_1612_perm_0"], x: ["qkv_47_cast_fp16"])
0.04% N|C  var_1614_cast_fp16_0, var_1614_cast_fp16_1, var_1614_cast_fp16_2 = split(x: ["transpose_51"], axis: ["var_1614_axis_0"], split_sizes: ["var_1613"])
           var_10_promoted_to_fp16_45 = const()
0.03% N|C  var_1620_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_45"], x: ["var_1614_cast_fp16_0"])
           var_1621 = const()
0.02% N|C  var_1622_cast_fp16 = ios16.reduce_mean(axes: ["var_1621"], keep_dims: ["var_22"], x: ["var_1620_cast_fp16"])
           var_1623_to_fp16 = const()
0.00% N|C  var_1624_cast_fp16 = ios17.add(y: ["var_1623_to_fp16"], x: ["var_1622_cast_fp16"])
           var_1625_epsilon_0 = const()
0.00% N|C  var_1625_cast_fp16 = ios17.rsqrt(epsilon: ["var_1625_epsilon_0"], x: ["var_1624_cast_fp16"])
0.03% N|C  var_1626_cast_fp16 = ios17.mul(x: ["var_1614_cast_fp16_0"], y: ["var_1625_cast_fp16"])
           model_transformer_layers_11_attn_q_norm_weight_to_fp16 = const()
0.03% N|C  query_23_cast_fp16 = ios17.mul(x: ["var_1626_cast_fp16"], y: ["model_transformer_layers_11_attn_q_norm_weight_to_fp16"])
           var_10_promoted_to_fp16_46 = const()
0.01% N|C  var_1631_cast_fp16 = ios17.pow(x: ["var_1614_cast_fp16_1"], y: ["var_10_promoted_to_fp16_46"])
           var_1632 = const()
0.00% N|C  var_1633_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_1631_cast_fp16"], axes: ["var_1632"])
           var_1634_to_fp16 = const()
0.00% N|C  var_1635_cast_fp16 = ios17.add(x: ["var_1633_cast_fp16"], y: ["var_1634_to_fp16"])
           var_1636_epsilon_0 = const()
0.00% N|C  var_1636_cast_fp16 = ios17.rsqrt(x: ["var_1635_cast_fp16"], epsilon: ["var_1636_epsilon_0"])
0.01% N|C  var_1637_cast_fp16 = ios17.mul(x: ["var_1614_cast_fp16_1"], y: ["var_1636_cast_fp16"])
           model_transformer_layers_11_attn_k_norm_weight_to_fp16 = const()
0.01% N|C  key_23_cast_fp16 = ios17.mul(y: ["model_transformer_layers_11_attn_k_norm_weight_to_fp16"], x: ["var_1637_cast_fp16"])
0.03% N|C  var_1654_cast_fp16 = ios17.mul(x: ["query_23_cast_fp16"], y: ["var_156_to_fp16"])
           var_1655_split_sizes_0 = const()
           var_1655_axis_0 = const()
0.03% N|C  var_1655_cast_fp16_0, var_1655_cast_fp16_1 = split(split_sizes: ["var_1655_split_sizes_0"], x: ["query_23_cast_fp16"], axis: ["var_1655_axis_0"])
           const_83_promoted_to_fp16 = const()
0.02% N|C  var_1657_cast_fp16 = ios17.mul(y: ["const_83_promoted_to_fp16"], x: ["var_1655_cast_fp16_1"])
           var_1659_interleave_0 = const()
0.01% N|C  var_1659_cast_fp16 = ios17.concat(axis: ["var_5"], values: ["var_1657_cast_fp16", "var_1655_cast_fp16_0"], interleave: ["var_1659_interleave_0"])
0.03% N|C  var_1660_cast_fp16 = ios17.mul(x: ["var_1659_cast_fp16"], y: ["var_152_to_fp16"])
0.08% N|C  query_float_23_cast_fp16 = ios17.add(y: ["var_1660_cast_fp16"], x: ["var_1654_cast_fp16"])
0.01% N|C  var_1666_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["key_23_cast_fp16"])
           var_1667_split_sizes_0 = const()
           var_1667_axis_0 = const()
0.01% N|C  var_1667_cast_fp16_0, var_1667_cast_fp16_1 = split(axis: ["var_1667_axis_0"], split_sizes: ["var_1667_split_sizes_0"], x: ["key_23_cast_fp16"])
           const_84_promoted_to_fp16 = const()
0.00% N|C  var_1669_cast_fp16 = ios17.mul(x: ["var_1667_cast_fp16_1"], y: ["const_84_promoted_to_fp16"])
           var_1671_interleave_0 = const()
0.00% N|C  var_1671_cast_fp16 = ios17.concat(values: ["var_1669_cast_fp16", "var_1667_cast_fp16_0"], interleave: ["var_1671_interleave_0"], axis: ["var_5"])
0.01% N|C  var_1672_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_1671_cast_fp16"])
0.02% N|C  key_float_23_cast_fp16 = ios17.add(x: ["var_1666_cast_fp16"], y: ["var_1672_cast_fp16"])
           reshape_44_shape_0 = const()
0.01% N|C  reshape_44_cast_fp16 = ios17.reshape(x: ["key_float_23_cast_fp16"], shape: ["reshape_44_shape_0"])
           tile_22_reps_0 = const()
0.02% N|C  tile_22_cast_fp16 = tile(reps: ["tile_22_reps_0"], x: ["reshape_44_cast_fp16"])
           reshape_45_shape_0 = const()
0.03% N|C  reshape_45_cast_fp16 = ios17.reshape(x: ["tile_22_cast_fp16"], shape: ["reshape_45_shape_0"])
           transpose_22_perm_0 = const()
           keys_47_shape_0 = const()
0.07% N|C  transpose_50 = ios17.transpose(perm: ["transpose_22_perm_0"], x: ["reshape_45_cast_fp16"])
0.03% N|C  keys_47_cast_fp16 = ios17.reshape(x: ["transpose_50"], shape: ["keys_47_shape_0"])
           reshape_46_shape_0 = const()
0.01% N|C  reshape_46_cast_fp16 = ios17.reshape(x: ["var_1614_cast_fp16_2"], shape: ["reshape_46_shape_0"])
           tile_23_reps_0 = const()
0.02% N|C  tile_23_cast_fp16 = tile(x: ["reshape_46_cast_fp16"], reps: ["tile_23_reps_0"])
           reshape_47_shape_0 = const()
0.03% N|C  reshape_47_cast_fp16 = ios17.reshape(x: ["tile_23_cast_fp16"], shape: ["reshape_47_shape_0"])
           transpose_23_perm_0 = const()
           values_47_shape_0 = const()
0.07% N|C  transpose_49 = ios17.transpose(x: ["reshape_47_cast_fp16"], perm: ["transpose_23_perm_0"])
0.03% N|C  values_47_cast_fp16 = ios17.reshape(x: ["transpose_49"], shape: ["values_47_shape_0"])
           mul_11_y_0_to_fp16 = const()
0.03% N|C  mul_11_cast_fp16 = ios17.mul(y: ["mul_11_y_0_to_fp16"], x: ["query_float_23_cast_fp16"])
           matmul_11_transpose_y_0 = const()
           matmul_11_transpose_x_0 = const()
0.07% N|C  matmul_11_cast_fp16 = ios17.matmul(transpose_x: ["matmul_11_transpose_x_0"], transpose_y: ["matmul_11_transpose_y_0"], y: ["keys_47_cast_fp16"], x: ["mul_11_cast_fp16"])
0.08% N|C  add_23_cast_fp16 = ios17.add(y: ["causal_mask_3_to_fp16"], x: ["matmul_11_cast_fp16"])
           softmax_11_axis_0 = const()
0.07% N|C  softmax_11_cast_fp16 = ios16.softmax(x: ["add_23_cast_fp16"], axis: ["softmax_11_axis_0"])
           attn_output_45_transpose_x_0 = const()
           attn_output_45_transpose_y_0 = const()
0.07% N|C  attn_output_45_cast_fp16 = ios17.matmul(x: ["softmax_11_cast_fp16"], transpose_x: ["attn_output_45_transpose_x_0"], y: ["values_47_cast_fp16"], transpose_y: ["attn_output_45_transpose_y_0"])
           var_1688_perm_0 = const()
           var_1690 = const()
0.07% N|C  transpose_48 = ios17.transpose(x: ["attn_output_45_cast_fp16"], perm: ["var_1688_perm_0"])
0.03% N|C  input_91_cast_fp16 = ios17.reshape(x: ["transpose_48"], shape: ["var_1690"])
           model_transformer_layers_11_attn_out_proj_weight_to_fp16 = const()
0.24% N|C  linear_45_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_11_attn_out_proj_weight_to_fp16"], x: ["input_91_cast_fp16"], bias: ["linear_1_bias_0_to_fp16"])
0.09% N|C  x_235_cast_fp16 = ios17.add(y: ["linear_45_cast_fp16"], x: ["x_219_cast_fp16"])
           var_10_promoted_to_fp16_47 = const()
0.04% N|C  var_1697_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_47"], x: ["x_235_cast_fp16"])
           var_1698 = const()
0.02% N|C  var_1699_cast_fp16 = ios16.reduce_mean(axes: ["var_1698"], x: ["var_1697_cast_fp16"], keep_dims: ["var_22"])
           var_1700_to_fp16 = const()
0.00% N|C  var_1701_cast_fp16 = ios17.add(x: ["var_1699_cast_fp16"], y: ["var_1700_to_fp16"])
           var_1702_epsilon_0 = const()
0.00% N|C  var_1702_cast_fp16 = ios17.rsqrt(epsilon: ["var_1702_epsilon_0"], x: ["var_1701_cast_fp16"])
0.04% N|C  var_1703_cast_fp16 = ios17.mul(y: ["var_1702_cast_fp16"], x: ["x_235_cast_fp16"])
           model_transformer_layers_11_ffn_norm_weight_to_fp16 = const()
0.04% N|C  input_93_cast_fp16 = ios17.mul(y: ["model_transformer_layers_11_ffn_norm_weight_to_fp16"], x: ["var_1703_cast_fp16"])
           model_transformer_layers_11_ffn_proj_1_weight_to_fp16 = const()
           linear_46_bias_0_to_fp16 = const()
0.15% N|C  linear_46_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_11_ffn_proj_1_weight_to_fp16"], bias: ["linear_46_bias_0_to_fp16"], x: ["input_93_cast_fp16"])
           var_1713_split_sizes_0 = const()
           var_1713_axis_0 = const()
0.19% N|C  var_1713_cast_fp16_0, var_1713_cast_fp16_1 = split(x: ["linear_46_cast_fp16"], split_sizes: ["var_1713_split_sizes_0"], axis: ["var_1713_axis_0"])
0.13% N|C  var_1715_cast_fp16 = ios16.silu(x: ["var_1713_cast_fp16_0"])
0.23% N|C  input_97_cast_fp16 = ios17.mul(y: ["var_1713_cast_fp16_1"], x: ["var_1715_cast_fp16"])
           model_transformer_layers_11_ffn_proj_2_weight_to_fp16 = const()
3.38% N|C  linear_47_cast_fp16 = ios17.linear(bias: ["linear_1_bias_0_to_fp16"], weight: ["model_transformer_layers_11_ffn_proj_2_weight_to_fp16"], x: ["input_97_cast_fp16"])
0.09% N|C  x_239_cast_fp16 = ios17.add(y: ["linear_47_cast_fp16"], x: ["x_235_cast_fp16"])
           var_10_promoted_to_fp16_48 = const()
0.04% N|C  var_1726_cast_fp16 = ios17.pow(x: ["x_239_cast_fp16"], y: ["var_10_promoted_to_fp16_48"])
           var_1727 = const()
0.02% N|C  var_1728_cast_fp16 = ios16.reduce_mean(x: ["var_1726_cast_fp16"], keep_dims: ["var_22"], axes: ["var_1727"])
           var_1729_to_fp16 = const()
0.00% N|C  var_1730_cast_fp16 = ios17.add(x: ["var_1728_cast_fp16"], y: ["var_1729_to_fp16"])
           var_1731_epsilon_0 = const()
0.00% N|C  var_1731_cast_fp16 = ios17.rsqrt(epsilon: ["var_1731_epsilon_0"], x: ["var_1730_cast_fp16"])
0.04% N|C  var_1732_cast_fp16 = ios17.mul(y: ["var_1731_cast_fp16"], x: ["x_239_cast_fp16"])
           model_transformer_layers_12_attn_norm_weight_to_fp16 = const()
0.04% N|C  hidden_states_73_cast_fp16 = ios17.mul(y: ["model_transformer_layers_12_attn_norm_weight_to_fp16"], x: ["var_1732_cast_fp16"])
           model_transformer_layers_12_attn_qkv_proj_weight_to_fp16 = const()
           linear_48_bias_0_to_fp16 = const()
0.38% N|C  linear_48_cast_fp16 = ios17.linear(bias: ["linear_48_bias_0_to_fp16"], x: ["hidden_states_73_cast_fp16"], weight: ["model_transformer_layers_12_attn_qkv_proj_weight_to_fp16"])
           var_1746 = const()
0.06% N|C  qkv_51_cast_fp16 = ios17.reshape(shape: ["var_1746"], x: ["linear_48_cast_fp16"])
           var_1748_perm_0 = const()
           var_1749 = const()
           var_1750_axis_0 = const()
0.13% N|C  transpose_47 = ios17.transpose(x: ["qkv_51_cast_fp16"], perm: ["var_1748_perm_0"])
0.05% N|C  var_1750_cast_fp16_0, var_1750_cast_fp16_1, var_1750_cast_fp16_2 = split(x: ["transpose_47"], split_sizes: ["var_1749"], axis: ["var_1750_axis_0"])
           var_10_promoted_to_fp16_49 = const()
0.04% N|C  var_1756_cast_fp16 = ios17.pow(x: ["var_1750_cast_fp16_0"], y: ["var_10_promoted_to_fp16_49"])
           var_1757 = const()
0.02% N|C  var_1758_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_1756_cast_fp16"], axes: ["var_1757"])
           var_1759_to_fp16 = const()
0.00% N|C  var_1760_cast_fp16 = ios17.add(x: ["var_1758_cast_fp16"], y: ["var_1759_to_fp16"])
           var_1761_epsilon_0 = const()
0.00% N|C  var_1761_cast_fp16 = ios17.rsqrt(epsilon: ["var_1761_epsilon_0"], x: ["var_1760_cast_fp16"])
0.04% N|C  var_1762_cast_fp16 = ios17.mul(y: ["var_1761_cast_fp16"], x: ["var_1750_cast_fp16_0"])
           model_transformer_layers_12_attn_q_norm_weight_to_fp16 = const()
0.04% N|C  query_25_cast_fp16 = ios17.mul(y: ["model_transformer_layers_12_attn_q_norm_weight_to_fp16"], x: ["var_1762_cast_fp16"])
           var_10_promoted_to_fp16_50 = const()
0.01% N|C  var_1767_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_50"], x: ["var_1750_cast_fp16_1"])
           var_1768 = const()
0.01% N|C  var_1769_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_1767_cast_fp16"], axes: ["var_1768"])
           var_1770_to_fp16 = const()
0.00% N|C  var_1771_cast_fp16 = ios17.add(y: ["var_1770_to_fp16"], x: ["var_1769_cast_fp16"])
           var_1772_epsilon_0 = const()
0.00% N|C  var_1772_cast_fp16 = ios17.rsqrt(x: ["var_1771_cast_fp16"], epsilon: ["var_1772_epsilon_0"])
0.01% N|C  var_1773_cast_fp16 = ios17.mul(y: ["var_1772_cast_fp16"], x: ["var_1750_cast_fp16_1"])
           model_transformer_layers_12_attn_k_norm_weight_to_fp16 = const()
0.01% N|C  key_25_cast_fp16 = ios17.mul(x: ["var_1773_cast_fp16"], y: ["model_transformer_layers_12_attn_k_norm_weight_to_fp16"])
0.04% N|C  var_1790_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["query_25_cast_fp16"])
           var_1791_split_sizes_0 = const()
           var_1791_axis_0 = const()
0.03% N|C  var_1791_cast_fp16_0, var_1791_cast_fp16_1 = split(axis: ["var_1791_axis_0"], split_sizes: ["var_1791_split_sizes_0"], x: ["query_25_cast_fp16"])
           const_90_promoted_to_fp16 = const()
0.02% N|C  var_1793_cast_fp16 = ios17.mul(y: ["const_90_promoted_to_fp16"], x: ["var_1791_cast_fp16_1"])
           var_1795_interleave_0 = const()
0.02% N|C  var_1795_cast_fp16 = ios17.concat(axis: ["var_5"], interleave: ["var_1795_interleave_0"], values: ["var_1793_cast_fp16", "var_1791_cast_fp16_0"])
0.04% N|C  var_1796_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_1795_cast_fp16"])
0.09% N|C  query_float_25_cast_fp16 = ios17.add(y: ["var_1796_cast_fp16"], x: ["var_1790_cast_fp16"])
0.01% N|C  var_1802_cast_fp16 = ios17.mul(x: ["key_25_cast_fp16"], y: ["var_156_to_fp16"])
           var_1803_split_sizes_0 = const()
           var_1803_axis_0 = const()
0.01% N|C  var_1803_cast_fp16_0, var_1803_cast_fp16_1 = split(x: ["key_25_cast_fp16"], axis: ["var_1803_axis_0"], split_sizes: ["var_1803_split_sizes_0"])
           const_91_promoted_to_fp16 = const()
0.00% N|C  var_1805_cast_fp16 = ios17.mul(y: ["const_91_promoted_to_fp16"], x: ["var_1803_cast_fp16_1"])
           var_1807_interleave_0 = const()
0.00% N|C  var_1807_cast_fp16 = ios17.concat(values: ["var_1805_cast_fp16", "var_1803_cast_fp16_0"], axis: ["var_5"], interleave: ["var_1807_interleave_0"])
0.01% N|C  var_1808_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_1807_cast_fp16"])
0.02% N|C  key_float_25_cast_fp16 = ios17.add(y: ["var_1808_cast_fp16"], x: ["var_1802_cast_fp16"])
           reshape_48_shape_0 = const()
0.01% N|C  reshape_48_cast_fp16 = ios17.reshape(x: ["key_float_25_cast_fp16"], shape: ["reshape_48_shape_0"])
           tile_24_reps_0 = const()
0.03% N|C  tile_24_cast_fp16 = tile(reps: ["tile_24_reps_0"], x: ["reshape_48_cast_fp16"])
           reshape_49_shape_0 = const()
0.04% N|C  reshape_49_cast_fp16 = ios17.reshape(shape: ["reshape_49_shape_0"], x: ["tile_24_cast_fp16"])
           transpose_24_perm_0 = const()
           keys_51_shape_0 = const()
0.09% N|C  transpose_46 = ios17.transpose(perm: ["transpose_24_perm_0"], x: ["reshape_49_cast_fp16"])
0.04% N|C  keys_51_cast_fp16 = ios17.reshape(x: ["transpose_46"], shape: ["keys_51_shape_0"])
           reshape_50_shape_0 = const()
0.01% N|C  reshape_50_cast_fp16 = ios17.reshape(x: ["var_1750_cast_fp16_2"], shape: ["reshape_50_shape_0"])
           tile_25_reps_0 = const()
0.03% N|C  tile_25_cast_fp16 = tile(reps: ["tile_25_reps_0"], x: ["reshape_50_cast_fp16"])
           reshape_51_shape_0 = const()
0.04% N|C  reshape_51_cast_fp16 = ios17.reshape(x: ["tile_25_cast_fp16"], shape: ["reshape_51_shape_0"])
           transpose_25_perm_0 = const()
           values_51_shape_0 = const()
0.09% N|C  transpose_45 = ios17.transpose(x: ["reshape_51_cast_fp16"], perm: ["transpose_25_perm_0"])
0.04% N|C  values_51_cast_fp16 = ios17.reshape(x: ["transpose_45"], shape: ["values_51_shape_0"])
           mul_12_y_0_to_fp16 = const()
0.04% N|C  mul_12_cast_fp16 = ios17.mul(x: ["query_float_25_cast_fp16"], y: ["mul_12_y_0_to_fp16"])
           matmul_12_transpose_y_0 = const()
           matmul_12_transpose_x_0 = const()
0.09% N|C  matmul_12_cast_fp16 = ios17.matmul(transpose_y: ["matmul_12_transpose_y_0"], y: ["keys_51_cast_fp16"], transpose_x: ["matmul_12_transpose_x_0"], x: ["mul_12_cast_fp16"])
0.10% N|C  add_25_cast_fp16 = ios17.add(x: ["matmul_12_cast_fp16"], y: ["causal_mask_3_to_fp16"])
           softmax_12_axis_0 = const()
0.09% N|C  softmax_12_cast_fp16 = ios16.softmax(axis: ["softmax_12_axis_0"], x: ["add_25_cast_fp16"])
           attn_output_49_transpose_x_0 = const()
           attn_output_49_transpose_y_0 = const()
0.09% N|C  attn_output_49_cast_fp16 = ios17.matmul(x: ["softmax_12_cast_fp16"], y: ["values_51_cast_fp16"], transpose_x: ["attn_output_49_transpose_x_0"], transpose_y: ["attn_output_49_transpose_y_0"])
           var_1824_perm_0 = const()
           var_1826 = const()
0.09% N|C  transpose_44 = ios17.transpose(perm: ["var_1824_perm_0"], x: ["attn_output_49_cast_fp16"])
0.04% N|C  input_99_cast_fp16 = ios17.reshape(x: ["transpose_44"], shape: ["var_1826"])
           model_transformer_layers_12_attn_out_proj_weight_to_fp16 = const()
0.38% N|C  linear_49_cast_fp16 = ios17.linear(bias: ["linear_1_bias_0_to_fp16"], weight: ["model_transformer_layers_12_attn_out_proj_weight_to_fp16"], x: ["input_99_cast_fp16"])
0.09% N|C  x_255_cast_fp16 = ios17.add(y: ["linear_49_cast_fp16"], x: ["x_239_cast_fp16"])
           var_10_promoted_to_fp16_51 = const()
0.04% N|C  var_1833_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_51"], x: ["x_255_cast_fp16"])
           var_1834 = const()
0.02% N|C  var_1835_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_1833_cast_fp16"], axes: ["var_1834"])
           var_1836_to_fp16 = const()
0.00% N|C  var_1837_cast_fp16 = ios17.add(x: ["var_1835_cast_fp16"], y: ["var_1836_to_fp16"])
           var_1838_epsilon_0 = const()
0.00% N|C  var_1838_cast_fp16 = ios17.rsqrt(x: ["var_1837_cast_fp16"], epsilon: ["var_1838_epsilon_0"])
0.04% N|C  var_1839_cast_fp16 = ios17.mul(y: ["var_1838_cast_fp16"], x: ["x_255_cast_fp16"])
           model_transformer_layers_12_ffn_norm_weight_to_fp16 = const()
0.04% N|C  input_101_cast_fp16 = ios17.mul(y: ["model_transformer_layers_12_ffn_norm_weight_to_fp16"], x: ["var_1839_cast_fp16"])
           model_transformer_layers_12_ffn_proj_1_weight_to_fp16 = const()
           linear_50_bias_0_to_fp16 = const()
0.15% N|C  linear_50_cast_fp16 = ios17.linear(bias: ["linear_50_bias_0_to_fp16"], weight: ["model_transformer_layers_12_ffn_proj_1_weight_to_fp16"], x: ["input_101_cast_fp16"])
           var_1849_split_sizes_0 = const()
           var_1849_axis_0 = const()
0.22% N|C  var_1849_cast_fp16_0, var_1849_cast_fp16_1 = split(axis: ["var_1849_axis_0"], x: ["linear_50_cast_fp16"], split_sizes: ["var_1849_split_sizes_0"])
0.14% N|C  var_1851_cast_fp16 = ios16.silu(x: ["var_1849_cast_fp16_0"])
0.26% N|C  input_105_cast_fp16 = ios17.mul(y: ["var_1849_cast_fp16_1"], x: ["var_1851_cast_fp16"])
           model_transformer_layers_12_ffn_proj_2_weight_to_fp16 = const()
4.34% N|C  linear_51_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_12_ffn_proj_2_weight_to_fp16"], bias: ["linear_1_bias_0_to_fp16"], x: ["input_105_cast_fp16"])
0.09% N|C  x_259_cast_fp16 = ios17.add(x: ["x_255_cast_fp16"], y: ["linear_51_cast_fp16"])
           var_10_promoted_to_fp16_52 = const()
0.04% N|C  var_1862_cast_fp16 = ios17.pow(x: ["x_259_cast_fp16"], y: ["var_10_promoted_to_fp16_52"])
           var_1863 = const()
0.02% N|C  var_1864_cast_fp16 = ios16.reduce_mean(x: ["var_1862_cast_fp16"], axes: ["var_1863"], keep_dims: ["var_22"])
           var_1865_to_fp16 = const()
0.00% N|C  var_1866_cast_fp16 = ios17.add(y: ["var_1865_to_fp16"], x: ["var_1864_cast_fp16"])
           var_1867_epsilon_0 = const()
0.00% N|C  var_1867_cast_fp16 = ios17.rsqrt(x: ["var_1866_cast_fp16"], epsilon: ["var_1867_epsilon_0"])
0.04% N|C  var_1868_cast_fp16 = ios17.mul(y: ["var_1867_cast_fp16"], x: ["x_259_cast_fp16"])
           model_transformer_layers_13_attn_norm_weight_to_fp16 = const()
0.04% N|C  hidden_states_79_cast_fp16 = ios17.mul(x: ["var_1868_cast_fp16"], y: ["model_transformer_layers_13_attn_norm_weight_to_fp16"])
           model_transformer_layers_13_attn_qkv_proj_weight_to_fp16 = const()
0.38% N|C  linear_52_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_13_attn_qkv_proj_weight_to_fp16"], x: ["hidden_states_79_cast_fp16"], bias: ["linear_48_bias_0_to_fp16"])
           var_1882 = const()
0.06% N|C  qkv_55_cast_fp16 = ios17.reshape(shape: ["var_1882"], x: ["linear_52_cast_fp16"])
           var_1884_perm_0 = const()
           var_1885 = const()
           var_1886_axis_0 = const()
0.13% N|C  transpose_43 = ios17.transpose(x: ["qkv_55_cast_fp16"], perm: ["var_1884_perm_0"])
0.05% N|C  var_1886_cast_fp16_0, var_1886_cast_fp16_1, var_1886_cast_fp16_2 = split(split_sizes: ["var_1885"], axis: ["var_1886_axis_0"], x: ["transpose_43"])
           var_10_promoted_to_fp16_53 = const()
0.04% N|C  var_1892_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_53"], x: ["var_1886_cast_fp16_0"])
           var_1893 = const()
0.02% N|C  var_1894_cast_fp16 = ios16.reduce_mean(axes: ["var_1893"], x: ["var_1892_cast_fp16"], keep_dims: ["var_22"])
           var_1895_to_fp16 = const()
0.00% N|C  var_1896_cast_fp16 = ios17.add(x: ["var_1894_cast_fp16"], y: ["var_1895_to_fp16"])
           var_1897_epsilon_0 = const()
0.00% N|C  var_1897_cast_fp16 = ios17.rsqrt(x: ["var_1896_cast_fp16"], epsilon: ["var_1897_epsilon_0"])
0.04% N|C  var_1898_cast_fp16 = ios17.mul(x: ["var_1886_cast_fp16_0"], y: ["var_1897_cast_fp16"])
           model_transformer_layers_13_attn_q_norm_weight_to_fp16 = const()
0.04% N|C  query_27_cast_fp16 = ios17.mul(y: ["model_transformer_layers_13_attn_q_norm_weight_to_fp16"], x: ["var_1898_cast_fp16"])
           var_10_promoted_to_fp16_54 = const()
0.01% N|C  var_1903_cast_fp16 = ios17.pow(x: ["var_1886_cast_fp16_1"], y: ["var_10_promoted_to_fp16_54"])
           var_1904 = const()
0.01% N|C  var_1905_cast_fp16 = ios16.reduce_mean(x: ["var_1903_cast_fp16"], axes: ["var_1904"], keep_dims: ["var_22"])
           var_1906_to_fp16 = const()
0.00% N|C  var_1907_cast_fp16 = ios17.add(y: ["var_1906_to_fp16"], x: ["var_1905_cast_fp16"])
           var_1908_epsilon_0 = const()
0.00% N|C  var_1908_cast_fp16 = ios17.rsqrt(x: ["var_1907_cast_fp16"], epsilon: ["var_1908_epsilon_0"])
0.01% N|C  var_1909_cast_fp16 = ios17.mul(y: ["var_1908_cast_fp16"], x: ["var_1886_cast_fp16_1"])
           model_transformer_layers_13_attn_k_norm_weight_to_fp16 = const()
0.01% N|C  key_27_cast_fp16 = ios17.mul(x: ["var_1909_cast_fp16"], y: ["model_transformer_layers_13_attn_k_norm_weight_to_fp16"])
0.04% N|C  var_1926_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["query_27_cast_fp16"])
           var_1927_split_sizes_0 = const()
           var_1927_axis_0 = const()
0.03% N|C  var_1927_cast_fp16_0, var_1927_cast_fp16_1 = split(split_sizes: ["var_1927_split_sizes_0"], x: ["query_27_cast_fp16"], axis: ["var_1927_axis_0"])
           const_97_promoted_to_fp16 = const()
0.02% N|C  var_1929_cast_fp16 = ios17.mul(y: ["const_97_promoted_to_fp16"], x: ["var_1927_cast_fp16_1"])
           var_1931_interleave_0 = const()
0.02% N|C  var_1931_cast_fp16 = ios17.concat(values: ["var_1929_cast_fp16", "var_1927_cast_fp16_0"], interleave: ["var_1931_interleave_0"], axis: ["var_5"])
0.04% N|C  var_1932_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_1931_cast_fp16"])
0.09% N|C  query_float_27_cast_fp16 = ios17.add(y: ["var_1932_cast_fp16"], x: ["var_1926_cast_fp16"])
0.01% N|C  var_1938_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["key_27_cast_fp16"])
           var_1939_split_sizes_0 = const()
           var_1939_axis_0 = const()
0.01% N|C  var_1939_cast_fp16_0, var_1939_cast_fp16_1 = split(axis: ["var_1939_axis_0"], split_sizes: ["var_1939_split_sizes_0"], x: ["key_27_cast_fp16"])
           const_98_promoted_to_fp16 = const()
0.00% N|C  var_1941_cast_fp16 = ios17.mul(x: ["var_1939_cast_fp16_1"], y: ["const_98_promoted_to_fp16"])
           var_1943_interleave_0 = const()
0.00% N|C  var_1943_cast_fp16 = ios17.concat(interleave: ["var_1943_interleave_0"], values: ["var_1941_cast_fp16", "var_1939_cast_fp16_0"], axis: ["var_5"])
0.01% N|C  var_1944_cast_fp16 = ios17.mul(x: ["var_1943_cast_fp16"], y: ["var_152_to_fp16"])
0.02% N|C  key_float_27_cast_fp16 = ios17.add(y: ["var_1944_cast_fp16"], x: ["var_1938_cast_fp16"])
           reshape_52_shape_0 = const()
0.01% N|C  reshape_52_cast_fp16 = ios17.reshape(shape: ["reshape_52_shape_0"], x: ["key_float_27_cast_fp16"])
           tile_26_reps_0 = const()
0.03% N|C  tile_26_cast_fp16 = tile(reps: ["tile_26_reps_0"], x: ["reshape_52_cast_fp16"])
           reshape_53_shape_0 = const()
0.04% N|C  reshape_53_cast_fp16 = ios17.reshape(x: ["tile_26_cast_fp16"], shape: ["reshape_53_shape_0"])
           transpose_26_perm_0 = const()
           keys_55_shape_0 = const()
0.09% N|C  transpose_42 = ios17.transpose(perm: ["transpose_26_perm_0"], x: ["reshape_53_cast_fp16"])
0.04% N|C  keys_55_cast_fp16 = ios17.reshape(shape: ["keys_55_shape_0"], x: ["transpose_42"])
           reshape_54_shape_0 = const()
0.01% N|C  reshape_54_cast_fp16 = ios17.reshape(shape: ["reshape_54_shape_0"], x: ["var_1886_cast_fp16_2"])
           tile_27_reps_0 = const()
0.03% N|C  tile_27_cast_fp16 = tile(x: ["reshape_54_cast_fp16"], reps: ["tile_27_reps_0"])
           reshape_55_shape_0 = const()
0.04% N|C  reshape_55_cast_fp16 = ios17.reshape(x: ["tile_27_cast_fp16"], shape: ["reshape_55_shape_0"])
           transpose_27_perm_0 = const()
           values_55_shape_0 = const()
0.09% N|C  transpose_41 = ios17.transpose(x: ["reshape_55_cast_fp16"], perm: ["transpose_27_perm_0"])
0.04% N|C  values_55_cast_fp16 = ios17.reshape(shape: ["values_55_shape_0"], x: ["transpose_41"])
           mul_13_y_0_to_fp16 = const()
0.04% N|C  mul_13_cast_fp16 = ios17.mul(x: ["query_float_27_cast_fp16"], y: ["mul_13_y_0_to_fp16"])
           matmul_13_transpose_y_0 = const()
           matmul_13_transpose_x_0 = const()
0.09% N|C  matmul_13_cast_fp16 = ios17.matmul(y: ["keys_55_cast_fp16"], x: ["mul_13_cast_fp16"], transpose_x: ["matmul_13_transpose_x_0"], transpose_y: ["matmul_13_transpose_y_0"])
0.10% N|C  add_27_cast_fp16 = ios17.add(y: ["causal_mask_3_to_fp16"], x: ["matmul_13_cast_fp16"])
           softmax_13_axis_0 = const()
0.09% N|C  softmax_13_cast_fp16 = ios16.softmax(x: ["add_27_cast_fp16"], axis: ["softmax_13_axis_0"])
           attn_output_53_transpose_x_0 = const()
           attn_output_53_transpose_y_0 = const()
0.09% N|C  attn_output_53_cast_fp16 = ios17.matmul(x: ["softmax_13_cast_fp16"], y: ["values_55_cast_fp16"], transpose_y: ["attn_output_53_transpose_y_0"], transpose_x: ["attn_output_53_transpose_x_0"])
           var_1960_perm_0 = const()
           var_1962 = const()
0.09% N|C  transpose_40 = ios17.transpose(x: ["attn_output_53_cast_fp16"], perm: ["var_1960_perm_0"])
0.04% N|C  input_107_cast_fp16 = ios17.reshape(x: ["transpose_40"], shape: ["var_1962"])
           model_transformer_layers_13_attn_out_proj_weight_to_fp16 = const()
0.38% N|C  linear_53_cast_fp16 = ios17.linear(bias: ["linear_1_bias_0_to_fp16"], weight: ["model_transformer_layers_13_attn_out_proj_weight_to_fp16"], x: ["input_107_cast_fp16"])
0.09% N|C  x_275_cast_fp16 = ios17.add(y: ["linear_53_cast_fp16"], x: ["x_259_cast_fp16"])
           var_10_promoted_to_fp16_55 = const()
0.04% N|C  var_1969_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_55"], x: ["x_275_cast_fp16"])
           var_1970 = const()
0.02% N|C  var_1971_cast_fp16 = ios16.reduce_mean(axes: ["var_1970"], keep_dims: ["var_22"], x: ["var_1969_cast_fp16"])
           var_1972_to_fp16 = const()
0.00% N|C  var_1973_cast_fp16 = ios17.add(y: ["var_1972_to_fp16"], x: ["var_1971_cast_fp16"])
           var_1974_epsilon_0 = const()
0.00% N|C  var_1974_cast_fp16 = ios17.rsqrt(x: ["var_1973_cast_fp16"], epsilon: ["var_1974_epsilon_0"])
0.04% N|C  var_1975_cast_fp16 = ios17.mul(x: ["x_275_cast_fp16"], y: ["var_1974_cast_fp16"])
           model_transformer_layers_13_ffn_norm_weight_to_fp16 = const()
0.04% N|C  input_109_cast_fp16 = ios17.mul(y: ["model_transformer_layers_13_ffn_norm_weight_to_fp16"], x: ["var_1975_cast_fp16"])
           model_transformer_layers_13_ffn_proj_1_weight_to_fp16 = const()
           linear_54_bias_0_to_fp16 = const()
0.15% N|C  linear_54_cast_fp16 = ios17.linear(bias: ["linear_54_bias_0_to_fp16"], weight: ["model_transformer_layers_13_ffn_proj_1_weight_to_fp16"], x: ["input_109_cast_fp16"])
           var_1985_split_sizes_0 = const()
           var_1985_axis_0 = const()
0.23% N|C  var_1985_cast_fp16_0, var_1985_cast_fp16_1 = split(axis: ["var_1985_axis_0"], split_sizes: ["var_1985_split_sizes_0"], x: ["linear_54_cast_fp16"])
0.15% N|C  var_1987_cast_fp16 = ios16.silu(x: ["var_1985_cast_fp16_0"])
0.27% N|C  input_113_cast_fp16 = ios17.mul(x: ["var_1987_cast_fp16"], y: ["var_1985_cast_fp16_1"])
           model_transformer_layers_13_ffn_proj_2_weight_to_fp16 = const()
4.87% N|C  linear_55_cast_fp16 = ios17.linear(bias: ["linear_1_bias_0_to_fp16"], weight: ["model_transformer_layers_13_ffn_proj_2_weight_to_fp16"], x: ["input_113_cast_fp16"])
0.09% N|C  x_279_cast_fp16 = ios17.add(y: ["linear_55_cast_fp16"], x: ["x_275_cast_fp16"])
           var_10_promoted_to_fp16_56 = const()
0.04% N|C  var_1998_cast_fp16 = ios17.pow(x: ["x_279_cast_fp16"], y: ["var_10_promoted_to_fp16_56"])
           var_1999 = const()
0.02% N|C  var_2000_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], axes: ["var_1999"], x: ["var_1998_cast_fp16"])
           var_2001_to_fp16 = const()
0.00% N|C  var_2002_cast_fp16 = ios17.add(y: ["var_2001_to_fp16"], x: ["var_2000_cast_fp16"])
           var_2003_epsilon_0 = const()
0.00% N|C  var_2003_cast_fp16 = ios17.rsqrt(x: ["var_2002_cast_fp16"], epsilon: ["var_2003_epsilon_0"])
0.04% N|C  var_2004_cast_fp16 = ios17.mul(y: ["var_2003_cast_fp16"], x: ["x_279_cast_fp16"])
           model_transformer_layers_14_attn_norm_weight_to_fp16 = const()
0.04% N|C  hidden_states_85_cast_fp16 = ios17.mul(y: ["model_transformer_layers_14_attn_norm_weight_to_fp16"], x: ["var_2004_cast_fp16"])
           model_transformer_layers_14_attn_qkv_proj_weight_to_fp16 = const()
0.38% N|C  linear_56_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_14_attn_qkv_proj_weight_to_fp16"], bias: ["linear_48_bias_0_to_fp16"], x: ["hidden_states_85_cast_fp16"])
           var_2018 = const()
0.06% N|C  qkv_59_cast_fp16 = ios17.reshape(x: ["linear_56_cast_fp16"], shape: ["var_2018"])
           var_2020_perm_0 = const()
           var_2021 = const()
           var_2022_axis_0 = const()
0.13% N|C  transpose_39 = ios17.transpose(perm: ["var_2020_perm_0"], x: ["qkv_59_cast_fp16"])
0.05% N|C  var_2022_cast_fp16_0, var_2022_cast_fp16_1, var_2022_cast_fp16_2 = split(axis: ["var_2022_axis_0"], x: ["transpose_39"], split_sizes: ["var_2021"])
           var_10_promoted_to_fp16_57 = const()
0.04% N|C  var_2028_cast_fp16 = ios17.pow(x: ["var_2022_cast_fp16_0"], y: ["var_10_promoted_to_fp16_57"])
           var_2029 = const()
0.02% N|C  var_2030_cast_fp16 = ios16.reduce_mean(axes: ["var_2029"], keep_dims: ["var_22"], x: ["var_2028_cast_fp16"])
           var_2031_to_fp16 = const()
0.00% N|C  var_2032_cast_fp16 = ios17.add(x: ["var_2030_cast_fp16"], y: ["var_2031_to_fp16"])
           var_2033_epsilon_0 = const()
0.00% N|C  var_2033_cast_fp16 = ios17.rsqrt(x: ["var_2032_cast_fp16"], epsilon: ["var_2033_epsilon_0"])
0.04% N|C  var_2034_cast_fp16 = ios17.mul(y: ["var_2033_cast_fp16"], x: ["var_2022_cast_fp16_0"])
           model_transformer_layers_14_attn_q_norm_weight_to_fp16 = const()
0.04% N|C  query_29_cast_fp16 = ios17.mul(x: ["var_2034_cast_fp16"], y: ["model_transformer_layers_14_attn_q_norm_weight_to_fp16"])
           var_10_promoted_to_fp16_58 = const()
0.01% N|C  var_2039_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_58"], x: ["var_2022_cast_fp16_1"])
           var_2040 = const()
0.01% N|C  var_2041_cast_fp16 = ios16.reduce_mean(x: ["var_2039_cast_fp16"], axes: ["var_2040"], keep_dims: ["var_22"])
           var_2042_to_fp16 = const()
0.00% N|C  var_2043_cast_fp16 = ios17.add(x: ["var_2041_cast_fp16"], y: ["var_2042_to_fp16"])
           var_2044_epsilon_0 = const()
0.00% N|C  var_2044_cast_fp16 = ios17.rsqrt(x: ["var_2043_cast_fp16"], epsilon: ["var_2044_epsilon_0"])
0.01% N|C  var_2045_cast_fp16 = ios17.mul(y: ["var_2044_cast_fp16"], x: ["var_2022_cast_fp16_1"])
           model_transformer_layers_14_attn_k_norm_weight_to_fp16 = const()
0.01% N|C  key_29_cast_fp16 = ios17.mul(y: ["model_transformer_layers_14_attn_k_norm_weight_to_fp16"], x: ["var_2045_cast_fp16"])
0.04% N|C  var_2062_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["query_29_cast_fp16"])
           var_2063_split_sizes_0 = const()
           var_2063_axis_0 = const()
0.03% N|C  var_2063_cast_fp16_0, var_2063_cast_fp16_1 = split(split_sizes: ["var_2063_split_sizes_0"], axis: ["var_2063_axis_0"], x: ["query_29_cast_fp16"])
           const_104_promoted_to_fp16 = const()
0.02% N|C  var_2065_cast_fp16 = ios17.mul(y: ["const_104_promoted_to_fp16"], x: ["var_2063_cast_fp16_1"])
           var_2067_interleave_0 = const()
0.02% N|C  var_2067_cast_fp16 = ios17.concat(values: ["var_2065_cast_fp16", "var_2063_cast_fp16_0"], axis: ["var_5"], interleave: ["var_2067_interleave_0"])
0.04% N|C  var_2068_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_2067_cast_fp16"])
0.09% N|C  query_float_29_cast_fp16 = ios17.add(x: ["var_2062_cast_fp16"], y: ["var_2068_cast_fp16"])
0.01% N|C  var_2074_cast_fp16 = ios17.mul(x: ["key_29_cast_fp16"], y: ["var_156_to_fp16"])
           var_2075_split_sizes_0 = const()
           var_2075_axis_0 = const()
0.01% N|C  var_2075_cast_fp16_0, var_2075_cast_fp16_1 = split(split_sizes: ["var_2075_split_sizes_0"], x: ["key_29_cast_fp16"], axis: ["var_2075_axis_0"])
           const_105_promoted_to_fp16 = const()
0.00% N|C  var_2077_cast_fp16 = ios17.mul(x: ["var_2075_cast_fp16_1"], y: ["const_105_promoted_to_fp16"])
           var_2079_interleave_0 = const()
0.00% N|C  var_2079_cast_fp16 = ios17.concat(values: ["var_2077_cast_fp16", "var_2075_cast_fp16_0"], interleave: ["var_2079_interleave_0"], axis: ["var_5"])
0.01% N|C  var_2080_cast_fp16 = ios17.mul(x: ["var_2079_cast_fp16"], y: ["var_152_to_fp16"])
0.02% N|C  key_float_29_cast_fp16 = ios17.add(x: ["var_2074_cast_fp16"], y: ["var_2080_cast_fp16"])
           reshape_56_shape_0 = const()
0.01% N|C  reshape_56_cast_fp16 = ios17.reshape(x: ["key_float_29_cast_fp16"], shape: ["reshape_56_shape_0"])
           tile_28_reps_0 = const()
0.03% N|C  tile_28_cast_fp16 = tile(reps: ["tile_28_reps_0"], x: ["reshape_56_cast_fp16"])
           reshape_57_shape_0 = const()
0.04% N|C  reshape_57_cast_fp16 = ios17.reshape(shape: ["reshape_57_shape_0"], x: ["tile_28_cast_fp16"])
           transpose_28_perm_0 = const()
           keys_59_shape_0 = const()
0.09% N|C  transpose_38 = ios17.transpose(perm: ["transpose_28_perm_0"], x: ["reshape_57_cast_fp16"])
0.04% N|C  keys_59_cast_fp16 = ios17.reshape(x: ["transpose_38"], shape: ["keys_59_shape_0"])
           reshape_58_shape_0 = const()
0.01% N|C  reshape_58_cast_fp16 = ios17.reshape(x: ["var_2022_cast_fp16_2"], shape: ["reshape_58_shape_0"])
           tile_29_reps_0 = const()
0.03% N|C  tile_29_cast_fp16 = tile(x: ["reshape_58_cast_fp16"], reps: ["tile_29_reps_0"])
           reshape_59_shape_0 = const()
0.04% N|C  reshape_59_cast_fp16 = ios17.reshape(x: ["tile_29_cast_fp16"], shape: ["reshape_59_shape_0"])
           transpose_29_perm_0 = const()
           values_59_shape_0 = const()
0.09% N|C  transpose_37 = ios17.transpose(x: ["reshape_59_cast_fp16"], perm: ["transpose_29_perm_0"])
0.04% N|C  values_59_cast_fp16 = ios17.reshape(shape: ["values_59_shape_0"], x: ["transpose_37"])
           mul_14_y_0_to_fp16 = const()
0.04% N|C  mul_14_cast_fp16 = ios17.mul(y: ["mul_14_y_0_to_fp16"], x: ["query_float_29_cast_fp16"])
           matmul_14_transpose_y_0 = const()
           matmul_14_transpose_x_0 = const()
0.09% N|C  matmul_14_cast_fp16 = ios17.matmul(x: ["mul_14_cast_fp16"], transpose_y: ["matmul_14_transpose_y_0"], y: ["keys_59_cast_fp16"], transpose_x: ["matmul_14_transpose_x_0"])
0.10% N|C  add_29_cast_fp16 = ios17.add(x: ["matmul_14_cast_fp16"], y: ["causal_mask_3_to_fp16"])
           softmax_14_axis_0 = const()
0.09% N|C  softmax_14_cast_fp16 = ios16.softmax(axis: ["softmax_14_axis_0"], x: ["add_29_cast_fp16"])
           attn_output_57_transpose_x_0 = const()
           attn_output_57_transpose_y_0 = const()
0.09% N|C  attn_output_57_cast_fp16 = ios17.matmul(x: ["softmax_14_cast_fp16"], transpose_y: ["attn_output_57_transpose_y_0"], y: ["values_59_cast_fp16"], transpose_x: ["attn_output_57_transpose_x_0"])
           var_2096_perm_0 = const()
           var_2098 = const()
0.09% N|C  transpose_36 = ios17.transpose(x: ["attn_output_57_cast_fp16"], perm: ["var_2096_perm_0"])
0.04% N|C  input_115_cast_fp16 = ios17.reshape(shape: ["var_2098"], x: ["transpose_36"])
           model_transformer_layers_14_attn_out_proj_weight_to_fp16 = const()
0.38% N|C  linear_57_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_14_attn_out_proj_weight_to_fp16"], x: ["input_115_cast_fp16"], bias: ["linear_1_bias_0_to_fp16"])
0.09% N|C  x_295_cast_fp16 = ios17.add(x: ["x_279_cast_fp16"], y: ["linear_57_cast_fp16"])
           var_10_promoted_to_fp16_59 = const()
0.04% N|C  var_2105_cast_fp16 = ios17.pow(x: ["x_295_cast_fp16"], y: ["var_10_promoted_to_fp16_59"])
           var_2106 = const()
0.02% N|C  var_2107_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_2105_cast_fp16"], axes: ["var_2106"])
           var_2108_to_fp16 = const()
0.00% N|C  var_2109_cast_fp16 = ios17.add(y: ["var_2108_to_fp16"], x: ["var_2107_cast_fp16"])
           var_2110_epsilon_0 = const()
0.00% N|C  var_2110_cast_fp16 = ios17.rsqrt(epsilon: ["var_2110_epsilon_0"], x: ["var_2109_cast_fp16"])
0.04% N|C  var_2111_cast_fp16 = ios17.mul(x: ["x_295_cast_fp16"], y: ["var_2110_cast_fp16"])
           model_transformer_layers_14_ffn_norm_weight_to_fp16 = const()
0.04% N|C  input_117_cast_fp16 = ios17.mul(y: ["model_transformer_layers_14_ffn_norm_weight_to_fp16"], x: ["var_2111_cast_fp16"])
           model_transformer_layers_14_ffn_proj_1_weight_to_fp16 = const()
           linear_58_bias_0_to_fp16 = const()
0.15% N|C  linear_58_cast_fp16 = ios17.linear(x: ["input_117_cast_fp16"], weight: ["model_transformer_layers_14_ffn_proj_1_weight_to_fp16"], bias: ["linear_58_bias_0_to_fp16"])
           var_2121_split_sizes_0 = const()
           var_2121_axis_0 = const()
0.24% N|C  var_2121_cast_fp16_0, var_2121_cast_fp16_1 = split(x: ["linear_58_cast_fp16"], axis: ["var_2121_axis_0"], split_sizes: ["var_2121_split_sizes_0"])
0.16% N|C  var_2123_cast_fp16 = ios16.silu(x: ["var_2121_cast_fp16_0"])
0.29% N|C  input_121_cast_fp16 = ios17.mul(x: ["var_2123_cast_fp16"], y: ["var_2121_cast_fp16_1"])
           model_transformer_layers_14_ffn_proj_2_weight_to_fp16 = const()
5.43% N|C  linear_59_cast_fp16 = ios17.linear(x: ["input_121_cast_fp16"], bias: ["linear_1_bias_0_to_fp16"], weight: ["model_transformer_layers_14_ffn_proj_2_weight_to_fp16"])
0.09% N|C  x_299_cast_fp16 = ios17.add(y: ["linear_59_cast_fp16"], x: ["x_295_cast_fp16"])
           var_10_promoted_to_fp16_60 = const()
0.04% N|C  var_2134_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_60"], x: ["x_299_cast_fp16"])
           var_2135 = const()
0.02% N|C  var_2136_cast_fp16 = ios16.reduce_mean(axes: ["var_2135"], x: ["var_2134_cast_fp16"], keep_dims: ["var_22"])
           var_2137_to_fp16 = const()
0.00% N|C  var_2138_cast_fp16 = ios17.add(x: ["var_2136_cast_fp16"], y: ["var_2137_to_fp16"])
           var_2139_epsilon_0 = const()
0.00% N|C  var_2139_cast_fp16 = ios17.rsqrt(epsilon: ["var_2139_epsilon_0"], x: ["var_2138_cast_fp16"])
0.04% N|C  var_2140_cast_fp16 = ios17.mul(y: ["var_2139_cast_fp16"], x: ["x_299_cast_fp16"])
           model_transformer_layers_15_attn_norm_weight_to_fp16 = const()
0.04% N|C  hidden_states_91_cast_fp16 = ios17.mul(y: ["model_transformer_layers_15_attn_norm_weight_to_fp16"], x: ["var_2140_cast_fp16"])
           model_transformer_layers_15_attn_qkv_proj_weight_to_fp16 = const()
0.38% N|C  linear_60_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_15_attn_qkv_proj_weight_to_fp16"], bias: ["linear_48_bias_0_to_fp16"], x: ["hidden_states_91_cast_fp16"])
           var_2154 = const()
0.06% N|C  qkv_cast_fp16 = ios17.reshape(shape: ["var_2154"], x: ["linear_60_cast_fp16"])
           var_2156_perm_0 = const()
           var_2157 = const()
           var_2158_axis_0 = const()
0.13% N|C  transpose_35 = ios17.transpose(x: ["qkv_cast_fp16"], perm: ["var_2156_perm_0"])
0.05% N|C  var_2158_cast_fp16_0, var_2158_cast_fp16_1, var_2158_cast_fp16_2 = split(axis: ["var_2158_axis_0"], split_sizes: ["var_2157"], x: ["transpose_35"])
           var_10_promoted_to_fp16_61 = const()
0.04% N|C  var_2164_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_61"], x: ["var_2158_cast_fp16_0"])
           var_2165 = const()
0.02% N|C  var_2166_cast_fp16 = ios16.reduce_mean(x: ["var_2164_cast_fp16"], keep_dims: ["var_22"], axes: ["var_2165"])
           var_2167_to_fp16 = const()
0.00% N|C  var_2168_cast_fp16 = ios17.add(x: ["var_2166_cast_fp16"], y: ["var_2167_to_fp16"])
           var_2169_epsilon_0 = const()
0.00% N|C  var_2169_cast_fp16 = ios17.rsqrt(x: ["var_2168_cast_fp16"], epsilon: ["var_2169_epsilon_0"])
0.04% N|C  var_2170_cast_fp16 = ios17.mul(x: ["var_2158_cast_fp16_0"], y: ["var_2169_cast_fp16"])
           model_transformer_layers_15_attn_q_norm_weight_to_fp16 = const()
0.04% N|C  query_cast_fp16 = ios17.mul(x: ["var_2170_cast_fp16"], y: ["model_transformer_layers_15_attn_q_norm_weight_to_fp16"])
           var_10_promoted_to_fp16_62 = const()
0.01% N|C  var_2175_cast_fp16 = ios17.pow(x: ["var_2158_cast_fp16_1"], y: ["var_10_promoted_to_fp16_62"])
           var_2176 = const()
0.01% N|C  var_2177_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_2175_cast_fp16"], axes: ["var_2176"])
           var_2178_to_fp16 = const()
0.00% N|C  var_2179_cast_fp16 = ios17.add(y: ["var_2178_to_fp16"], x: ["var_2177_cast_fp16"])
           var_2180_epsilon_0 = const()
0.00% N|C  var_2180_cast_fp16 = ios17.rsqrt(epsilon: ["var_2180_epsilon_0"], x: ["var_2179_cast_fp16"])
0.01% N|C  var_2181_cast_fp16 = ios17.mul(y: ["var_2180_cast_fp16"], x: ["var_2158_cast_fp16_1"])
           model_transformer_layers_15_attn_k_norm_weight_to_fp16 = const()
0.01% N|C  key_cast_fp16 = ios17.mul(x: ["var_2181_cast_fp16"], y: ["model_transformer_layers_15_attn_k_norm_weight_to_fp16"])
0.04% N|C  var_2198_cast_fp16 = ios17.mul(x: ["query_cast_fp16"], y: ["var_156_to_fp16"])
           var_2199_split_sizes_0 = const()
           var_2199_axis_0 = const()
0.03% N|C  var_2199_cast_fp16_0, var_2199_cast_fp16_1 = split(split_sizes: ["var_2199_split_sizes_0"], x: ["query_cast_fp16"], axis: ["var_2199_axis_0"])
           const_111_promoted_to_fp16 = const()
0.02% N|C  var_2201_cast_fp16 = ios17.mul(x: ["var_2199_cast_fp16_1"], y: ["const_111_promoted_to_fp16"])
           var_2203_interleave_0 = const()
0.02% N|C  var_2203_cast_fp16 = ios17.concat(values: ["var_2201_cast_fp16", "var_2199_cast_fp16_0"], interleave: ["var_2203_interleave_0"], axis: ["var_5"])
0.04% N|C  var_2204_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_2203_cast_fp16"])
0.09% N|C  query_float_cast_fp16 = ios17.add(x: ["var_2198_cast_fp16"], y: ["var_2204_cast_fp16"])
0.01% N|C  var_2210_cast_fp16 = ios17.mul(y: ["var_156_to_fp16"], x: ["key_cast_fp16"])
           var_2211_split_sizes_0 = const()
           var_2211_axis_0 = const()
0.01% N|C  var_2211_cast_fp16_0, var_2211_cast_fp16_1 = split(axis: ["var_2211_axis_0"], x: ["key_cast_fp16"], split_sizes: ["var_2211_split_sizes_0"])
           const_112_promoted_to_fp16 = const()
0.00% N|C  var_2213_cast_fp16 = ios17.mul(x: ["var_2211_cast_fp16_1"], y: ["const_112_promoted_to_fp16"])
           var_2215_interleave_0 = const()
0.00% N|C  var_2215_cast_fp16 = ios17.concat(interleave: ["var_2215_interleave_0"], values: ["var_2213_cast_fp16", "var_2211_cast_fp16_0"], axis: ["var_5"])
0.01% N|C  var_2216_cast_fp16 = ios17.mul(y: ["var_152_to_fp16"], x: ["var_2215_cast_fp16"])
0.02% N|C  key_float_cast_fp16 = ios17.add(y: ["var_2216_cast_fp16"], x: ["var_2210_cast_fp16"])
           reshape_60_shape_0 = const()
0.01% N|C  reshape_60_cast_fp16 = ios17.reshape(shape: ["reshape_60_shape_0"], x: ["key_float_cast_fp16"])
           tile_30_reps_0 = const()
0.03% N|C  tile_30_cast_fp16 = tile(x: ["reshape_60_cast_fp16"], reps: ["tile_30_reps_0"])
           reshape_61_shape_0 = const()
0.04% N|C  reshape_61_cast_fp16 = ios17.reshape(shape: ["reshape_61_shape_0"], x: ["tile_30_cast_fp16"])
           transpose_30_perm_0 = const()
           keys_shape_0 = const()
0.09% N|C  transpose_34 = ios17.transpose(x: ["reshape_61_cast_fp16"], perm: ["transpose_30_perm_0"])
0.04% N|C  keys_cast_fp16 = ios17.reshape(x: ["transpose_34"], shape: ["keys_shape_0"])
           reshape_62_shape_0 = const()
0.01% N|C  reshape_62_cast_fp16 = ios17.reshape(shape: ["reshape_62_shape_0"], x: ["var_2158_cast_fp16_2"])
           tile_31_reps_0 = const()
0.03% N|C  tile_31_cast_fp16 = tile(x: ["reshape_62_cast_fp16"], reps: ["tile_31_reps_0"])
           reshape_63_shape_0 = const()
0.04% N|C  reshape_63_cast_fp16 = ios17.reshape(x: ["tile_31_cast_fp16"], shape: ["reshape_63_shape_0"])
           transpose_31_perm_0 = const()
           values_shape_0 = const()
0.09% N|C  transpose_33 = ios17.transpose(x: ["reshape_63_cast_fp16"], perm: ["transpose_31_perm_0"])
0.04% N|C  values_cast_fp16 = ios17.reshape(x: ["transpose_33"], shape: ["values_shape_0"])
           mul_15_y_0_to_fp16 = const()
0.04% N|C  mul_15_cast_fp16 = ios17.mul(x: ["query_float_cast_fp16"], y: ["mul_15_y_0_to_fp16"])
           matmul_15_transpose_y_0 = const()
           matmul_15_transpose_x_0 = const()
0.09% N|C  matmul_15_cast_fp16 = ios17.matmul(transpose_x: ["matmul_15_transpose_x_0"], x: ["mul_15_cast_fp16"], y: ["keys_cast_fp16"], transpose_y: ["matmul_15_transpose_y_0"])
0.10% N|C  add_31_cast_fp16 = ios17.add(x: ["matmul_15_cast_fp16"], y: ["causal_mask_3_to_fp16"])
           softmax_15_axis_0 = const()
0.09% N|C  softmax_15_cast_fp16 = ios16.softmax(axis: ["softmax_15_axis_0"], x: ["add_31_cast_fp16"])
           attn_output_61_transpose_x_0 = const()
           attn_output_61_transpose_y_0 = const()
0.09% N|C  attn_output_61_cast_fp16 = ios17.matmul(transpose_y: ["attn_output_61_transpose_y_0"], transpose_x: ["attn_output_61_transpose_x_0"], y: ["values_cast_fp16"], x: ["softmax_15_cast_fp16"])
           var_2232_perm_0 = const()
           var_2234 = const()
0.09% N|C  transpose_32 = ios17.transpose(x: ["attn_output_61_cast_fp16"], perm: ["var_2232_perm_0"])
0.04% N|C  input_123_cast_fp16 = ios17.reshape(shape: ["var_2234"], x: ["transpose_32"])
           model_transformer_layers_15_attn_out_proj_weight_to_fp16 = const()
0.38% N|C  linear_61_cast_fp16 = ios17.linear(bias: ["linear_1_bias_0_to_fp16"], weight: ["model_transformer_layers_15_attn_out_proj_weight_to_fp16"], x: ["input_123_cast_fp16"])
0.09% N|C  x_315_cast_fp16 = ios17.add(x: ["x_299_cast_fp16"], y: ["linear_61_cast_fp16"])
           var_10_promoted_to_fp16_63 = const()
0.04% N|C  var_2241_cast_fp16 = ios17.pow(y: ["var_10_promoted_to_fp16_63"], x: ["x_315_cast_fp16"])
           var_2242 = const()
0.02% N|C  var_2243_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_2241_cast_fp16"], axes: ["var_2242"])
           var_2244_to_fp16 = const()
0.00% N|C  var_2245_cast_fp16 = ios17.add(y: ["var_2244_to_fp16"], x: ["var_2243_cast_fp16"])
           var_2246_epsilon_0 = const()
0.00% N|C  var_2246_cast_fp16 = ios17.rsqrt(epsilon: ["var_2246_epsilon_0"], x: ["var_2245_cast_fp16"])
0.04% N|C  var_2247_cast_fp16 = ios17.mul(y: ["var_2246_cast_fp16"], x: ["x_315_cast_fp16"])
           model_transformer_layers_15_ffn_norm_weight_to_fp16 = const()
0.04% N|C  input_125_cast_fp16 = ios17.mul(y: ["model_transformer_layers_15_ffn_norm_weight_to_fp16"], x: ["var_2247_cast_fp16"])
           model_transformer_layers_15_ffn_proj_1_weight_to_fp16 = const()
           linear_62_bias_0_to_fp16 = const()
0.15% N|C  linear_62_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_15_ffn_proj_1_weight_to_fp16"], x: ["input_125_cast_fp16"], bias: ["linear_62_bias_0_to_fp16"])
           var_2257_split_sizes_0 = const()
           var_2257_axis_0 = const()
0.26% N|C  var_2257_cast_fp16_0, var_2257_cast_fp16_1 = split(axis: ["var_2257_axis_0"], split_sizes: ["var_2257_split_sizes_0"], x: ["linear_62_cast_fp16"])
0.17% N|C  var_2259_cast_fp16 = ios16.silu(x: ["var_2257_cast_fp16_0"])
0.30% N|C  input_cast_fp16 = ios17.mul(x: ["var_2259_cast_fp16"], y: ["var_2257_cast_fp16_1"])
           model_transformer_layers_15_ffn_proj_2_weight_to_fp16 = const()
6.01% N|C  linear_63_cast_fp16 = ios17.linear(weight: ["model_transformer_layers_15_ffn_proj_2_weight_to_fp16"], x: ["input_cast_fp16"], bias: ["linear_1_bias_0_to_fp16"])
0.09% N|C  x_319_cast_fp16 = ios17.add(y: ["linear_63_cast_fp16"], x: ["x_315_cast_fp16"])
           var_10_promoted_to_fp16_64 = const()
0.04% N|C  var_2266_cast_fp16 = ios17.pow(x: ["x_319_cast_fp16"], y: ["var_10_promoted_to_fp16_64"])
           var_2267 = const()
0.02% N|C  var_2268_cast_fp16 = ios16.reduce_mean(keep_dims: ["var_22"], x: ["var_2266_cast_fp16"], axes: ["var_2267"])
           var_2269_to_fp16 = const()
0.00% N|C  var_2270_cast_fp16 = ios17.add(x: ["var_2268_cast_fp16"], y: ["var_2269_to_fp16"])
           var_2271_epsilon_0 = const()
0.00% N|C  var_2271_cast_fp16 = ios17.rsqrt(x: ["var_2270_cast_fp16"], epsilon: ["var_2271_epsilon_0"])
0.04% N|C  var_2272_cast_fp16 = ios17.mul(y: ["var_2271_cast_fp16"], x: ["x_319_cast_fp16"])
           model_transformer_norm_weight_to_fp16 = const()
0.04% N|C  hidden_states_cast_fp16 = ios17.mul(x: ["var_2272_cast_fp16"], y: ["model_transformer_norm_weight_to_fp16"])
           linear_64_bias_0_to_fp16 = const()
4.26% C|   linear_64_cast_fp16 = ios17.linear(x: ["hidden_states_cast_fp16"], bias: ["linear_64_bias_0_to_fp16"], weight: ["model_transformer_token_embeddings_weight_to_fp16"])
           var_2277_cast_fp16_to_fp32_dtype_0 = const()
1.92% C|   logits = ios17.cast(x: ["linear_64_cast_fp16"], dtype: ["var_2277_cast_fp16_to_fp32_dtype_0"])
           -> (logits)
}