diff --git "a/openai_whisper-tiny.en/TextDecoder.mlmodelc/model.mil" "b/openai_whisper-tiny.en/TextDecoder.mlmodelc/model.mil"
--- "a/openai_whisper-tiny.en/TextDecoder.mlmodelc/model.mil"
+++ "b/openai_whisper-tiny.en/TextDecoder.mlmodelc/model.mil"
@@ -12,8 +12,8 @@ program(1.0)
             tensor<bool, []> var_28_validate_indices_0 = const()[name = tensor<string, []>("op_28_validate_indices_0"), val = tensor<bool, []>(false)];
             tensor<fp16, [448, 384]> embed_positions_weight_to_fp16 = const()[name = tensor<string, []>("embed_positions_weight_to_fp16"), val = tensor<fp16, [448, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39831680)))];
             tensor<string, []> cache_length_to_int16_dtype_0 = const()[name = tensor<string, []>("cache_length_to_int16_dtype_0"), val = tensor<string, []>("int16")];
-            tensor<int16, [1]> cast_68 = cast(dtype = cache_length_to_int16_dtype_0, x = cache_length)[name = tensor<string, []>("cast_68")];
-            tensor<fp16, [1, 384]> var_28_cast_fp16_cast_int16 = gather(axis = var_28_axis_0, batch_dims = var_28_batch_dims_0, indices = cast_68, validate_indices = var_28_validate_indices_0, x = embed_positions_weight_to_fp16)[name = tensor<string, []>("op_28_cast_fp16_cast_int16")];
+            tensor<int16, [1]> cast_69 = cast(dtype = cache_length_to_int16_dtype_0, x = cache_length)[name = tensor<string, []>("cast_69")];
+            tensor<fp16, [1, 384]> var_28_cast_fp16_cast_int16 = gather(axis = var_28_axis_0, batch_dims = var_28_batch_dims_0, indices = cast_69, validate_indices = var_28_validate_indices_0, x = embed_positions_weight_to_fp16)[name = tensor<string, []>("op_28_cast_fp16_cast_int16")];
             tensor<fp16, [1, 384]> hidden_states_1_cast_fp16 = add(x = var_24_cast_fp16, y = var_28_cast_fp16_cast_int16)[name = tensor<string, []>("hidden_states_1_cast_fp16")];
             tensor<int32, [1]> var_42_axes_0 = const()[name = tensor<string, []>("op_42_axes_0"), val = tensor<int32, [1]>([2])];
             tensor<fp16, [1, 384, 1]> var_42_cast_fp16 = expand_dims(axes = var_42_axes_0, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("op_42_cast_fp16")];
@@ -151,12 +151,12 @@ program(1.0)
             tensor<bool, []> mh_w_5_transpose_x_0 = const()[name = tensor<string, []>("mh_w_5_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_5_transpose_y_0 = const()[name = tensor<string, []>("mh_w_5_transpose_y_0"), val = tensor<bool, []>(false)];
             tensor<fp16, [1, 6, 1, 1500]> mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_215_cast_fp16, y = var_217_cast_fp16)[name = tensor<string, []>("mh_w_5_cast_fp16")];
-            tensor<fp16, [1, 6, 1, 1500]> var_220_cast_fp16 = softmax(axis = var_64, x = mh_w_5_cast_fp16)[name = tensor<string, []>("op_220_cast_fp16")];
+            tensor<fp16, [1, 6, 1, 1500]> obj_13_cast_fp16 = softmax(axis = var_64, x = mh_w_5_cast_fp16)[name = tensor<string, []>("obj_13_cast_fp16")];
             tensor<int32, [4]> var_221 = const()[name = tensor<string, []>("op_221"), val = tensor<int32, [4]>([1, 6, 64, -1])];
             tensor<fp16, [1, 6, 64, 1500]> var_222_cast_fp16 = reshape(shape = var_221, x = value_3_cast_fp16)[name = tensor<string, []>("op_222_cast_fp16")];
             tensor<bool, []> attn_3_transpose_x_0 = const()[name = tensor<string, []>("attn_3_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_3_transpose_y_0 = const()[name = tensor<string, []>("attn_3_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 6, 64, 1]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_222_cast_fp16, y = var_220_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")];
+            tensor<fp16, [1, 6, 64, 1]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_222_cast_fp16, y = obj_13_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")];
             tensor<int32, [4]> var_225 = const()[name = tensor<string, []>("op_225"), val = tensor<int32, [4]>([1, 384, 1, -1])];
             tensor<fp16, [1, 384, 1, 1]> input_3_cast_fp16 = reshape(shape = var_225, x = attn_3_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
             tensor<int32, [2]> var_229 = const()[name = tensor<string, []>("op_229"), val = tensor<int32, [2]>([1, 1])];
@@ -213,30 +213,30 @@ program(1.0)
             tensor<fp32, []> denom_7_epsilon_0 = const()[name = tensor<string, []>("denom_7_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
             tensor<fp16, [1, 1, 1, 1]> denom_7_cast_fp16 = rsqrt(epsilon = denom_7_epsilon_0, x = var_305_cast_fp16)[name = tensor<string, []>("denom_7_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> out_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = denom_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
-            tensor<fp16, [384]> obj_13_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_13_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44910656)))];
-            tensor<fp16, [384]> obj_13_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_13_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44911488)))];
-            tensor<fp16, []> obj_13_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_13_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 384, 1, 1]> obj_13_cast_fp16 = batch_norm(beta = obj_13_beta_0_to_fp16, epsilon = obj_13_epsilon_0_to_fp16, gamma = obj_13_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor<string, []>("obj_13_cast_fp16")];
+            tensor<fp16, [384]> obj_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_15_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44910656)))];
+            tensor<fp16, [384]> obj_15_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_15_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44911488)))];
+            tensor<fp16, []> obj_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 384, 1, 1]> obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor<string, []>("obj_15_cast_fp16")];
             tensor<int32, [2]> var_320 = const()[name = tensor<string, []>("op_320"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [2]> var_322 = const()[name = tensor<string, []>("op_322"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> query_5_pad_type_0 = const()[name = tensor<string, []>("query_5_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> query_5_pad_0 = const()[name = tensor<string, []>("query_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44912320)))];
             tensor<fp16, [384]> layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45207296)))];
-            tensor<fp16, [1, 384, 1, 1]> query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = var_322, groups = var_285, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = var_320, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = var_322, groups = var_285, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = var_320, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")];
             tensor<int32, [2]> var_326 = const()[name = tensor<string, []>("op_326"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [2]> var_328 = const()[name = tensor<string, []>("op_328"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> current_key_3_pad_type_0 = const()[name = tensor<string, []>("current_key_3_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> current_key_3_pad_0 = const()[name = tensor<string, []>("current_key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45208128)))];
-            tensor<fp16, [1, 384, 1, 1]> current_key_3_cast_fp16 = conv(dilations = var_328, groups = var_285, pad = current_key_3_pad_0, pad_type = current_key_3_pad_type_0, strides = var_326, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor<string, []>("current_key_3_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> current_key_3_cast_fp16 = conv(dilations = var_328, groups = var_285, pad = current_key_3_pad_0, pad_type = current_key_3_pad_type_0, strides = var_326, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("current_key_3_cast_fp16")];
             tensor<int32, [2]> var_333 = const()[name = tensor<string, []>("op_333"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [2]> var_335 = const()[name = tensor<string, []>("op_335"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> current_value_3_pad_type_0 = const()[name = tensor<string, []>("current_value_3_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> current_value_3_pad_0 = const()[name = tensor<string, []>("current_value_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45503104)))];
             tensor<fp16, [384]> layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45798080)))];
-            tensor<fp16, [1, 384, 1, 1]> current_value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = var_335, groups = var_285, pad = current_value_3_pad_0, pad_type = current_value_3_pad_type_0, strides = var_333, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor<string, []>("current_value_3_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> current_value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = var_335, groups = var_285, pad = current_value_3_pad_0, pad_type = current_value_3_pad_type_0, strides = var_333, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("current_value_3_cast_fp16")];
             tensor<fp16, [1, 384, 1, 448]> var_342_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_342_cast_fp16")];
             tensor<fp16, [1, 384, 1, 448]> var_344_cast_fp16 = mul(x = var_47_cast_fp16_1, y = var_129_cast_fp16)[name = tensor<string, []>("op_344_cast_fp16")];
             tensor<fp16, [1, 384, 1, 448]> key_5_cast_fp16 = add(x = var_342_cast_fp16, y = var_344_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")];
@@ -263,12 +263,12 @@ program(1.0)
             tensor<fp16, [1, 384, 1, 1]> input_11_cast_fp16 = reshape(shape = var_369, x = attn_5_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
             tensor<int32, [2]> var_373 = const()[name = tensor<string, []>("op_373"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [2]> var_375 = const()[name = tensor<string, []>("op_375"), val = tensor<int32, [2]>([1, 1])];
-            tensor<string, []> obj_19_pad_type_0 = const()[name = tensor<string, []>("obj_19_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> obj_19_pad_0 = const()[name = tensor<string, []>("obj_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<string, []> obj_21_pad_type_0 = const()[name = tensor<string, []>("obj_21_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_21_pad_0 = const()[name = tensor<string, []>("obj_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45798912)))];
             tensor<fp16, [384]> layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46093888)))];
-            tensor<fp16, [1, 384, 1, 1]> obj_19_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = var_375, groups = var_285, pad = obj_19_pad_0, pad_type = obj_19_pad_type_0, strides = var_373, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("obj_19_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_19_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> obj_21_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = var_375, groups = var_285, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = var_373, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("obj_21_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")];
             tensor<int32, [1]> var_385 = const()[name = tensor<string, []>("op_385"), val = tensor<int32, [1]>([1])];
             tensor<fp16, [1, 1, 1, 1]> channels_mean_9_cast_fp16 = reduce_mean(axes = var_385, keep_dims = var_286, x = inputs_9_cast_fp16)[name = tensor<string, []>("channels_mean_9_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> zero_mean_9_cast_fp16 = sub(x = inputs_9_cast_fp16, y = channels_mean_9_cast_fp16)[name = tensor<string, []>("zero_mean_9_cast_fp16")];
@@ -280,17 +280,17 @@ program(1.0)
             tensor<fp32, []> denom_9_epsilon_0 = const()[name = tensor<string, []>("denom_9_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
             tensor<fp16, [1, 1, 1, 1]> denom_9_cast_fp16 = rsqrt(epsilon = denom_9_epsilon_0, x = var_392_cast_fp16)[name = tensor<string, []>("denom_9_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> out_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = denom_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
-            tensor<fp16, [384]> obj_21_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_21_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46094720)))];
-            tensor<fp16, [384]> obj_21_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_21_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46095552)))];
-            tensor<fp16, []> obj_21_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_21_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 384, 1, 1]> obj_21_cast_fp16 = batch_norm(beta = obj_21_beta_0_to_fp16, epsilon = obj_21_epsilon_0_to_fp16, gamma = obj_21_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor<string, []>("obj_21_cast_fp16")];
+            tensor<fp16, [384]> obj_23_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_23_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46094720)))];
+            tensor<fp16, [384]> obj_23_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_23_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46095552)))];
+            tensor<fp16, []> obj_23_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_23_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 384, 1, 1]> obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor<string, []>("obj_23_cast_fp16")];
             tensor<int32, [2]> var_407 = const()[name = tensor<string, []>("op_407"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [2]> var_409 = const()[name = tensor<string, []>("op_409"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> query_7_pad_type_0 = const()[name = tensor<string, []>("query_7_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> query_7_pad_0 = const()[name = tensor<string, []>("query_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_1_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46096384)))];
             tensor<fp16, [384]> layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46391360)))];
-            tensor<fp16, [1, 384, 1, 1]> query_7_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = var_409, groups = var_285, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = var_407, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_21_cast_fp16)[name = tensor<string, []>("query_7_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> query_7_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = var_409, groups = var_285, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = var_407, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor<string, []>("query_7_cast_fp16")];
             tensor<int32, [2]> var_413 = const()[name = tensor<string, []>("op_413"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [2]> var_415 = const()[name = tensor<string, []>("op_415"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> key_7_pad_type_0 = const()[name = tensor<string, []>("key_7_pad_type_0"), val = tensor<string, []>("custom")];
@@ -313,405 +313,483 @@ program(1.0)
             tensor<bool, []> mh_w_11_transpose_x_0 = const()[name = tensor<string, []>("mh_w_11_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_11_transpose_y_0 = const()[name = tensor<string, []>("mh_w_11_transpose_y_0"), val = tensor<bool, []>(false)];
             tensor<fp16, [1, 6, 1, 1500]> mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_429_cast_fp16, y = var_431_cast_fp16)[name = tensor<string, []>("mh_w_11_cast_fp16")];
-            tensor<fp16, [1, 6, 1, 1500]> var_434_cast_fp16 = softmax(axis = var_278, x = mh_w_11_cast_fp16)[name = tensor<string, []>("op_434_cast_fp16")];
+            tensor<fp16, [1, 6, 1, 1500]> obj_27_cast_fp16 = softmax(axis = var_278, x = mh_w_11_cast_fp16)[name = tensor<string, []>("obj_27_cast_fp16")];
             tensor<int32, [4]> var_435 = const()[name = tensor<string, []>("op_435"), val = tensor<int32, [4]>([1, 6, 64, -1])];
             tensor<fp16, [1, 6, 64, 1500]> var_436_cast_fp16 = reshape(shape = var_435, x = value_7_cast_fp16)[name = tensor<string, []>("op_436_cast_fp16")];
             tensor<bool, []> attn_7_transpose_x_0 = const()[name = tensor<string, []>("attn_7_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_7_transpose_y_0 = const()[name = tensor<string, []>("attn_7_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 6, 64, 1]> attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_436_cast_fp16, y = var_434_cast_fp16)[name = tensor<string, []>("attn_7_cast_fp16")];
+            tensor<fp16, [1, 6, 64, 1]> attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_436_cast_fp16, y = obj_27_cast_fp16)[name = tensor<string, []>("attn_7_cast_fp16")];
             tensor<int32, [4]> var_439 = const()[name = tensor<string, []>("op_439"), val = tensor<int32, [4]>([1, 384, 1, -1])];
             tensor<fp16, [1, 384, 1, 1]> input_13_cast_fp16 = reshape(shape = var_439, x = attn_7_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
             tensor<int32, [2]> var_443 = const()[name = tensor<string, []>("op_443"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [2]> var_445 = const()[name = tensor<string, []>("op_445"), val = tensor<int32, [2]>([1, 1])];
-            tensor<string, []> obj_23_pad_type_0 = const()[name = tensor<string, []>("obj_23_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> obj_23_pad_0 = const()[name = tensor<string, []>("obj_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<string, []> obj_25_pad_type_0 = const()[name = tensor<string, []>("obj_25_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_25_pad_0 = const()[name = tensor<string, []>("obj_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_1_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46982976)))];
             tensor<fp16, [384]> layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47277952)))];
-            tensor<fp16, [1, 384, 1, 1]> obj_23_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = var_445, groups = var_285, pad = obj_23_pad_0, pad_type = obj_23_pad_type_0, strides = var_443, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("obj_23_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_23_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")];
-            tensor<int32, [1]> var_451 = const()[name = tensor<string, []>("op_451"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_11_cast_fp16 = reduce_mean(axes = var_451, keep_dims = var_286, x = inputs_11_cast_fp16)[name = tensor<string, []>("channels_mean_11_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> obj_25_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = var_445, groups = var_285, pad = obj_25_pad_0, pad_type = obj_25_pad_type_0, strides = var_443, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("obj_25_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")];
+            tensor<int32, [1]> var_454 = const()[name = tensor<string, []>("op_454"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> channels_mean_11_cast_fp16 = reduce_mean(axes = var_454, keep_dims = var_286, x = inputs_11_cast_fp16)[name = tensor<string, []>("channels_mean_11_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> zero_mean_11_cast_fp16 = sub(x = inputs_11_cast_fp16, y = channels_mean_11_cast_fp16)[name = tensor<string, []>("zero_mean_11_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = zero_mean_11_cast_fp16)[name = tensor<string, []>("zero_mean_sq_11_cast_fp16")];
-            tensor<int32, [1]> var_455 = const()[name = tensor<string, []>("op_455"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_456_cast_fp16 = reduce_mean(axes = var_455, keep_dims = var_286, x = zero_mean_sq_11_cast_fp16)[name = tensor<string, []>("op_456_cast_fp16")];
-            tensor<fp16, []> var_457_to_fp16 = const()[name = tensor<string, []>("op_457_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_458_cast_fp16 = add(x = var_456_cast_fp16, y = var_457_to_fp16)[name = tensor<string, []>("op_458_cast_fp16")];
+            tensor<int32, [1]> var_458 = const()[name = tensor<string, []>("op_458"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> var_459_cast_fp16 = reduce_mean(axes = var_458, keep_dims = var_286, x = zero_mean_sq_11_cast_fp16)[name = tensor<string, []>("op_459_cast_fp16")];
+            tensor<fp16, []> var_460_to_fp16 = const()[name = tensor<string, []>("op_460_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1]> var_461_cast_fp16 = add(x = var_459_cast_fp16, y = var_460_to_fp16)[name = tensor<string, []>("op_461_cast_fp16")];
             tensor<fp32, []> denom_11_epsilon_0 = const()[name = tensor<string, []>("denom_11_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
-            tensor<fp16, [1, 1, 1, 1]> denom_11_cast_fp16 = rsqrt(epsilon = denom_11_epsilon_0, x = var_458_cast_fp16)[name = tensor<string, []>("denom_11_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> denom_11_cast_fp16 = rsqrt(epsilon = denom_11_epsilon_0, x = var_461_cast_fp16)[name = tensor<string, []>("denom_11_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> out_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = denom_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
             tensor<fp16, [384]> input_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_15_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47278784)))];
             tensor<fp16, [384]> input_15_beta_0_to_fp16 = const()[name = tensor<string, []>("input_15_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47279616)))];
             tensor<fp16, []> input_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
             tensor<fp16, [1, 384, 1, 1]> input_15_cast_fp16 = batch_norm(beta = input_15_beta_0_to_fp16, epsilon = input_15_epsilon_0_to_fp16, gamma = input_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
-            tensor<int32, [2]> var_469 = const()[name = tensor<string, []>("op_469"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_471 = const()[name = tensor<string, []>("op_471"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_472 = const()[name = tensor<string, []>("op_472"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_474 = const()[name = tensor<string, []>("op_474"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> input_17_pad_type_0 = const()[name = tensor<string, []>("input_17_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [1536, 384, 1, 1]> layers_1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_weight_to_fp16"), val = tensor<fp16, [1536, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47280448)))];
             tensor<fp16, [1536]> layers_1_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48460160)))];
-            tensor<fp16, [1, 1536, 1, 1]> input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = var_471, groups = var_285, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = var_469, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
+            tensor<fp16, [1, 1536, 1, 1]> input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = var_474, groups = var_285, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = var_472, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
             tensor<string, []> input_19_mode_0 = const()[name = tensor<string, []>("input_19_mode_0"), val = tensor<string, []>("EXACT")];
             tensor<fp16, [1, 1536, 1, 1]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
-            tensor<int32, [2]> var_477 = const()[name = tensor<string, []>("op_477"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_479 = const()[name = tensor<string, []>("op_479"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_480 = const()[name = tensor<string, []>("op_480"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_482 = const()[name = tensor<string, []>("op_482"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> hidden_states_5_pad_type_0 = const()[name = tensor<string, []>("hidden_states_5_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> hidden_states_5_pad_0 = const()[name = tensor<string, []>("hidden_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 1536, 1, 1]> layers_1_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_weight_to_fp16"), val = tensor<fp16, [384, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48463296)))];
             tensor<fp16, [384]> layers_1_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49643008)))];
-            tensor<fp16, [1, 384, 1, 1]> hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = var_479, groups = var_285, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = var_477, weight = layers_1_fc2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("hidden_states_5_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = var_482, groups = var_285, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = var_480, weight = layers_1_fc2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("hidden_states_5_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")];
-            tensor<int32, []> var_492 = const()[name = tensor<string, []>("op_492"), val = tensor<int32, []>(3)];
-            tensor<int32, []> var_499 = const()[name = tensor<string, []>("op_499"), val = tensor<int32, []>(1)];
-            tensor<bool, []> var_500 = const()[name = tensor<string, []>("op_500"), val = tensor<bool, []>(true)];
-            tensor<int32, [1]> var_512 = const()[name = tensor<string, []>("op_512"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_13_cast_fp16 = reduce_mean(axes = var_512, keep_dims = var_500, x = inputs_13_cast_fp16)[name = tensor<string, []>("channels_mean_13_cast_fp16")];
+            tensor<int32, []> var_496 = const()[name = tensor<string, []>("op_496"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_503 = const()[name = tensor<string, []>("op_503"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_504 = const()[name = tensor<string, []>("op_504"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_516 = const()[name = tensor<string, []>("op_516"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> channels_mean_13_cast_fp16 = reduce_mean(axes = var_516, keep_dims = var_504, x = inputs_13_cast_fp16)[name = tensor<string, []>("channels_mean_13_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> zero_mean_13_cast_fp16 = sub(x = inputs_13_cast_fp16, y = channels_mean_13_cast_fp16)[name = tensor<string, []>("zero_mean_13_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = zero_mean_13_cast_fp16)[name = tensor<string, []>("zero_mean_sq_13_cast_fp16")];
-            tensor<int32, [1]> var_516 = const()[name = tensor<string, []>("op_516"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_517_cast_fp16 = reduce_mean(axes = var_516, keep_dims = var_500, x = zero_mean_sq_13_cast_fp16)[name = tensor<string, []>("op_517_cast_fp16")];
-            tensor<fp16, []> var_518_to_fp16 = const()[name = tensor<string, []>("op_518_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_519_cast_fp16 = add(x = var_517_cast_fp16, y = var_518_to_fp16)[name = tensor<string, []>("op_519_cast_fp16")];
+            tensor<int32, [1]> var_520 = const()[name = tensor<string, []>("op_520"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> var_521_cast_fp16 = reduce_mean(axes = var_520, keep_dims = var_504, x = zero_mean_sq_13_cast_fp16)[name = tensor<string, []>("op_521_cast_fp16")];
+            tensor<fp16, []> var_522_to_fp16 = const()[name = tensor<string, []>("op_522_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1]> var_523_cast_fp16 = add(x = var_521_cast_fp16, y = var_522_to_fp16)[name = tensor<string, []>("op_523_cast_fp16")];
             tensor<fp32, []> denom_13_epsilon_0 = const()[name = tensor<string, []>("denom_13_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
-            tensor<fp16, [1, 1, 1, 1]> denom_13_cast_fp16 = rsqrt(epsilon = denom_13_epsilon_0, x = var_519_cast_fp16)[name = tensor<string, []>("denom_13_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> denom_13_cast_fp16 = rsqrt(epsilon = denom_13_epsilon_0, x = var_523_cast_fp16)[name = tensor<string, []>("denom_13_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> out_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = denom_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
-            tensor<fp16, [384]> obj_25_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_25_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49643840)))];
-            tensor<fp16, [384]> obj_25_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_25_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49644672)))];
-            tensor<fp16, []> obj_25_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_25_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 384, 1, 1]> obj_25_cast_fp16 = batch_norm(beta = obj_25_beta_0_to_fp16, epsilon = obj_25_epsilon_0_to_fp16, gamma = obj_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor<string, []>("obj_25_cast_fp16")];
-            tensor<int32, [2]> var_534 = const()[name = tensor<string, []>("op_534"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_536 = const()[name = tensor<string, []>("op_536"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [384]> obj_29_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_29_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49643840)))];
+            tensor<fp16, [384]> obj_29_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_29_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49644672)))];
+            tensor<fp16, []> obj_29_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_29_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 384, 1, 1]> obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor<string, []>("obj_29_cast_fp16")];
+            tensor<int32, [2]> var_538 = const()[name = tensor<string, []>("op_538"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_540 = const()[name = tensor<string, []>("op_540"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> query_9_pad_type_0 = const()[name = tensor<string, []>("query_9_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> query_9_pad_0 = const()[name = tensor<string, []>("query_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49645504)))];
             tensor<fp16, [384]> layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49940480)))];
-            tensor<fp16, [1, 384, 1, 1]> query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = var_536, groups = var_499, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = var_534, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")];
-            tensor<int32, [2]> var_540 = const()[name = tensor<string, []>("op_540"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_542 = const()[name = tensor<string, []>("op_542"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 384, 1, 1]> query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = var_540, groups = var_503, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = var_538, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")];
+            tensor<int32, [2]> var_544 = const()[name = tensor<string, []>("op_544"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_546 = const()[name = tensor<string, []>("op_546"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> current_key_5_pad_type_0 = const()[name = tensor<string, []>("current_key_5_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> current_key_5_pad_0 = const()[name = tensor<string, []>("current_key_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49941312)))];
-            tensor<fp16, [1, 384, 1, 1]> current_key_5_cast_fp16 = conv(dilations = var_542, groups = var_499, pad = current_key_5_pad_0, pad_type = current_key_5_pad_type_0, strides = var_540, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor<string, []>("current_key_5_cast_fp16")];
-            tensor<int32, [2]> var_547 = const()[name = tensor<string, []>("op_547"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_549 = const()[name = tensor<string, []>("op_549"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 384, 1, 1]> current_key_5_cast_fp16 = conv(dilations = var_546, groups = var_503, pad = current_key_5_pad_0, pad_type = current_key_5_pad_type_0, strides = var_544, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("current_key_5_cast_fp16")];
+            tensor<int32, [2]> var_551 = const()[name = tensor<string, []>("op_551"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_553 = const()[name = tensor<string, []>("op_553"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> current_value_5_pad_type_0 = const()[name = tensor<string, []>("current_value_5_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> current_value_5_pad_0 = const()[name = tensor<string, []>("current_value_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50236288)))];
             tensor<fp16, [384]> layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50531264)))];
-            tensor<fp16, [1, 384, 1, 1]> current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = var_549, groups = var_499, pad = current_value_5_pad_0, pad_type = current_value_5_pad_type_0, strides = var_547, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor<string, []>("current_value_5_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 448]> var_556_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_556_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 448]> var_558_cast_fp16 = mul(x = var_47_cast_fp16_2, y = var_129_cast_fp16)[name = tensor<string, []>("op_558_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 448]> key_9_cast_fp16 = add(x = var_556_cast_fp16, y = var_558_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 448]> var_560_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_560_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 448]> var_562_cast_fp16 = mul(x = var_54_cast_fp16_2, y = var_129_cast_fp16)[name = tensor<string, []>("op_562_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 448]> value_9_cast_fp16 = add(x = var_560_cast_fp16, y = var_562_cast_fp16)[name = tensor<string, []>("value_9_cast_fp16")];
-            tensor<int32, [4]> var_565 = const()[name = tensor<string, []>("op_565"), val = tensor<int32, [4]>([1, 6, 64, -1])];
-            tensor<fp16, [1, 6, 64, 1]> var_566_cast_fp16 = reshape(shape = var_565, x = query_9_cast_fp16)[name = tensor<string, []>("op_566_cast_fp16")];
-            tensor<fp16, []> var_567_to_fp16 = const()[name = tensor<string, []>("op_567_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
-            tensor<fp16, [1, 6, 64, 1]> var_568_cast_fp16 = mul(x = var_566_cast_fp16, y = var_567_to_fp16)[name = tensor<string, []>("op_568_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = var_553, groups = var_503, pad = current_value_5_pad_0, pad_type = current_value_5_pad_type_0, strides = var_551, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("current_value_5_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 448]> var_560_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_560_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 448]> var_562_cast_fp16 = mul(x = var_47_cast_fp16_2, y = var_129_cast_fp16)[name = tensor<string, []>("op_562_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 448]> key_9_cast_fp16 = add(x = var_560_cast_fp16, y = var_562_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 448]> var_564_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_564_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 448]> var_566_cast_fp16 = mul(x = var_54_cast_fp16_2, y = var_129_cast_fp16)[name = tensor<string, []>("op_566_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 448]> value_9_cast_fp16 = add(x = var_564_cast_fp16, y = var_566_cast_fp16)[name = tensor<string, []>("value_9_cast_fp16")];
             tensor<int32, [4]> var_569 = const()[name = tensor<string, []>("op_569"), val = tensor<int32, [4]>([1, 6, 64, -1])];
-            tensor<fp16, [1, 6, 64, 448]> var_570_cast_fp16 = reshape(shape = var_569, x = key_9_cast_fp16)[name = tensor<string, []>("op_570_cast_fp16")];
+            tensor<fp16, [1, 6, 64, 1]> var_570_cast_fp16 = reshape(shape = var_569, x = query_9_cast_fp16)[name = tensor<string, []>("op_570_cast_fp16")];
+            tensor<fp16, []> var_571_to_fp16 = const()[name = tensor<string, []>("op_571_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 6, 64, 1]> var_572_cast_fp16 = mul(x = var_570_cast_fp16, y = var_571_to_fp16)[name = tensor<string, []>("op_572_cast_fp16")];
+            tensor<int32, [4]> var_573 = const()[name = tensor<string, []>("op_573"), val = tensor<int32, [4]>([1, 6, 64, -1])];
+            tensor<fp16, [1, 6, 64, 448]> var_574_cast_fp16 = reshape(shape = var_573, x = key_9_cast_fp16)[name = tensor<string, []>("op_574_cast_fp16")];
             tensor<bool, []> mh_w_13_transpose_x_0 = const()[name = tensor<string, []>("mh_w_13_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_13_transpose_y_0 = const()[name = tensor<string, []>("mh_w_13_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 6, 1, 448]> mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_568_cast_fp16, y = var_570_cast_fp16)[name = tensor<string, []>("mh_w_13_cast_fp16")];
+            tensor<fp16, [1, 6, 1, 448]> mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_572_cast_fp16, y = var_574_cast_fp16)[name = tensor<string, []>("mh_w_13_cast_fp16")];
             tensor<fp16, [1, 6, 1, 448]> mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_147_cast_fp16)[name = tensor<string, []>("mh_w_15_cast_fp16")];
-            tensor<fp16, [1, 6, 1, 448]> var_578_cast_fp16 = softmax(axis = var_492, x = mh_w_15_cast_fp16)[name = tensor<string, []>("op_578_cast_fp16")];
-            tensor<int32, [4]> var_579 = const()[name = tensor<string, []>("op_579"), val = tensor<int32, [4]>([1, 6, 64, -1])];
-            tensor<fp16, [1, 6, 64, 448]> var_580_cast_fp16 = reshape(shape = var_579, x = value_9_cast_fp16)[name = tensor<string, []>("op_580_cast_fp16")];
+            tensor<fp16, [1, 6, 1, 448]> var_582_cast_fp16 = softmax(axis = var_496, x = mh_w_15_cast_fp16)[name = tensor<string, []>("op_582_cast_fp16")];
+            tensor<int32, [4]> var_583 = const()[name = tensor<string, []>("op_583"), val = tensor<int32, [4]>([1, 6, 64, -1])];
+            tensor<fp16, [1, 6, 64, 448]> var_584_cast_fp16 = reshape(shape = var_583, x = value_9_cast_fp16)[name = tensor<string, []>("op_584_cast_fp16")];
             tensor<bool, []> attn_9_transpose_x_0 = const()[name = tensor<string, []>("attn_9_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_9_transpose_y_0 = const()[name = tensor<string, []>("attn_9_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 6, 64, 1]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_580_cast_fp16, y = var_578_cast_fp16)[name = tensor<string, []>("attn_9_cast_fp16")];
-            tensor<int32, [4]> var_583 = const()[name = tensor<string, []>("op_583"), val = tensor<int32, [4]>([1, 384, 1, -1])];
-            tensor<fp16, [1, 384, 1, 1]> input_21_cast_fp16 = reshape(shape = var_583, x = attn_9_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
-            tensor<int32, [2]> var_587 = const()[name = tensor<string, []>("op_587"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_589 = const()[name = tensor<string, []>("op_589"), val = tensor<int32, [2]>([1, 1])];
-            tensor<string, []> obj_31_pad_type_0 = const()[name = tensor<string, []>("obj_31_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> obj_31_pad_0 = const()[name = tensor<string, []>("obj_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1, 6, 64, 1]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_584_cast_fp16, y = var_582_cast_fp16)[name = tensor<string, []>("attn_9_cast_fp16")];
+            tensor<int32, [4]> var_587 = const()[name = tensor<string, []>("op_587"), val = tensor<int32, [4]>([1, 384, 1, -1])];
+            tensor<fp16, [1, 384, 1, 1]> input_21_cast_fp16 = reshape(shape = var_587, x = attn_9_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
+            tensor<int32, [2]> var_591 = const()[name = tensor<string, []>("op_591"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_593 = const()[name = tensor<string, []>("op_593"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_35_pad_type_0 = const()[name = tensor<string, []>("obj_35_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_35_pad_0 = const()[name = tensor<string, []>("obj_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50532096)))];
             tensor<fp16, [384]> layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50827072)))];
-            tensor<fp16, [1, 384, 1, 1]> obj_31_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = var_589, groups = var_499, pad = obj_31_pad_0, pad_type = obj_31_pad_type_0, strides = var_587, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("obj_31_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_31_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")];
-            tensor<int32, [1]> var_599 = const()[name = tensor<string, []>("op_599"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_15_cast_fp16 = reduce_mean(axes = var_599, keep_dims = var_500, x = inputs_15_cast_fp16)[name = tensor<string, []>("channels_mean_15_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> obj_35_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = var_593, groups = var_503, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = var_591, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("obj_35_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")];
+            tensor<int32, [1]> var_603 = const()[name = tensor<string, []>("op_603"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> channels_mean_15_cast_fp16 = reduce_mean(axes = var_603, keep_dims = var_504, x = inputs_15_cast_fp16)[name = tensor<string, []>("channels_mean_15_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> zero_mean_15_cast_fp16 = sub(x = inputs_15_cast_fp16, y = channels_mean_15_cast_fp16)[name = tensor<string, []>("zero_mean_15_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = zero_mean_15_cast_fp16)[name = tensor<string, []>("zero_mean_sq_15_cast_fp16")];
-            tensor<int32, [1]> var_603 = const()[name = tensor<string, []>("op_603"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_604_cast_fp16 = reduce_mean(axes = var_603, keep_dims = var_500, x = zero_mean_sq_15_cast_fp16)[name = tensor<string, []>("op_604_cast_fp16")];
-            tensor<fp16, []> var_605_to_fp16 = const()[name = tensor<string, []>("op_605_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_606_cast_fp16 = add(x = var_604_cast_fp16, y = var_605_to_fp16)[name = tensor<string, []>("op_606_cast_fp16")];
+            tensor<int32, [1]> var_607 = const()[name = tensor<string, []>("op_607"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> var_608_cast_fp16 = reduce_mean(axes = var_607, keep_dims = var_504, x = zero_mean_sq_15_cast_fp16)[name = tensor<string, []>("op_608_cast_fp16")];
+            tensor<fp16, []> var_609_to_fp16 = const()[name = tensor<string, []>("op_609_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1]> var_610_cast_fp16 = add(x = var_608_cast_fp16, y = var_609_to_fp16)[name = tensor<string, []>("op_610_cast_fp16")];
             tensor<fp32, []> denom_15_epsilon_0 = const()[name = tensor<string, []>("denom_15_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
-            tensor<fp16, [1, 1, 1, 1]> denom_15_cast_fp16 = rsqrt(epsilon = denom_15_epsilon_0, x = var_606_cast_fp16)[name = tensor<string, []>("denom_15_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> denom_15_cast_fp16 = rsqrt(epsilon = denom_15_epsilon_0, x = var_610_cast_fp16)[name = tensor<string, []>("denom_15_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> out_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = denom_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
-            tensor<fp16, [384]> obj_33_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_33_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50827904)))];
-            tensor<fp16, [384]> obj_33_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_33_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50828736)))];
-            tensor<fp16, []> obj_33_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_33_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 384, 1, 1]> obj_33_cast_fp16 = batch_norm(beta = obj_33_beta_0_to_fp16, epsilon = obj_33_epsilon_0_to_fp16, gamma = obj_33_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor<string, []>("obj_33_cast_fp16")];
-            tensor<int32, [2]> var_621 = const()[name = tensor<string, []>("op_621"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_623 = const()[name = tensor<string, []>("op_623"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [384]> obj_37_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_37_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50827904)))];
+            tensor<fp16, [384]> obj_37_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_37_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50828736)))];
+            tensor<fp16, []> obj_37_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_37_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 384, 1, 1]> obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor<string, []>("obj_37_cast_fp16")];
+            tensor<int32, [2]> var_625 = const()[name = tensor<string, []>("op_625"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_627 = const()[name = tensor<string, []>("op_627"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> query_11_pad_type_0 = const()[name = tensor<string, []>("query_11_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> query_11_pad_0 = const()[name = tensor<string, []>("query_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_2_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50829568)))];
             tensor<fp16, [384]> layers_2_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51124544)))];
-            tensor<fp16, [1, 384, 1, 1]> query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = var_623, groups = var_499, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = var_621, weight = layers_2_encoder_attn_q_proj_weight_to_fp16, x = obj_33_cast_fp16)[name = tensor<string, []>("query_11_cast_fp16")];
-            tensor<int32, [2]> var_627 = const()[name = tensor<string, []>("op_627"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_629 = const()[name = tensor<string, []>("op_629"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 384, 1, 1]> query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = var_627, groups = var_503, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = var_625, weight = layers_2_encoder_attn_q_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor<string, []>("query_11_cast_fp16")];
+            tensor<int32, [2]> var_631 = const()[name = tensor<string, []>("op_631"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_633 = const()[name = tensor<string, []>("op_633"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> key_11_pad_type_0 = const()[name = tensor<string, []>("key_11_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> key_11_pad_0 = const()[name = tensor<string, []>("key_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_2_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51125376)))];
-            tensor<fp16, [1, 384, 1, 1500]> key_11_cast_fp16 = conv(dilations = var_629, groups = var_499, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = var_627, weight = layers_2_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_11_cast_fp16")];
-            tensor<int32, [2]> var_634 = const()[name = tensor<string, []>("op_634"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_636 = const()[name = tensor<string, []>("op_636"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 384, 1, 1500]> key_11_cast_fp16 = conv(dilations = var_633, groups = var_503, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = var_631, weight = layers_2_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_11_cast_fp16")];
+            tensor<int32, [2]> var_638 = const()[name = tensor<string, []>("op_638"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_640 = const()[name = tensor<string, []>("op_640"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> value_11_pad_type_0 = const()[name = tensor<string, []>("value_11_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> value_11_pad_0 = const()[name = tensor<string, []>("value_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_2_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51420352)))];
             tensor<fp16, [384]> layers_2_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51715328)))];
-            tensor<fp16, [1, 384, 1, 1500]> value_11_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_bias_to_fp16, dilations = var_636, groups = var_499, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = var_634, weight = layers_2_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_11_cast_fp16")];
-            tensor<int32, [4]> var_640 = const()[name = tensor<string, []>("op_640"), val = tensor<int32, [4]>([1, 6, 64, -1])];
-            tensor<fp16, [1, 6, 64, 1]> var_641_cast_fp16 = reshape(shape = var_640, x = query_11_cast_fp16)[name = tensor<string, []>("op_641_cast_fp16")];
-            tensor<fp16, []> var_642_to_fp16 = const()[name = tensor<string, []>("op_642_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
-            tensor<fp16, [1, 6, 64, 1]> var_643_cast_fp16 = mul(x = var_641_cast_fp16, y = var_642_to_fp16)[name = tensor<string, []>("op_643_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1500]> value_11_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_bias_to_fp16, dilations = var_640, groups = var_503, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = var_638, weight = layers_2_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_11_cast_fp16")];
             tensor<int32, [4]> var_644 = const()[name = tensor<string, []>("op_644"), val = tensor<int32, [4]>([1, 6, 64, -1])];
-            tensor<fp16, [1, 6, 64, 1500]> var_645_cast_fp16 = reshape(shape = var_644, x = key_11_cast_fp16)[name = tensor<string, []>("op_645_cast_fp16")];
+            tensor<fp16, [1, 6, 64, 1]> var_645_cast_fp16 = reshape(shape = var_644, x = query_11_cast_fp16)[name = tensor<string, []>("op_645_cast_fp16")];
+            tensor<fp16, []> var_646_to_fp16 = const()[name = tensor<string, []>("op_646_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 6, 64, 1]> var_647_cast_fp16 = mul(x = var_645_cast_fp16, y = var_646_to_fp16)[name = tensor<string, []>("op_647_cast_fp16")];
+            tensor<int32, [4]> var_648 = const()[name = tensor<string, []>("op_648"), val = tensor<int32, [4]>([1, 6, 64, -1])];
+            tensor<fp16, [1, 6, 64, 1500]> var_649_cast_fp16 = reshape(shape = var_648, x = key_11_cast_fp16)[name = tensor<string, []>("op_649_cast_fp16")];
             tensor<bool, []> mh_w_17_transpose_x_0 = const()[name = tensor<string, []>("mh_w_17_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_17_transpose_y_0 = const()[name = tensor<string, []>("mh_w_17_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 6, 1, 1500]> mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_643_cast_fp16, y = var_645_cast_fp16)[name = tensor<string, []>("mh_w_17_cast_fp16")];
-            tensor<fp16, [1, 6, 1, 1500]> var_648_cast_fp16 = softmax(axis = var_492, x = mh_w_17_cast_fp16)[name = tensor<string, []>("op_648_cast_fp16")];
-            tensor<int32, [4]> var_649 = const()[name = tensor<string, []>("op_649"), val = tensor<int32, [4]>([1, 6, 64, -1])];
-            tensor<fp16, [1, 6, 64, 1500]> var_650_cast_fp16 = reshape(shape = var_649, x = value_11_cast_fp16)[name = tensor<string, []>("op_650_cast_fp16")];
+            tensor<fp16, [1, 6, 1, 1500]> mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_647_cast_fp16, y = var_649_cast_fp16)[name = tensor<string, []>("mh_w_17_cast_fp16")];
+            tensor<fp16, [1, 6, 1, 1500]> obj_41_cast_fp16 = softmax(axis = var_496, x = mh_w_17_cast_fp16)[name = tensor<string, []>("obj_41_cast_fp16")];
+            tensor<int32, [4]> var_653 = const()[name = tensor<string, []>("op_653"), val = tensor<int32, [4]>([1, 6, 64, -1])];
+            tensor<fp16, [1, 6, 64, 1500]> var_654_cast_fp16 = reshape(shape = var_653, x = value_11_cast_fp16)[name = tensor<string, []>("op_654_cast_fp16")];
             tensor<bool, []> attn_11_transpose_x_0 = const()[name = tensor<string, []>("attn_11_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_11_transpose_y_0 = const()[name = tensor<string, []>("attn_11_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 6, 64, 1]> attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_650_cast_fp16, y = var_648_cast_fp16)[name = tensor<string, []>("attn_11_cast_fp16")];
-            tensor<int32, [4]> var_653 = const()[name = tensor<string, []>("op_653"), val = tensor<int32, [4]>([1, 384, 1, -1])];
-            tensor<fp16, [1, 384, 1, 1]> input_23_cast_fp16 = reshape(shape = var_653, x = attn_11_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
-            tensor<int32, [2]> var_657 = const()[name = tensor<string, []>("op_657"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_659 = const()[name = tensor<string, []>("op_659"), val = tensor<int32, [2]>([1, 1])];
-            tensor<string, []> obj_35_pad_type_0 = const()[name = tensor<string, []>("obj_35_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> obj_35_pad_0 = const()[name = tensor<string, []>("obj_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1, 6, 64, 1]> attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_654_cast_fp16, y = obj_41_cast_fp16)[name = tensor<string, []>("attn_11_cast_fp16")];
+            tensor<int32, [4]> var_657 = const()[name = tensor<string, []>("op_657"), val = tensor<int32, [4]>([1, 384, 1, -1])];
+            tensor<fp16, [1, 384, 1, 1]> input_23_cast_fp16 = reshape(shape = var_657, x = attn_11_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
+            tensor<int32, [2]> var_661 = const()[name = tensor<string, []>("op_661"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_663 = const()[name = tensor<string, []>("op_663"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_39_pad_type_0 = const()[name = tensor<string, []>("obj_39_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_39_pad_0 = const()[name = tensor<string, []>("obj_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_2_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51716160)))];
             tensor<fp16, [384]> layers_2_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52011136)))];
-            tensor<fp16, [1, 384, 1, 1]> obj_35_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = var_659, groups = var_499, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = var_657, weight = layers_2_encoder_attn_o_proj_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("obj_35_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_35_cast_fp16)[name = tensor<string, []>("inputs_17_cast_fp16")];
-            tensor<int32, [1]> var_665 = const()[name = tensor<string, []>("op_665"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_17_cast_fp16 = reduce_mean(axes = var_665, keep_dims = var_500, x = inputs_17_cast_fp16)[name = tensor<string, []>("channels_mean_17_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> obj_39_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = var_663, groups = var_503, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = var_661, weight = layers_2_encoder_attn_o_proj_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("obj_39_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor<string, []>("inputs_17_cast_fp16")];
+            tensor<int32, [1]> var_672 = const()[name = tensor<string, []>("op_672"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> channels_mean_17_cast_fp16 = reduce_mean(axes = var_672, keep_dims = var_504, x = inputs_17_cast_fp16)[name = tensor<string, []>("channels_mean_17_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> zero_mean_17_cast_fp16 = sub(x = inputs_17_cast_fp16, y = channels_mean_17_cast_fp16)[name = tensor<string, []>("zero_mean_17_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = zero_mean_17_cast_fp16)[name = tensor<string, []>("zero_mean_sq_17_cast_fp16")];
-            tensor<int32, [1]> var_669 = const()[name = tensor<string, []>("op_669"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_670_cast_fp16 = reduce_mean(axes = var_669, keep_dims = var_500, x = zero_mean_sq_17_cast_fp16)[name = tensor<string, []>("op_670_cast_fp16")];
-            tensor<fp16, []> var_671_to_fp16 = const()[name = tensor<string, []>("op_671_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_672_cast_fp16 = add(x = var_670_cast_fp16, y = var_671_to_fp16)[name = tensor<string, []>("op_672_cast_fp16")];
+            tensor<int32, [1]> var_676 = const()[name = tensor<string, []>("op_676"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> var_677_cast_fp16 = reduce_mean(axes = var_676, keep_dims = var_504, x = zero_mean_sq_17_cast_fp16)[name = tensor<string, []>("op_677_cast_fp16")];
+            tensor<fp16, []> var_678_to_fp16 = const()[name = tensor<string, []>("op_678_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1]> var_679_cast_fp16 = add(x = var_677_cast_fp16, y = var_678_to_fp16)[name = tensor<string, []>("op_679_cast_fp16")];
             tensor<fp32, []> denom_17_epsilon_0 = const()[name = tensor<string, []>("denom_17_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
-            tensor<fp16, [1, 1, 1, 1]> denom_17_cast_fp16 = rsqrt(epsilon = denom_17_epsilon_0, x = var_672_cast_fp16)[name = tensor<string, []>("denom_17_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> denom_17_cast_fp16 = rsqrt(epsilon = denom_17_epsilon_0, x = var_679_cast_fp16)[name = tensor<string, []>("denom_17_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> out_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = denom_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")];
             tensor<fp16, [384]> input_25_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_25_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52011968)))];
             tensor<fp16, [384]> input_25_beta_0_to_fp16 = const()[name = tensor<string, []>("input_25_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52012800)))];
             tensor<fp16, []> input_25_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_25_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
             tensor<fp16, [1, 384, 1, 1]> input_25_cast_fp16 = batch_norm(beta = input_25_beta_0_to_fp16, epsilon = input_25_epsilon_0_to_fp16, gamma = input_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
-            tensor<int32, [2]> var_683 = const()[name = tensor<string, []>("op_683"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_685 = const()[name = tensor<string, []>("op_685"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_690 = const()[name = tensor<string, []>("op_690"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_692 = const()[name = tensor<string, []>("op_692"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> input_27_pad_type_0 = const()[name = tensor<string, []>("input_27_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> input_27_pad_0 = const()[name = tensor<string, []>("input_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [1536, 384, 1, 1]> layers_2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_weight_to_fp16"), val = tensor<fp16, [1536, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52013632)))];
             tensor<fp16, [1536]> layers_2_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53193344)))];
-            tensor<fp16, [1, 1536, 1, 1]> input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = var_685, groups = var_499, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = var_683, weight = layers_2_fc1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
+            tensor<fp16, [1, 1536, 1, 1]> input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = var_692, groups = var_503, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = var_690, weight = layers_2_fc1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
             tensor<string, []> input_29_mode_0 = const()[name = tensor<string, []>("input_29_mode_0"), val = tensor<string, []>("EXACT")];
             tensor<fp16, [1, 1536, 1, 1]> input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = input_27_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
-            tensor<int32, [2]> var_691 = const()[name = tensor<string, []>("op_691"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_693 = const()[name = tensor<string, []>("op_693"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_698 = const()[name = tensor<string, []>("op_698"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_700 = const()[name = tensor<string, []>("op_700"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> hidden_states_7_pad_type_0 = const()[name = tensor<string, []>("hidden_states_7_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> hidden_states_7_pad_0 = const()[name = tensor<string, []>("hidden_states_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 1536, 1, 1]> layers_2_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_weight_to_fp16"), val = tensor<fp16, [384, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53196480)))];
             tensor<fp16, [384]> layers_2_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54376192)))];
-            tensor<fp16, [1, 384, 1, 1]> hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = var_693, groups = var_499, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = var_691, weight = layers_2_fc2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("hidden_states_7_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = var_700, groups = var_503, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = var_698, weight = layers_2_fc2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("hidden_states_7_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor<string, []>("inputs_19_cast_fp16")];
-            tensor<int32, []> var_706 = const()[name = tensor<string, []>("op_706"), val = tensor<int32, []>(3)];
-            tensor<int32, []> var_713 = const()[name = tensor<string, []>("op_713"), val = tensor<int32, []>(1)];
-            tensor<bool, []> var_714 = const()[name = tensor<string, []>("op_714"), val = tensor<bool, []>(true)];
-            tensor<int32, [1]> var_726 = const()[name = tensor<string, []>("op_726"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_19_cast_fp16 = reduce_mean(axes = var_726, keep_dims = var_714, x = inputs_19_cast_fp16)[name = tensor<string, []>("channels_mean_19_cast_fp16")];
+            tensor<int32, []> var_714 = const()[name = tensor<string, []>("op_714"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_721 = const()[name = tensor<string, []>("op_721"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_722 = const()[name = tensor<string, []>("op_722"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_734 = const()[name = tensor<string, []>("op_734"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> channels_mean_19_cast_fp16 = reduce_mean(axes = var_734, keep_dims = var_722, x = inputs_19_cast_fp16)[name = tensor<string, []>("channels_mean_19_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> zero_mean_19_cast_fp16 = sub(x = inputs_19_cast_fp16, y = channels_mean_19_cast_fp16)[name = tensor<string, []>("zero_mean_19_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = zero_mean_19_cast_fp16)[name = tensor<string, []>("zero_mean_sq_19_cast_fp16")];
-            tensor<int32, [1]> var_730 = const()[name = tensor<string, []>("op_730"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_731_cast_fp16 = reduce_mean(axes = var_730, keep_dims = var_714, x = zero_mean_sq_19_cast_fp16)[name = tensor<string, []>("op_731_cast_fp16")];
-            tensor<fp16, []> var_732_to_fp16 = const()[name = tensor<string, []>("op_732_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_733_cast_fp16 = add(x = var_731_cast_fp16, y = var_732_to_fp16)[name = tensor<string, []>("op_733_cast_fp16")];
+            tensor<int32, [1]> var_738 = const()[name = tensor<string, []>("op_738"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> var_739_cast_fp16 = reduce_mean(axes = var_738, keep_dims = var_722, x = zero_mean_sq_19_cast_fp16)[name = tensor<string, []>("op_739_cast_fp16")];
+            tensor<fp16, []> var_740_to_fp16 = const()[name = tensor<string, []>("op_740_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1]> var_741_cast_fp16 = add(x = var_739_cast_fp16, y = var_740_to_fp16)[name = tensor<string, []>("op_741_cast_fp16")];
             tensor<fp32, []> denom_19_epsilon_0 = const()[name = tensor<string, []>("denom_19_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
-            tensor<fp16, [1, 1, 1, 1]> denom_19_cast_fp16 = rsqrt(epsilon = denom_19_epsilon_0, x = var_733_cast_fp16)[name = tensor<string, []>("denom_19_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> denom_19_cast_fp16 = rsqrt(epsilon = denom_19_epsilon_0, x = var_741_cast_fp16)[name = tensor<string, []>("denom_19_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> out_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = denom_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")];
-            tensor<fp16, [384]> obj_37_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_37_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54377024)))];
-            tensor<fp16, [384]> obj_37_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_37_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54377856)))];
-            tensor<fp16, []> obj_37_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_37_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 384, 1, 1]> obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor<string, []>("obj_37_cast_fp16")];
-            tensor<int32, [2]> var_748 = const()[name = tensor<string, []>("op_748"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_750 = const()[name = tensor<string, []>("op_750"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [384]> obj_43_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_43_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54377024)))];
+            tensor<fp16, [384]> obj_43_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_43_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54377856)))];
+            tensor<fp16, []> obj_43_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_43_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 384, 1, 1]> obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor<string, []>("obj_43_cast_fp16")];
+            tensor<int32, [2]> var_756 = const()[name = tensor<string, []>("op_756"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_758 = const()[name = tensor<string, []>("op_758"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> query_13_pad_type_0 = const()[name = tensor<string, []>("query_13_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> query_13_pad_0 = const()[name = tensor<string, []>("query_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54378688)))];
             tensor<fp16, [384]> layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54673664)))];
-            tensor<fp16, [1, 384, 1, 1]> query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = var_750, groups = var_713, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = var_748, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")];
-            tensor<int32, [2]> var_754 = const()[name = tensor<string, []>("op_754"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_756 = const()[name = tensor<string, []>("op_756"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 384, 1, 1]> query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = var_758, groups = var_721, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = var_756, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")];
+            tensor<int32, [2]> var_762 = const()[name = tensor<string, []>("op_762"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_764 = const()[name = tensor<string, []>("op_764"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> current_key_pad_type_0 = const()[name = tensor<string, []>("current_key_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> current_key_pad_0 = const()[name = tensor<string, []>("current_key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54674496)))];
-            tensor<fp16, [1, 384, 1, 1]> current_key_cast_fp16 = conv(dilations = var_756, groups = var_713, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = var_754, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor<string, []>("current_key_cast_fp16")];
-            tensor<int32, [2]> var_761 = const()[name = tensor<string, []>("op_761"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_763 = const()[name = tensor<string, []>("op_763"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 384, 1, 1]> current_key_cast_fp16 = conv(dilations = var_764, groups = var_721, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = var_762, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("current_key_cast_fp16")];
+            tensor<int32, [2]> var_769 = const()[name = tensor<string, []>("op_769"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_771 = const()[name = tensor<string, []>("op_771"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> current_value_pad_type_0 = const()[name = tensor<string, []>("current_value_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> current_value_pad_0 = const()[name = tensor<string, []>("current_value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54969472)))];
             tensor<fp16, [384]> layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55264448)))];
-            tensor<fp16, [1, 384, 1, 1]> current_value_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = var_763, groups = var_713, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = var_761, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor<string, []>("current_value_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 448]> var_770_cast_fp16 = mul(x = current_key_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_770_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 448]> var_772_cast_fp16 = mul(x = var_47_cast_fp16_3, y = var_129_cast_fp16)[name = tensor<string, []>("op_772_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 448]> key_13_cast_fp16 = add(x = var_770_cast_fp16, y = var_772_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 448]> var_774_cast_fp16 = mul(x = current_value_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_774_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 448]> var_776_cast_fp16 = mul(x = var_54_cast_fp16_3, y = var_129_cast_fp16)[name = tensor<string, []>("op_776_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 448]> value_13_cast_fp16 = add(x = var_774_cast_fp16, y = var_776_cast_fp16)[name = tensor<string, []>("value_13_cast_fp16")];
-            tensor<int32, [4]> var_779 = const()[name = tensor<string, []>("op_779"), val = tensor<int32, [4]>([1, 6, 64, -1])];
-            tensor<fp16, [1, 6, 64, 1]> var_780_cast_fp16 = reshape(shape = var_779, x = query_13_cast_fp16)[name = tensor<string, []>("op_780_cast_fp16")];
-            tensor<fp16, []> var_781_to_fp16 = const()[name = tensor<string, []>("op_781_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
-            tensor<fp16, [1, 6, 64, 1]> var_782_cast_fp16 = mul(x = var_780_cast_fp16, y = var_781_to_fp16)[name = tensor<string, []>("op_782_cast_fp16")];
-            tensor<int32, [4]> var_783 = const()[name = tensor<string, []>("op_783"), val = tensor<int32, [4]>([1, 6, 64, -1])];
-            tensor<fp16, [1, 6, 64, 448]> var_784_cast_fp16 = reshape(shape = var_783, x = key_13_cast_fp16)[name = tensor<string, []>("op_784_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> current_value_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = var_771, groups = var_721, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = var_769, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("current_value_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 448]> var_778_cast_fp16 = mul(x = current_key_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_778_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 448]> var_780_cast_fp16 = mul(x = var_47_cast_fp16_3, y = var_129_cast_fp16)[name = tensor<string, []>("op_780_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 448]> key_13_cast_fp16 = add(x = var_778_cast_fp16, y = var_780_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 448]> var_782_cast_fp16 = mul(x = current_value_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("op_782_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 448]> var_784_cast_fp16 = mul(x = var_54_cast_fp16_3, y = var_129_cast_fp16)[name = tensor<string, []>("op_784_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 448]> value_13_cast_fp16 = add(x = var_782_cast_fp16, y = var_784_cast_fp16)[name = tensor<string, []>("value_13_cast_fp16")];
+            tensor<int32, [4]> var_787 = const()[name = tensor<string, []>("op_787"), val = tensor<int32, [4]>([1, 6, 64, -1])];
+            tensor<fp16, [1, 6, 64, 1]> var_788_cast_fp16 = reshape(shape = var_787, x = query_13_cast_fp16)[name = tensor<string, []>("op_788_cast_fp16")];
+            tensor<fp16, []> var_789_to_fp16 = const()[name = tensor<string, []>("op_789_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 6, 64, 1]> var_790_cast_fp16 = mul(x = var_788_cast_fp16, y = var_789_to_fp16)[name = tensor<string, []>("op_790_cast_fp16")];
+            tensor<int32, [4]> var_791 = const()[name = tensor<string, []>("op_791"), val = tensor<int32, [4]>([1, 6, 64, -1])];
+            tensor<fp16, [1, 6, 64, 448]> var_792_cast_fp16 = reshape(shape = var_791, x = key_13_cast_fp16)[name = tensor<string, []>("op_792_cast_fp16")];
             tensor<bool, []> mh_w_19_transpose_x_0 = const()[name = tensor<string, []>("mh_w_19_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_19_transpose_y_0 = const()[name = tensor<string, []>("mh_w_19_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 6, 1, 448]> mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_782_cast_fp16, y = var_784_cast_fp16)[name = tensor<string, []>("mh_w_19_cast_fp16")];
+            tensor<fp16, [1, 6, 1, 448]> mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_790_cast_fp16, y = var_792_cast_fp16)[name = tensor<string, []>("mh_w_19_cast_fp16")];
             tensor<fp16, [1, 6, 1, 448]> mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_147_cast_fp16)[name = tensor<string, []>("mh_w_21_cast_fp16")];
-            tensor<fp16, [1, 6, 1, 448]> var_792_cast_fp16 = softmax(axis = var_706, x = mh_w_21_cast_fp16)[name = tensor<string, []>("op_792_cast_fp16")];
-            tensor<int32, [4]> var_793 = const()[name = tensor<string, []>("op_793"), val = tensor<int32, [4]>([1, 6, 64, -1])];
-            tensor<fp16, [1, 6, 64, 448]> var_794_cast_fp16 = reshape(shape = var_793, x = value_13_cast_fp16)[name = tensor<string, []>("op_794_cast_fp16")];
+            tensor<fp16, [1, 6, 1, 448]> var_800_cast_fp16 = softmax(axis = var_714, x = mh_w_21_cast_fp16)[name = tensor<string, []>("op_800_cast_fp16")];
+            tensor<int32, [4]> var_801 = const()[name = tensor<string, []>("op_801"), val = tensor<int32, [4]>([1, 6, 64, -1])];
+            tensor<fp16, [1, 6, 64, 448]> var_802_cast_fp16 = reshape(shape = var_801, x = value_13_cast_fp16)[name = tensor<string, []>("op_802_cast_fp16")];
             tensor<bool, []> attn_13_transpose_x_0 = const()[name = tensor<string, []>("attn_13_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_13_transpose_y_0 = const()[name = tensor<string, []>("attn_13_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 6, 64, 1]> attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_794_cast_fp16, y = var_792_cast_fp16)[name = tensor<string, []>("attn_13_cast_fp16")];
-            tensor<int32, [4]> var_797 = const()[name = tensor<string, []>("op_797"), val = tensor<int32, [4]>([1, 384, 1, -1])];
-            tensor<fp16, [1, 384, 1, 1]> input_31_cast_fp16 = reshape(shape = var_797, x = attn_13_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
-            tensor<int32, [2]> var_801 = const()[name = tensor<string, []>("op_801"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_803 = const()[name = tensor<string, []>("op_803"), val = tensor<int32, [2]>([1, 1])];
-            tensor<string, []> obj_43_pad_type_0 = const()[name = tensor<string, []>("obj_43_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> obj_43_pad_0 = const()[name = tensor<string, []>("obj_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1, 6, 64, 1]> attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_802_cast_fp16, y = var_800_cast_fp16)[name = tensor<string, []>("attn_13_cast_fp16")];
+            tensor<int32, [4]> var_805 = const()[name = tensor<string, []>("op_805"), val = tensor<int32, [4]>([1, 384, 1, -1])];
+            tensor<fp16, [1, 384, 1, 1]> input_31_cast_fp16 = reshape(shape = var_805, x = attn_13_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
+            tensor<int32, [2]> var_809 = const()[name = tensor<string, []>("op_809"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_811 = const()[name = tensor<string, []>("op_811"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_49_pad_type_0 = const()[name = tensor<string, []>("obj_49_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_49_pad_0 = const()[name = tensor<string, []>("obj_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55265280)))];
             tensor<fp16, [384]> layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55560256)))];
-            tensor<fp16, [1, 384, 1, 1]> obj_43_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = var_803, groups = var_713, pad = obj_43_pad_0, pad_type = obj_43_pad_type_0, strides = var_801, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("obj_43_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_43_cast_fp16)[name = tensor<string, []>("inputs_21_cast_fp16")];
-            tensor<int32, [1]> var_813 = const()[name = tensor<string, []>("op_813"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_21_cast_fp16 = reduce_mean(axes = var_813, keep_dims = var_714, x = inputs_21_cast_fp16)[name = tensor<string, []>("channels_mean_21_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> obj_49_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = var_811, groups = var_721, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = var_809, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("obj_49_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor<string, []>("inputs_21_cast_fp16")];
+            tensor<int32, [1]> var_821 = const()[name = tensor<string, []>("op_821"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> channels_mean_21_cast_fp16 = reduce_mean(axes = var_821, keep_dims = var_722, x = inputs_21_cast_fp16)[name = tensor<string, []>("channels_mean_21_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> zero_mean_21_cast_fp16 = sub(x = inputs_21_cast_fp16, y = channels_mean_21_cast_fp16)[name = tensor<string, []>("zero_mean_21_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = zero_mean_21_cast_fp16)[name = tensor<string, []>("zero_mean_sq_21_cast_fp16")];
-            tensor<int32, [1]> var_817 = const()[name = tensor<string, []>("op_817"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_818_cast_fp16 = reduce_mean(axes = var_817, keep_dims = var_714, x = zero_mean_sq_21_cast_fp16)[name = tensor<string, []>("op_818_cast_fp16")];
-            tensor<fp16, []> var_819_to_fp16 = const()[name = tensor<string, []>("op_819_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_820_cast_fp16 = add(x = var_818_cast_fp16, y = var_819_to_fp16)[name = tensor<string, []>("op_820_cast_fp16")];
+            tensor<int32, [1]> var_825 = const()[name = tensor<string, []>("op_825"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> var_826_cast_fp16 = reduce_mean(axes = var_825, keep_dims = var_722, x = zero_mean_sq_21_cast_fp16)[name = tensor<string, []>("op_826_cast_fp16")];
+            tensor<fp16, []> var_827_to_fp16 = const()[name = tensor<string, []>("op_827_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1]> var_828_cast_fp16 = add(x = var_826_cast_fp16, y = var_827_to_fp16)[name = tensor<string, []>("op_828_cast_fp16")];
             tensor<fp32, []> denom_21_epsilon_0 = const()[name = tensor<string, []>("denom_21_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
-            tensor<fp16, [1, 1, 1, 1]> denom_21_cast_fp16 = rsqrt(epsilon = denom_21_epsilon_0, x = var_820_cast_fp16)[name = tensor<string, []>("denom_21_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> denom_21_cast_fp16 = rsqrt(epsilon = denom_21_epsilon_0, x = var_828_cast_fp16)[name = tensor<string, []>("denom_21_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> out_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = denom_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")];
-            tensor<fp16, [384]> obj_45_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_45_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55561088)))];
-            tensor<fp16, [384]> obj_45_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_45_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55561920)))];
-            tensor<fp16, []> obj_45_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_45_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 384, 1, 1]> obj_45_cast_fp16 = batch_norm(beta = obj_45_beta_0_to_fp16, epsilon = obj_45_epsilon_0_to_fp16, gamma = obj_45_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor<string, []>("obj_45_cast_fp16")];
-            tensor<int32, [2]> var_835 = const()[name = tensor<string, []>("op_835"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_837 = const()[name = tensor<string, []>("op_837"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [384]> obj_51_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_51_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55561088)))];
+            tensor<fp16, [384]> obj_51_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_51_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55561920)))];
+            tensor<fp16, []> obj_51_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_51_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 384, 1, 1]> obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor<string, []>("obj_51_cast_fp16")];
+            tensor<int32, [2]> var_843 = const()[name = tensor<string, []>("op_843"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_845 = const()[name = tensor<string, []>("op_845"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> query_pad_type_0 = const()[name = tensor<string, []>("query_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> query_pad_0 = const()[name = tensor<string, []>("query_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_3_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55562752)))];
             tensor<fp16, [384]> layers_3_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55857728)))];
-            tensor<fp16, [1, 384, 1, 1]> query_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = var_837, groups = var_713, pad = query_pad_0, pad_type = query_pad_type_0, strides = var_835, weight = layers_3_encoder_attn_q_proj_weight_to_fp16, x = obj_45_cast_fp16)[name = tensor<string, []>("query_cast_fp16")];
-            tensor<int32, [2]> var_841 = const()[name = tensor<string, []>("op_841"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_843 = const()[name = tensor<string, []>("op_843"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 384, 1, 1]> query_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = var_845, groups = var_721, pad = query_pad_0, pad_type = query_pad_type_0, strides = var_843, weight = layers_3_encoder_attn_q_proj_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor<string, []>("query_cast_fp16")];
+            tensor<int32, [2]> var_849 = const()[name = tensor<string, []>("op_849"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_851 = const()[name = tensor<string, []>("op_851"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> key_pad_type_0 = const()[name = tensor<string, []>("key_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> key_pad_0 = const()[name = tensor<string, []>("key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_3_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55858560)))];
-            tensor<fp16, [1, 384, 1, 1500]> key_cast_fp16 = conv(dilations = var_843, groups = var_713, pad = key_pad_0, pad_type = key_pad_type_0, strides = var_841, weight = layers_3_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_cast_fp16")];
-            tensor<int32, [2]> var_848 = const()[name = tensor<string, []>("op_848"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_850 = const()[name = tensor<string, []>("op_850"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 384, 1, 1500]> key_cast_fp16 = conv(dilations = var_851, groups = var_721, pad = key_pad_0, pad_type = key_pad_type_0, strides = var_849, weight = layers_3_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_cast_fp16")];
+            tensor<int32, [2]> var_856 = const()[name = tensor<string, []>("op_856"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_858 = const()[name = tensor<string, []>("op_858"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> value_pad_type_0 = const()[name = tensor<string, []>("value_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> value_pad_0 = const()[name = tensor<string, []>("value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_3_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56153536)))];
             tensor<fp16, [384]> layers_3_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56448512)))];
-            tensor<fp16, [1, 384, 1, 1500]> value_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_bias_to_fp16, dilations = var_850, groups = var_713, pad = value_pad_0, pad_type = value_pad_type_0, strides = var_848, weight = layers_3_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_cast_fp16")];
-            tensor<int32, [4]> var_854 = const()[name = tensor<string, []>("op_854"), val = tensor<int32, [4]>([1, 6, 64, -1])];
-            tensor<fp16, [1, 6, 64, 1]> var_855_cast_fp16 = reshape(shape = var_854, x = query_cast_fp16)[name = tensor<string, []>("op_855_cast_fp16")];
-            tensor<fp16, []> var_856_to_fp16 = const()[name = tensor<string, []>("op_856_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
-            tensor<fp16, [1, 6, 64, 1]> var_857_cast_fp16 = mul(x = var_855_cast_fp16, y = var_856_to_fp16)[name = tensor<string, []>("op_857_cast_fp16")];
-            tensor<int32, [4]> var_858 = const()[name = tensor<string, []>("op_858"), val = tensor<int32, [4]>([1, 6, 64, -1])];
-            tensor<fp16, [1, 6, 64, 1500]> var_859_cast_fp16 = reshape(shape = var_858, x = key_cast_fp16)[name = tensor<string, []>("op_859_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1500]> value_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_bias_to_fp16, dilations = var_858, groups = var_721, pad = value_pad_0, pad_type = value_pad_type_0, strides = var_856, weight = layers_3_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_cast_fp16")];
+            tensor<int32, [4]> var_862 = const()[name = tensor<string, []>("op_862"), val = tensor<int32, [4]>([1, 6, 64, -1])];
+            tensor<fp16, [1, 6, 64, 1]> var_863_cast_fp16 = reshape(shape = var_862, x = query_cast_fp16)[name = tensor<string, []>("op_863_cast_fp16")];
+            tensor<fp16, []> var_864_to_fp16 = const()[name = tensor<string, []>("op_864_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 6, 64, 1]> var_865_cast_fp16 = mul(x = var_863_cast_fp16, y = var_864_to_fp16)[name = tensor<string, []>("op_865_cast_fp16")];
+            tensor<int32, [4]> var_866 = const()[name = tensor<string, []>("op_866"), val = tensor<int32, [4]>([1, 6, 64, -1])];
+            tensor<fp16, [1, 6, 64, 1500]> var_867_cast_fp16 = reshape(shape = var_866, x = key_cast_fp16)[name = tensor<string, []>("op_867_cast_fp16")];
             tensor<bool, []> mh_w_transpose_x_0 = const()[name = tensor<string, []>("mh_w_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_transpose_y_0 = const()[name = tensor<string, []>("mh_w_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 6, 1, 1500]> mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_857_cast_fp16, y = var_859_cast_fp16)[name = tensor<string, []>("mh_w_cast_fp16")];
-            tensor<fp16, [1, 6, 1, 1500]> var_862_cast_fp16 = softmax(axis = var_706, x = mh_w_cast_fp16)[name = tensor<string, []>("op_862_cast_fp16")];
-            tensor<int32, [4]> var_863 = const()[name = tensor<string, []>("op_863"), val = tensor<int32, [4]>([1, 6, 64, -1])];
-            tensor<fp16, [1, 6, 64, 1500]> var_864_cast_fp16 = reshape(shape = var_863, x = value_cast_fp16)[name = tensor<string, []>("op_864_cast_fp16")];
+            tensor<fp16, [1, 6, 1, 1500]> mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_865_cast_fp16, y = var_867_cast_fp16)[name = tensor<string, []>("mh_w_cast_fp16")];
+            tensor<fp16, [1, 6, 1, 1500]> obj_55_cast_fp16 = softmax(axis = var_714, x = mh_w_cast_fp16)[name = tensor<string, []>("obj_55_cast_fp16")];
+            tensor<int32, [4]> var_871 = const()[name = tensor<string, []>("op_871"), val = tensor<int32, [4]>([1, 6, 64, -1])];
+            tensor<fp16, [1, 6, 64, 1500]> var_872_cast_fp16 = reshape(shape = var_871, x = value_cast_fp16)[name = tensor<string, []>("op_872_cast_fp16")];
             tensor<bool, []> attn_transpose_x_0 = const()[name = tensor<string, []>("attn_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_transpose_y_0 = const()[name = tensor<string, []>("attn_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 6, 64, 1]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_864_cast_fp16, y = var_862_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")];
-            tensor<int32, [4]> var_867 = const()[name = tensor<string, []>("op_867"), val = tensor<int32, [4]>([1, 384, 1, -1])];
-            tensor<fp16, [1, 384, 1, 1]> input_33_cast_fp16 = reshape(shape = var_867, x = attn_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
-            tensor<int32, [2]> var_871 = const()[name = tensor<string, []>("op_871"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_873 = const()[name = tensor<string, []>("op_873"), val = tensor<int32, [2]>([1, 1])];
-            tensor<string, []> obj_47_pad_type_0 = const()[name = tensor<string, []>("obj_47_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> obj_47_pad_0 = const()[name = tensor<string, []>("obj_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1, 6, 64, 1]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_872_cast_fp16, y = obj_55_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")];
+            tensor<int32, [4]> var_875 = const()[name = tensor<string, []>("op_875"), val = tensor<int32, [4]>([1, 384, 1, -1])];
+            tensor<fp16, [1, 384, 1, 1]> input_33_cast_fp16 = reshape(shape = var_875, x = attn_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
+            tensor<int32, [2]> var_879 = const()[name = tensor<string, []>("op_879"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_881 = const()[name = tensor<string, []>("op_881"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_53_pad_type_0 = const()[name = tensor<string, []>("obj_53_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_53_pad_0 = const()[name = tensor<string, []>("obj_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 384, 1, 1]> layers_3_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56449344)))];
             tensor<fp16, [384]> layers_3_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56744320)))];
-            tensor<fp16, [1, 384, 1, 1]> obj_47_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = var_873, groups = var_713, pad = obj_47_pad_0, pad_type = obj_47_pad_type_0, strides = var_871, weight = layers_3_encoder_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("obj_47_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_47_cast_fp16)[name = tensor<string, []>("inputs_23_cast_fp16")];
-            tensor<int32, [1]> var_879 = const()[name = tensor<string, []>("op_879"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_23_cast_fp16 = reduce_mean(axes = var_879, keep_dims = var_714, x = inputs_23_cast_fp16)[name = tensor<string, []>("channels_mean_23_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> obj_53_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = var_881, groups = var_721, pad = obj_53_pad_0, pad_type = obj_53_pad_type_0, strides = var_879, weight = layers_3_encoder_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("obj_53_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor<string, []>("inputs_23_cast_fp16")];
+            tensor<int32, [1]> var_890 = const()[name = tensor<string, []>("op_890"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> channels_mean_23_cast_fp16 = reduce_mean(axes = var_890, keep_dims = var_722, x = inputs_23_cast_fp16)[name = tensor<string, []>("channels_mean_23_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> zero_mean_23_cast_fp16 = sub(x = inputs_23_cast_fp16, y = channels_mean_23_cast_fp16)[name = tensor<string, []>("zero_mean_23_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = zero_mean_23_cast_fp16)[name = tensor<string, []>("zero_mean_sq_23_cast_fp16")];
-            tensor<int32, [1]> var_883 = const()[name = tensor<string, []>("op_883"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_884_cast_fp16 = reduce_mean(axes = var_883, keep_dims = var_714, x = zero_mean_sq_23_cast_fp16)[name = tensor<string, []>("op_884_cast_fp16")];
-            tensor<fp16, []> var_885_to_fp16 = const()[name = tensor<string, []>("op_885_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_886_cast_fp16 = add(x = var_884_cast_fp16, y = var_885_to_fp16)[name = tensor<string, []>("op_886_cast_fp16")];
+            tensor<int32, [1]> var_894 = const()[name = tensor<string, []>("op_894"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> var_895_cast_fp16 = reduce_mean(axes = var_894, keep_dims = var_722, x = zero_mean_sq_23_cast_fp16)[name = tensor<string, []>("op_895_cast_fp16")];
+            tensor<fp16, []> var_896_to_fp16 = const()[name = tensor<string, []>("op_896_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1]> var_897_cast_fp16 = add(x = var_895_cast_fp16, y = var_896_to_fp16)[name = tensor<string, []>("op_897_cast_fp16")];
             tensor<fp32, []> denom_23_epsilon_0 = const()[name = tensor<string, []>("denom_23_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
-            tensor<fp16, [1, 1, 1, 1]> denom_23_cast_fp16 = rsqrt(epsilon = denom_23_epsilon_0, x = var_886_cast_fp16)[name = tensor<string, []>("denom_23_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> denom_23_cast_fp16 = rsqrt(epsilon = denom_23_epsilon_0, x = var_897_cast_fp16)[name = tensor<string, []>("denom_23_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> out_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = denom_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")];
             tensor<fp16, [384]> input_35_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_35_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56745152)))];
             tensor<fp16, [384]> input_35_beta_0_to_fp16 = const()[name = tensor<string, []>("input_35_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56745984)))];
             tensor<fp16, []> input_35_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_35_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
             tensor<fp16, [1, 384, 1, 1]> input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
-            tensor<int32, [2]> var_897 = const()[name = tensor<string, []>("op_897"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_899 = const()[name = tensor<string, []>("op_899"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_908 = const()[name = tensor<string, []>("op_908"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_910 = const()[name = tensor<string, []>("op_910"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> input_37_pad_type_0 = const()[name = tensor<string, []>("input_37_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> input_37_pad_0 = const()[name = tensor<string, []>("input_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [1536, 384, 1, 1]> layers_3_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_weight_to_fp16"), val = tensor<fp16, [1536, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56746816)))];
             tensor<fp16, [1536]> layers_3_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57926528)))];
-            tensor<fp16, [1, 1536, 1, 1]> input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = var_899, groups = var_713, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = var_897, weight = layers_3_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
+            tensor<fp16, [1, 1536, 1, 1]> input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = var_910, groups = var_721, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = var_908, weight = layers_3_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
             tensor<string, []> input_mode_0 = const()[name = tensor<string, []>("input_mode_0"), val = tensor<string, []>("EXACT")];
             tensor<fp16, [1, 1536, 1, 1]> input_cast_fp16 = gelu(mode = input_mode_0, x = input_37_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
-            tensor<int32, [2]> var_905 = const()[name = tensor<string, []>("op_905"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> var_907 = const()[name = tensor<string, []>("op_907"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_916 = const()[name = tensor<string, []>("op_916"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_918 = const()[name = tensor<string, []>("op_918"), val = tensor<int32, [2]>([1, 1])];
             tensor<string, []> hidden_states_9_pad_type_0 = const()[name = tensor<string, []>("hidden_states_9_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> hidden_states_9_pad_0 = const()[name = tensor<string, []>("hidden_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<fp16, [384, 1536, 1, 1]> layers_3_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_weight_to_fp16"), val = tensor<fp16, [384, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57929664)))];
             tensor<fp16, [384]> layers_3_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59109376)))];
-            tensor<fp16, [1, 384, 1, 1]> hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = var_907, groups = var_713, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = var_905, weight = layers_3_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("hidden_states_9_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = var_918, groups = var_721, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = var_916, weight = layers_3_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("hidden_states_9_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> inputs_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")];
-            tensor<bool, []> var_917 = const()[name = tensor<string, []>("op_917"), val = tensor<bool, []>(true)];
-            tensor<int32, [1]> var_921 = const()[name = tensor<string, []>("op_921"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_cast_fp16 = reduce_mean(axes = var_921, keep_dims = var_917, x = inputs_cast_fp16)[name = tensor<string, []>("channels_mean_cast_fp16")];
+            tensor<bool, []> var_929 = const()[name = tensor<string, []>("op_929"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_933 = const()[name = tensor<string, []>("op_933"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> channels_mean_cast_fp16 = reduce_mean(axes = var_933, keep_dims = var_929, x = inputs_cast_fp16)[name = tensor<string, []>("channels_mean_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> zero_mean_cast_fp16 = sub(x = inputs_cast_fp16, y = channels_mean_cast_fp16)[name = tensor<string, []>("zero_mean_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_cast_fp16 = mul(x = zero_mean_cast_fp16, y = zero_mean_cast_fp16)[name = tensor<string, []>("zero_mean_sq_cast_fp16")];
-            tensor<int32, [1]> var_925 = const()[name = tensor<string, []>("op_925"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_926_cast_fp16 = reduce_mean(axes = var_925, keep_dims = var_917, x = zero_mean_sq_cast_fp16)[name = tensor<string, []>("op_926_cast_fp16")];
-            tensor<fp16, []> var_927_to_fp16 = const()[name = tensor<string, []>("op_927_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_928_cast_fp16 = add(x = var_926_cast_fp16, y = var_927_to_fp16)[name = tensor<string, []>("op_928_cast_fp16")];
+            tensor<int32, [1]> var_937 = const()[name = tensor<string, []>("op_937"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> var_938_cast_fp16 = reduce_mean(axes = var_937, keep_dims = var_929, x = zero_mean_sq_cast_fp16)[name = tensor<string, []>("op_938_cast_fp16")];
+            tensor<fp16, []> var_939_to_fp16 = const()[name = tensor<string, []>("op_939_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1]> var_940_cast_fp16 = add(x = var_938_cast_fp16, y = var_939_to_fp16)[name = tensor<string, []>("op_940_cast_fp16")];
             tensor<fp32, []> denom_epsilon_0 = const()[name = tensor<string, []>("denom_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
-            tensor<fp16, [1, 1, 1, 1]> denom_cast_fp16 = rsqrt(epsilon = denom_epsilon_0, x = var_928_cast_fp16)[name = tensor<string, []>("denom_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 1]> denom_cast_fp16 = rsqrt(epsilon = denom_epsilon_0, x = var_940_cast_fp16)[name = tensor<string, []>("denom_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> out_cast_fp16 = mul(x = zero_mean_cast_fp16, y = denom_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
             tensor<fp16, [384]> hidden_states_gamma_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59110208)))];
             tensor<fp16, [384]> hidden_states_beta_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59111040)))];
             tensor<fp16, []> hidden_states_epsilon_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
             tensor<fp16, [1, 384, 1, 1]> hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")];
-            tensor<int32, [1]> var_938_axes_0 = const()[name = tensor<string, []>("op_938_axes_0"), val = tensor<int32, [1]>([2])];
-            tensor<fp16, [1, 384, 1]> var_938_cast_fp16 = squeeze(axes = var_938_axes_0, x = hidden_states_cast_fp16)[name = tensor<string, []>("op_938_cast_fp16")];
-            tensor<int32, [3]> var_941_perm_0 = const()[name = tensor<string, []>("op_941_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_950_axes_0 = const()[name = tensor<string, []>("op_950_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 384, 1]> var_950_cast_fp16 = squeeze(axes = var_950_axes_0, x = hidden_states_cast_fp16)[name = tensor<string, []>("op_950_cast_fp16")];
+            tensor<int32, [3]> var_953_perm_0 = const()[name = tensor<string, []>("op_953_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [51864]> linear_0_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_0_bias_0_to_fp16"), val = tensor<fp16, [51864]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59111872)))];
-            tensor<fp16, [1, 1, 384]> transpose_0 = transpose(perm = var_941_perm_0, x = var_938_cast_fp16)[name = tensor<string, []>("transpose_0")];
+            tensor<fp16, [1, 1, 384]> transpose_0 = transpose(perm = var_953_perm_0, x = var_950_cast_fp16)[name = tensor<string, []>("transpose_0")];
             tensor<fp16, [1, 1, 51864]> logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = transpose_0)[name = tensor<string, []>("linear_0_cast_fp16")];
-            tensor<int32, []> var_945 = const()[name = tensor<string, []>("op_945"), val = tensor<int32, []>(1)];
-            tensor<bool, []> obj_51_interleave_0 = const()[name = tensor<string, []>("obj_51_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 1536, 1, 1]> key_cache_updates = concat(axis = var_945, interleave = obj_51_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_cast_fp16))[name = tensor<string, []>("obj_51_cast_fp16")];
-            tensor<int32, []> var_948 = const()[name = tensor<string, []>("op_948"), val = tensor<int32, []>(1)];
-            tensor<bool, []> obj_interleave_0 = const()[name = tensor<string, []>("obj_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 1536, 1, 1]> value_cache_updates = concat(axis = var_948, interleave = obj_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_cast_fp16))[name = tensor<string, []>("obj_cast_fp16")];
-        } -> (logits, key_cache_updates, value_cache_updates);
+            tensor<int32, []> var_957 = const()[name = tensor<string, []>("op_957"), val = tensor<int32, []>(1)];
+            tensor<bool, []> obj_59_interleave_0 = const()[name = tensor<string, []>("obj_59_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1536, 1, 1]> key_cache_updates = concat(axis = var_957, interleave = obj_59_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_cast_fp16))[name = tensor<string, []>("obj_59_cast_fp16")];
+            tensor<int32, []> var_960 = const()[name = tensor<string, []>("op_960"), val = tensor<int32, []>(1)];
+            tensor<bool, []> obj_61_interleave_0 = const()[name = tensor<string, []>("obj_61_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1536, 1, 1]> value_cache_updates = concat(axis = var_960, interleave = obj_61_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_cast_fp16))[name = tensor<string, []>("obj_61_cast_fp16")];
+            tensor<int32, [4]> var_971_begin_0 = const()[name = tensor<string, []>("op_971_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_971_end_0 = const()[name = tensor<string, []>("op_971_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
+            tensor<bool, [4]> var_971_end_mask_0 = const()[name = tensor<string, []>("op_971_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1500]> var_971_cast_fp16 = slice_by_index(begin = var_971_begin_0, end = var_971_end_0, end_mask = var_971_end_mask_0, x = obj_27_cast_fp16)[name = tensor<string, []>("op_971_cast_fp16")];
+            tensor<int32, [4]> var_974_begin_0 = const()[name = tensor<string, []>("op_974_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_974_end_0 = const()[name = tensor<string, []>("op_974_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
+            tensor<bool, [4]> var_974_end_mask_0 = const()[name = tensor<string, []>("op_974_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_974_squeeze_mask_0 = const()[name = tensor<string, []>("op_974_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1500]> var_974_cast_fp16 = slice_by_index(begin = var_974_begin_0, end = var_974_end_0, end_mask = var_974_end_mask_0, squeeze_mask = var_974_squeeze_mask_0, x = var_971_cast_fp16)[name = tensor<string, []>("op_974_cast_fp16")];
+            tensor<int32, [4]> var_989_begin_0 = const()[name = tensor<string, []>("op_989_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_989_end_0 = const()[name = tensor<string, []>("op_989_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
+            tensor<bool, [4]> var_989_end_mask_0 = const()[name = tensor<string, []>("op_989_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1500]> var_989_cast_fp16 = slice_by_index(begin = var_989_begin_0, end = var_989_end_0, end_mask = var_989_end_mask_0, x = obj_41_cast_fp16)[name = tensor<string, []>("op_989_cast_fp16")];
+            tensor<int32, [4]> var_992_begin_0 = const()[name = tensor<string, []>("op_992_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_992_end_0 = const()[name = tensor<string, []>("op_992_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
+            tensor<bool, [4]> var_992_end_mask_0 = const()[name = tensor<string, []>("op_992_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_992_squeeze_mask_0 = const()[name = tensor<string, []>("op_992_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1500]> var_992_cast_fp16 = slice_by_index(begin = var_992_begin_0, end = var_992_end_0, end_mask = var_992_end_mask_0, squeeze_mask = var_992_squeeze_mask_0, x = var_989_cast_fp16)[name = tensor<string, []>("op_992_cast_fp16")];
+            tensor<int32, [4]> var_1007_begin_0 = const()[name = tensor<string, []>("op_1007_begin_0"), val = tensor<int32, [4]>([0, 5, 0, 0])];
+            tensor<int32, [4]> var_1007_end_0 = const()[name = tensor<string, []>("op_1007_end_0"), val = tensor<int32, [4]>([1, 6, 1, 1500])];
+            tensor<bool, [4]> var_1007_end_mask_0 = const()[name = tensor<string, []>("op_1007_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1007_cast_fp16 = slice_by_index(begin = var_1007_begin_0, end = var_1007_end_0, end_mask = var_1007_end_mask_0, x = obj_41_cast_fp16)[name = tensor<string, []>("op_1007_cast_fp16")];
+            tensor<int32, [4]> var_1010_begin_0 = const()[name = tensor<string, []>("op_1010_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_1010_end_0 = const()[name = tensor<string, []>("op_1010_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
+            tensor<bool, [4]> var_1010_end_mask_0 = const()[name = tensor<string, []>("op_1010_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_1010_squeeze_mask_0 = const()[name = tensor<string, []>("op_1010_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1500]> var_1010_cast_fp16 = slice_by_index(begin = var_1010_begin_0, end = var_1010_end_0, end_mask = var_1010_end_mask_0, squeeze_mask = var_1010_squeeze_mask_0, x = var_1007_cast_fp16)[name = tensor<string, []>("op_1010_cast_fp16")];
+            tensor<int32, [4]> var_1025_begin_0 = const()[name = tensor<string, []>("op_1025_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_1025_end_0 = const()[name = tensor<string, []>("op_1025_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
+            tensor<bool, [4]> var_1025_end_mask_0 = const()[name = tensor<string, []>("op_1025_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1025_cast_fp16 = slice_by_index(begin = var_1025_begin_0, end = var_1025_end_0, end_mask = var_1025_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1025_cast_fp16")];
+            tensor<int32, [4]> var_1028_begin_0 = const()[name = tensor<string, []>("op_1028_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_1028_end_0 = const()[name = tensor<string, []>("op_1028_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
+            tensor<bool, [4]> var_1028_end_mask_0 = const()[name = tensor<string, []>("op_1028_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_1028_squeeze_mask_0 = const()[name = tensor<string, []>("op_1028_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1500]> var_1028_cast_fp16 = slice_by_index(begin = var_1028_begin_0, end = var_1028_end_0, end_mask = var_1028_end_mask_0, squeeze_mask = var_1028_squeeze_mask_0, x = var_1025_cast_fp16)[name = tensor<string, []>("op_1028_cast_fp16")];
+            tensor<int32, [4]> var_1043_begin_0 = const()[name = tensor<string, []>("op_1043_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
+            tensor<int32, [4]> var_1043_end_0 = const()[name = tensor<string, []>("op_1043_end_0"), val = tensor<int32, [4]>([1, 2, 1, 1500])];
+            tensor<bool, [4]> var_1043_end_mask_0 = const()[name = tensor<string, []>("op_1043_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1043_cast_fp16 = slice_by_index(begin = var_1043_begin_0, end = var_1043_end_0, end_mask = var_1043_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1043_cast_fp16")];
+            tensor<int32, [4]> var_1046_begin_0 = const()[name = tensor<string, []>("op_1046_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_1046_end_0 = const()[name = tensor<string, []>("op_1046_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
+            tensor<bool, [4]> var_1046_end_mask_0 = const()[name = tensor<string, []>("op_1046_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_1046_squeeze_mask_0 = const()[name = tensor<string, []>("op_1046_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1500]> var_1046_cast_fp16 = slice_by_index(begin = var_1046_begin_0, end = var_1046_end_0, end_mask = var_1046_end_mask_0, squeeze_mask = var_1046_squeeze_mask_0, x = var_1043_cast_fp16)[name = tensor<string, []>("op_1046_cast_fp16")];
+            tensor<int32, [4]> var_1061_begin_0 = const()[name = tensor<string, []>("op_1061_begin_0"), val = tensor<int32, [4]>([0, 2, 0, 0])];
+            tensor<int32, [4]> var_1061_end_0 = const()[name = tensor<string, []>("op_1061_end_0"), val = tensor<int32, [4]>([1, 3, 1, 1500])];
+            tensor<bool, [4]> var_1061_end_mask_0 = const()[name = tensor<string, []>("op_1061_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1061_cast_fp16 = slice_by_index(begin = var_1061_begin_0, end = var_1061_end_0, end_mask = var_1061_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1061_cast_fp16")];
+            tensor<int32, [4]> var_1064_begin_0 = const()[name = tensor<string, []>("op_1064_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_1064_end_0 = const()[name = tensor<string, []>("op_1064_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
+            tensor<bool, [4]> var_1064_end_mask_0 = const()[name = tensor<string, []>("op_1064_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_1064_squeeze_mask_0 = const()[name = tensor<string, []>("op_1064_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1500]> var_1064_cast_fp16 = slice_by_index(begin = var_1064_begin_0, end = var_1064_end_0, end_mask = var_1064_end_mask_0, squeeze_mask = var_1064_squeeze_mask_0, x = var_1061_cast_fp16)[name = tensor<string, []>("op_1064_cast_fp16")];
+            tensor<int32, [4]> var_1079_begin_0 = const()[name = tensor<string, []>("op_1079_begin_0"), val = tensor<int32, [4]>([0, 3, 0, 0])];
+            tensor<int32, [4]> var_1079_end_0 = const()[name = tensor<string, []>("op_1079_end_0"), val = tensor<int32, [4]>([1, 4, 1, 1500])];
+            tensor<bool, [4]> var_1079_end_mask_0 = const()[name = tensor<string, []>("op_1079_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1079_cast_fp16 = slice_by_index(begin = var_1079_begin_0, end = var_1079_end_0, end_mask = var_1079_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1079_cast_fp16")];
+            tensor<int32, [4]> var_1082_begin_0 = const()[name = tensor<string, []>("op_1082_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_1082_end_0 = const()[name = tensor<string, []>("op_1082_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
+            tensor<bool, [4]> var_1082_end_mask_0 = const()[name = tensor<string, []>("op_1082_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_1082_squeeze_mask_0 = const()[name = tensor<string, []>("op_1082_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1500]> var_1082_cast_fp16 = slice_by_index(begin = var_1082_begin_0, end = var_1082_end_0, end_mask = var_1082_end_mask_0, squeeze_mask = var_1082_squeeze_mask_0, x = var_1079_cast_fp16)[name = tensor<string, []>("op_1082_cast_fp16")];
+            tensor<int32, [4]> var_1097_begin_0 = const()[name = tensor<string, []>("op_1097_begin_0"), val = tensor<int32, [4]>([0, 4, 0, 0])];
+            tensor<int32, [4]> var_1097_end_0 = const()[name = tensor<string, []>("op_1097_end_0"), val = tensor<int32, [4]>([1, 5, 1, 1500])];
+            tensor<bool, [4]> var_1097_end_mask_0 = const()[name = tensor<string, []>("op_1097_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1500]> var_1097_cast_fp16 = slice_by_index(begin = var_1097_begin_0, end = var_1097_end_0, end_mask = var_1097_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1097_cast_fp16")];
+            tensor<int32, [4]> var_1100_begin_0 = const()[name = tensor<string, []>("op_1100_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_1100_end_0 = const()[name = tensor<string, []>("op_1100_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])];
+            tensor<bool, [4]> var_1100_end_mask_0 = const()[name = tensor<string, []>("op_1100_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_1100_squeeze_mask_0 = const()[name = tensor<string, []>("op_1100_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1500]> var_1100_cast_fp16 = slice_by_index(begin = var_1100_begin_0, end = var_1100_end_0, end_mask = var_1100_end_mask_0, squeeze_mask = var_1100_squeeze_mask_0, x = var_1097_cast_fp16)[name = tensor<string, []>("op_1100_cast_fp16")];
+            tensor<int32, []> var_1107 = const()[name = tensor<string, []>("op_1107"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_1108_interleave_0 = const()[name = tensor<string, []>("op_1108_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 1500]> var_1108_cast_fp16 = concat(axis = var_1107, interleave = var_1108_interleave_0, values = (var_974_cast_fp16, var_992_cast_fp16, var_1010_cast_fp16, var_1028_cast_fp16, var_1046_cast_fp16, var_1064_cast_fp16, var_1082_cast_fp16, var_1100_cast_fp16))[name = tensor<string, []>("op_1108_cast_fp16")];
+            tensor<int32, [1]> var_1110 = const()[name = tensor<string, []>("op_1110"), val = tensor<int32, [1]>([1])];
+            tensor<bool, []> var_1111 = const()[name = tensor<string, []>("op_1111"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1500]> alignment_heads_weights = reduce_mean(axes = var_1110, keep_dims = var_1111, x = var_1108_cast_fp16)[name = tensor<string, []>("obj_cast_fp16")];
+        } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights);
 }
\ No newline at end of file