diff --git "a/fr/1120ms/encoder.mlmodelc/model.mil" "b/fr/1120ms/encoder.mlmodelc/model.mil"
new file mode 100644--- /dev/null
+++ "b/fr/1120ms/encoder.mlmodelc/model.mil"
@@ -0,0 +1,4439 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})]
+{
+    func main<ios18>(tensor<fp32, [1, 24, 42, 1024]> cache_channel, tensor<int32, [1]> cache_len, tensor<fp32, [1, 24, 1024, 8]> cache_time, tensor<fp32, [1, 128, 121]> mel, tensor<int32, [1]> mel_length, tensor<int32, [1]> prompt_id) {
+            tensor<int32, [4]> value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string cache_channel_to_fp16_dtype_0 = const()[name = string("cache_channel_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<int32, [4]> value_5_perm_0 = const()[name = string("value_5_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string cache_time_to_fp16_dtype_0 = const()[name = string("cache_time_to_fp16_dtype_0"), val = string("fp16")];
+            int32 var_60 = const()[name = string("op_60"), val = int32(-1)];
+            int32 var_69 = const()[name = string("op_69"), val = int32(1)];
+            tensor<int32, [3]> x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string mel_to_fp16_dtype_0 = const()[name = string("mel_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<int32, [1]> tensor_1_axes_0 = const()[name = string("tensor_1_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 128, 121]> mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = string("cast_22")];
+            tensor<fp16, [1, 121, 128]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = mel_to_fp16)[name = string("transpose_367")];
+            tensor<fp16, [1, 1, 121, 128]> tensor_1_cast_fp16 = expand_dims(axes = tensor_1_axes_0, x = x_1_cast_fp16)[name = string("tensor_1_cast_fp16")];
+            tensor<int32, [1, 121]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1, 121]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<int32, [1]> var_138_axes_0 = const()[name = string("op_138_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_138 = expand_dims(axes = var_138_axes_0, x = mel_length)[name = string("op_138")];
+            tensor<bool, [1, 121]> time_mask_1 = less(x = expand_dims_0, y = var_138)[name = string("time_mask_1")];
+            tensor<int32, [1]> var_140_axes_0 = const()[name = string("op_140_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 121, 1]> var_140 = expand_dims(axes = var_140_axes_0, x = time_mask_1)[name = string("op_140")];
+            tensor<int32, [3]> var_142_reps_0 = const()[name = string("op_142_reps_0"), val = tensor<int32, [3]>([1, 1, 128])];
+            tensor<bool, [1, 121, 128]> var_142 = tile(reps = var_142_reps_0, x = var_140)[name = string("op_142")];
+            tensor<int32, [1]> var_148_axes_0 = const()[name = string("op_148_axes_0"), val = tensor<int32, [1]>([1])];
+            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 121, 128]> var_142_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_142)[name = string("cast_21")];
+            tensor<fp16, [1, 1, 121, 128]> var_148_cast_fp16 = expand_dims(axes = var_148_axes_0, x = var_142_to_fp16)[name = string("op_148_cast_fp16")];
+            tensor<fp16, [1, 1, 121, 128]> input_1_cast_fp16 = mul(x = tensor_1_cast_fp16, y = var_148_cast_fp16)[name = string("input_1_cast_fp16")];
+            tensor<int32, [8]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("constant")];
+            fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 124, 131]> input_3_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")];
+            string tensor_3_pad_type_0 = const()[name = string("tensor_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_3_strides_0 = const()[name = string("tensor_3_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, [4]> tensor_3_pad_0 = const()[name = string("tensor_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_3_dilations_0 = const()[name = string("tensor_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 tensor_3_groups_0 = const()[name = string("tensor_3_groups_0"), val = int32(1)];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_0_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3008))))[name = string("encoder_pre_encode_conv_0_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3584)))];
+            tensor<fp16, [1, 256, 61, 65]> tensor_3_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = tensor_3_dilations_0, groups = tensor_3_groups_0, pad = tensor_3_pad_0, pad_type = tensor_3_pad_type_0, strides = tensor_3_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("tensor_3_cast_fp16")];
+            string current_lengths_1_to_fp16_dtype_0 = const()[name = string("current_lengths_1_to_fp16_dtype_0"), val = string("fp16")];
+            fp16 var_161_promoted_to_fp16 = const()[name = string("op_161_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> mel_length_to_fp16 = cast(dtype = current_lengths_1_to_fp16_dtype_0, x = mel_length)[name = string("cast_20")];
+            tensor<fp16, [1]> var_162_cast_fp16 = add(x = mel_length_to_fp16, y = var_161_promoted_to_fp16)[name = string("op_162_cast_fp16")];
+            fp16 var_163_promoted_to_fp16 = const()[name = string("op_163_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> var_164_cast_fp16 = add(x = var_162_cast_fp16, y = var_163_promoted_to_fp16)[name = string("op_164_cast_fp16")];
+            fp16 var_165_promoted_to_fp16 = const()[name = string("op_165_promoted_to_fp16"), val = fp16(0x1.8p+1)];
+            tensor<fp16, [1]> var_166_cast_fp16 = sub(x = var_164_cast_fp16, y = var_165_promoted_to_fp16)[name = string("op_166_cast_fp16")];
+            fp16 var_57_promoted_to_fp16 = const()[name = string("op_57_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> floor_div_0_cast_fp16 = floor_div(x = var_166_cast_fp16, y = var_57_promoted_to_fp16)[name = string("floor_div_0_cast_fp16")];
+            fp16 var_168_promoted_to_fp16 = const()[name = string("op_168_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_3_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_168_promoted_to_fp16)[name = string("current_lengths_3_cast_fp16")];
+            string lengths_19_dtype_0 = const()[name = string("lengths_19_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 61]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1, 61]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4160)))];
+            tensor<int32, [1]> var_177_axes_0 = const()[name = string("op_177_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_3_cast_fp16_to_int32 = cast(dtype = lengths_19_dtype_0, x = current_lengths_3_cast_fp16)[name = string("cast_19")];
+            tensor<int32, [1, 1]> var_177 = expand_dims(axes = var_177_axes_0, x = current_lengths_3_cast_fp16_to_int32)[name = string("op_177")];
+            tensor<bool, [1, 61]> time_mask_3 = less(x = expand_dims_1, y = var_177)[name = string("time_mask_3")];
+            tensor<int32, [1]> var_179_axes_0 = const()[name = string("op_179_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 61, 1]> var_179 = expand_dims(axes = var_179_axes_0, x = time_mask_3)[name = string("op_179")];
+            tensor<int32, [3]> var_181_reps_0 = const()[name = string("op_181_reps_0"), val = tensor<int32, [3]>([1, 1, 65])];
+            tensor<bool, [1, 61, 65]> var_181 = tile(reps = var_181_reps_0, x = var_179)[name = string("op_181")];
+            tensor<int32, [1]> var_187_axes_0 = const()[name = string("op_187_axes_0"), val = tensor<int32, [1]>([1])];
+            string mask_3_to_fp16_dtype_0 = const()[name = string("mask_3_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 61, 65]> var_181_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_181)[name = string("cast_18")];
+            tensor<fp16, [1, 1, 61, 65]> var_187_cast_fp16 = expand_dims(axes = var_187_axes_0, x = var_181_to_fp16)[name = string("op_187_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_3_reps_0 = const()[name = string("expanded_mask_3_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 61, 65]> expanded_mask_3_cast_fp16 = tile(reps = expanded_mask_3_reps_0, x = var_187_cast_fp16)[name = string("expanded_mask_3_cast_fp16")];
+            tensor<fp16, [1, 256, 61, 65]> input_5_cast_fp16 = mul(x = tensor_3_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<fp16, [1, 256, 61, 65]> tensor_5_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("tensor_5_cast_fp16")];
+            tensor<fp16, [1, 256, 61, 65]> input_7_cast_fp16 = mul(x = tensor_5_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_7_cast_fp16")];
+            tensor<int32, [8]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("constant")];
+            fp16 const_23_to_fp16 = const()[name = string("const_23_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 256, 64, 68]> input_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")];
+            string tensor_7_pad_type_0 = const()[name = string("tensor_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_7_strides_0 = const()[name = string("tensor_7_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            int32 tensor_7_groups_0 = const()[name = string("tensor_7_groups_0"), val = int32(256)];
+            tensor<int32, [4]> tensor_7_pad_0 = const()[name = string("tensor_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_7_dilations_0 = const()[name = string("tensor_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4480))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6848))))[name = string("encoder_pre_encode_conv_2_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7424)))];
+            tensor<fp16, [1, 256, 31, 33]> tensor_7_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = tensor_7_dilations_0, groups = tensor_7_groups_0, pad = tensor_7_pad_0, pad_type = tensor_7_pad_type_0, strides = tensor_7_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("tensor_7_cast_fp16")];
+            fp16 var_209_promoted_to_fp16 = const()[name = string("op_209_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> var_210_cast_fp16 = add(x = current_lengths_3_cast_fp16, y = var_209_promoted_to_fp16)[name = string("op_210_cast_fp16")];
+            fp16 var_211_promoted_to_fp16 = const()[name = string("op_211_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> var_212_cast_fp16 = add(x = var_210_cast_fp16, y = var_211_promoted_to_fp16)[name = string("op_212_cast_fp16")];
+            fp16 var_213_promoted_to_fp16 = const()[name = string("op_213_promoted_to_fp16"), val = fp16(0x1.8p+1)];
+            tensor<fp16, [1]> var_214_cast_fp16 = sub(x = var_212_cast_fp16, y = var_213_promoted_to_fp16)[name = string("op_214_cast_fp16")];
+            fp16 var_57_promoted_1_to_fp16 = const()[name = string("op_57_promoted_1_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> floor_div_1_cast_fp16 = floor_div(x = var_214_cast_fp16, y = var_57_promoted_1_to_fp16)[name = string("floor_div_1_cast_fp16")];
+            fp16 var_216_promoted_to_fp16 = const()[name = string("op_216_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_5_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_216_promoted_to_fp16)[name = string("current_lengths_5_cast_fp16")];
+            string lengths_21_dtype_0 = const()[name = string("lengths_21_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 31]> expand_dims_2 = const()[name = string("expand_dims_2"), val = tensor<int32, [1, 31]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8000)))];
+            tensor<int32, [1]> var_225_axes_0 = const()[name = string("op_225_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_5_cast_fp16_to_int32 = cast(dtype = lengths_21_dtype_0, x = current_lengths_5_cast_fp16)[name = string("cast_17")];
+            tensor<int32, [1, 1]> var_225 = expand_dims(axes = var_225_axes_0, x = current_lengths_5_cast_fp16_to_int32)[name = string("op_225")];
+            tensor<bool, [1, 31]> time_mask_5 = less(x = expand_dims_2, y = var_225)[name = string("time_mask_5")];
+            tensor<int32, [1]> var_227_axes_0 = const()[name = string("op_227_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 31, 1]> var_227 = expand_dims(axes = var_227_axes_0, x = time_mask_5)[name = string("op_227")];
+            tensor<int32, [3]> var_229_reps_0 = const()[name = string("op_229_reps_0"), val = tensor<int32, [3]>([1, 1, 33])];
+            tensor<bool, [1, 31, 33]> var_229 = tile(reps = var_229_reps_0, x = var_227)[name = string("op_229")];
+            tensor<int32, [1]> var_235_axes_0 = const()[name = string("op_235_axes_0"), val = tensor<int32, [1]>([1])];
+            string mask_5_to_fp16_dtype_0 = const()[name = string("mask_5_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 31, 33]> var_229_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_229)[name = string("cast_16")];
+            tensor<fp16, [1, 1, 31, 33]> var_235_cast_fp16 = expand_dims(axes = var_235_axes_0, x = var_229_to_fp16)[name = string("op_235_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_7_reps_0 = const()[name = string("expanded_mask_7_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 31, 33]> expanded_mask_7_cast_fp16 = tile(reps = expanded_mask_7_reps_0, x = var_235_cast_fp16)[name = string("expanded_mask_7_cast_fp16")];
+            tensor<fp16, [1, 256, 31, 33]> input_11_cast_fp16 = mul(x = tensor_7_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_11_cast_fp16")];
+            string tensor_9_pad_type_0 = const()[name = string("tensor_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_9_strides_0 = const()[name = string("tensor_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_9_pad_0 = const()[name = string("tensor_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_9_dilations_0 = const()[name = string("tensor_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 tensor_9_groups_0 = const()[name = string("tensor_9_groups_0"), val = int32(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_3_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8192))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73792))))[name = string("encoder_pre_encode_conv_3_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74368)))];
+            tensor<fp16, [1, 256, 31, 33]> tensor_9_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = tensor_9_dilations_0, groups = tensor_9_groups_0, pad = tensor_9_pad_0, pad_type = tensor_9_pad_type_0, strides = tensor_9_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("tensor_9_cast_fp16")];
+            tensor<fp16, [1, 256, 31, 33]> input_13_cast_fp16 = mul(x = tensor_9_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_13_cast_fp16")];
+            tensor<fp16, [1, 256, 31, 33]> tensor_11_cast_fp16 = relu(x = input_13_cast_fp16)[name = string("tensor_11_cast_fp16")];
+            tensor<fp16, [1, 256, 31, 33]> input_15_cast_fp16 = mul(x = tensor_11_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_15_cast_fp16")];
+            tensor<int32, [8]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            string input_17_mode_0 = const()[name = string("input_17_mode_0"), val = string("constant")];
+            fp16 const_41_to_fp16 = const()[name = string("const_41_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 256, 34, 36]> input_17_cast_fp16 = pad(constant_val = const_41_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")];
+            string tensor_13_pad_type_0 = const()[name = string("tensor_13_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_13_strides_0 = const()[name = string("tensor_13_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            int32 tensor_13_groups_0 = const()[name = string("tensor_13_groups_0"), val = int32(256)];
+            tensor<int32, [4]> tensor_13_pad_0 = const()[name = string("tensor_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_13_dilations_0 = const()[name = string("tensor_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_5_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74944))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77312))))[name = string("encoder_pre_encode_conv_5_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77888)))];
+            tensor<fp16, [1, 256, 16, 17]> tensor_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = tensor_13_dilations_0, groups = tensor_13_groups_0, pad = tensor_13_pad_0, pad_type = tensor_13_pad_type_0, strides = tensor_13_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("tensor_13_cast_fp16")];
+            fp16 var_272_promoted_to_fp16 = const()[name = string("op_272_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> var_273_cast_fp16 = add(x = current_lengths_5_cast_fp16, y = var_272_promoted_to_fp16)[name = string("op_273_cast_fp16")];
+            fp16 var_274_promoted_to_fp16 = const()[name = string("op_274_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> var_275_cast_fp16 = add(x = var_273_cast_fp16, y = var_274_promoted_to_fp16)[name = string("op_275_cast_fp16")];
+            fp16 var_276_promoted_to_fp16 = const()[name = string("op_276_promoted_to_fp16"), val = fp16(0x1.8p+1)];
+            tensor<fp16, [1]> var_277_cast_fp16 = sub(x = var_275_cast_fp16, y = var_276_promoted_to_fp16)[name = string("op_277_cast_fp16")];
+            fp16 var_57_promoted_2_to_fp16 = const()[name = string("op_57_promoted_2_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> floor_div_2_cast_fp16 = floor_div(x = var_277_cast_fp16, y = var_57_promoted_2_to_fp16)[name = string("floor_div_2_cast_fp16")];
+            fp16 var_279_promoted_to_fp16 = const()[name = string("op_279_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_279_promoted_to_fp16)[name = string("current_lengths_cast_fp16")];
+            string lengths_dtype_0 = const()[name = string("lengths_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 16]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78464)))];
+            tensor<int32, [1]> var_288_axes_0 = const()[name = string("op_288_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_cast_fp16_to_int32 = cast(dtype = lengths_dtype_0, x = current_lengths_cast_fp16)[name = string("cast_15")];
+            tensor<int32, [1, 1]> var_288 = expand_dims(axes = var_288_axes_0, x = current_lengths_cast_fp16_to_int32)[name = string("op_288")];
+            tensor<bool, [1, 16]> time_mask = less(x = expand_dims_3, y = var_288)[name = string("time_mask")];
+            tensor<int32, [1]> var_290_axes_0 = const()[name = string("op_290_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 16, 1]> var_290 = expand_dims(axes = var_290_axes_0, x = time_mask)[name = string("op_290")];
+            tensor<int32, [3]> var_292_reps_0 = const()[name = string("op_292_reps_0"), val = tensor<int32, [3]>([1, 1, 17])];
+            tensor<bool, [1, 16, 17]> var_292 = tile(reps = var_292_reps_0, x = var_290)[name = string("op_292")];
+            tensor<int32, [1]> var_298_axes_0 = const()[name = string("op_298_axes_0"), val = tensor<int32, [1]>([1])];
+            string mask_7_to_fp16_dtype_0 = const()[name = string("mask_7_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 16, 17]> var_292_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_292)[name = string("cast_14")];
+            tensor<fp16, [1, 1, 16, 17]> var_298_cast_fp16 = expand_dims(axes = var_298_axes_0, x = var_292_to_fp16)[name = string("op_298_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_13_reps_0 = const()[name = string("expanded_mask_13_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 16, 17]> expanded_mask_13_cast_fp16 = tile(reps = expanded_mask_13_reps_0, x = var_298_cast_fp16)[name = string("expanded_mask_13_cast_fp16")];
+            tensor<fp16, [1, 256, 16, 17]> input_19_cast_fp16 = mul(x = tensor_13_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_19_cast_fp16")];
+            string tensor_15_pad_type_0 = const()[name = string("tensor_15_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_15_strides_0 = const()[name = string("tensor_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_15_pad_0 = const()[name = string("tensor_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_15_dilations_0 = const()[name = string("tensor_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 tensor_15_groups_0 = const()[name = string("tensor_15_groups_0"), val = int32(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_6_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78592))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144192))))[name = string("encoder_pre_encode_conv_6_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144768)))];
+            tensor<fp16, [1, 256, 16, 17]> tensor_15_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = tensor_15_dilations_0, groups = tensor_15_groups_0, pad = tensor_15_pad_0, pad_type = tensor_15_pad_type_0, strides = tensor_15_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("tensor_15_cast_fp16")];
+            tensor<fp16, [1, 256, 16, 17]> input_21_cast_fp16 = mul(x = tensor_15_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_21_cast_fp16")];
+            tensor<fp16, [1, 256, 16, 17]> tensor_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("tensor_cast_fp16")];
+            tensor<fp16, [1, 256, 16, 17]> x_3_cast_fp16 = mul(x = tensor_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("x_3_cast_fp16")];
+            tensor<int32, [4]> var_332_perm_0 = const()[name = string("op_332_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_333 = const()[name = string("op_333"), val = tensor<int32, [3]>([1, 16, -1])];
+            tensor<fp16, [1, 16, 256, 17]> var_332_cast_fp16 = transpose(perm = var_332_perm_0, x = x_3_cast_fp16)[name = string("transpose_366")];
+            tensor<fp16, [1, 16, 4352]> input_23_cast_fp16 = reshape(shape = var_333, x = var_332_cast_fp16)[name = string("input_23_cast_fp16")];
+            tensor<fp16, [1024, 4352]> encoder_pre_encode_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4352]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145344))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4601856))))[name = string("encoder_pre_encode_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_pre_encode_out_bias_to_fp16 = const()[name = string("encoder_pre_encode_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4603968)))];
+            tensor<fp16, [1, 16, 1024]> linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [3]> var_343_begin_0 = const()[name = string("op_343_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_343_end_0 = const()[name = string("op_343_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
+            tensor<bool, [3]> var_343_end_mask_0 = const()[name = string("op_343_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_343_cast_fp16 = slice_by_index(begin = var_343_begin_0, end = var_343_end_0, end_mask = var_343_end_mask_0, x = linear_0_cast_fp16)[name = string("op_343_cast_fp16")];
+            int32 var_345 = const()[name = string("op_345"), val = int32(2)];
+            tensor<int32, [1]> var_346 = sub(x = current_lengths_cast_fp16_to_int32, y = var_345)[name = string("op_346")];
+            string var_346_promoted_to_fp16_dtype_0 = const()[name = string("op_346_promoted_to_fp16_dtype_0"), val = string("fp16")];
+            fp16 var_63_promoted_to_fp16 = const()[name = string("op_63_promoted_to_fp16"), val = fp16(0x0p+0)];
+            fp16 const_61_to_fp16 = const()[name = string("const_61_to_fp16"), val = fp16(inf)];
+            tensor<fp16, [1]> var_346_to_fp16 = cast(dtype = var_346_promoted_to_fp16_dtype_0, x = var_346)[name = string("cast_13")];
+            tensor<fp16, [1]> clip_0_cast_fp16 = clip(alpha = var_63_promoted_to_fp16, beta = const_61_to_fp16, x = var_346_to_fp16)[name = string("clip_0_cast_fp16")];
+            tensor<int32, [1]> max_audio_length_1 = const()[name = string("max_audio_length_1"), val = tensor<int32, [1]>([14])];
+            fp16 var_362_promoted_to_fp16 = const()[name = string("op_362_promoted_to_fp16"), val = fp16(0x1.5p+5)];
+            tensor<fp16, [1]> padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_362_promoted_to_fp16)[name = string("padding_length_cast_fp16")];
+            int32 const_63 = const()[name = string("const_63"), val = int32(-1)];
+            tensor<int32, [1]> var_364 = mul(x = cache_len, y = const_63)[name = string("op_364")];
+            int32 var_365 = const()[name = string("op_365"), val = int32(42)];
+            tensor<int32, [1]> offset = add(x = var_364, y = var_365)[name = string("offset")];
+            tensor<int32, [1]> var_405_axes_0 = const()[name = string("op_405_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1]> var_405_cast_fp16 = expand_dims(axes = var_405_axes_0, x = padding_length_cast_fp16)[name = string("op_405_cast_fp16")];
+            tensor<fp16, [1, 56]> var_404_promoted_to_fp16 = const()[name = string("op_404_promoted_to_fp16"), val = tensor<fp16, [1, 56]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606080)))];
+            tensor<bool, [1, 56]> pad_mask_1_cast_fp16 = less(x = var_404_promoted_to_fp16, y = var_405_cast_fp16)[name = string("pad_mask_1_cast_fp16")];
+            tensor<int32, [1, 56]> expand_dims_5 = const()[name = string("expand_dims_5"), val = tensor<int32, [1, 56]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606272)))];
+            tensor<int32, [1]> var_411_axes_0 = const()[name = string("op_411_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 1]> var_411 = expand_dims(axes = var_411_axes_0, x = offset)[name = string("op_411")];
+            tensor<bool, [1, 56]> pad_mask_off = greater_equal(x = expand_dims_5, y = var_411)[name = string("pad_mask_off")];
+            tensor<bool, [1, 56]> pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = string("pad_mask_3")];
+            tensor<int32, [1]> var_414_axes_0 = const()[name = string("op_414_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 56]> var_414 = expand_dims(axes = var_414_axes_0, x = pad_mask_3)[name = string("op_414")];
+            tensor<int32, [3]> var_415 = const()[name = string("op_415"), val = tensor<int32, [3]>([1, 56, 1])];
+            tensor<bool, [1, 56, 56]> pad_mask_for_att_mask_1 = tile(reps = var_415, x = var_414)[name = string("pad_mask_for_att_mask_1")];
+            tensor<int32, [3]> var_417_perm_0 = const()[name = string("op_417_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<bool, [1, 56, 56]> var_417 = transpose(perm = var_417_perm_0, x = pad_mask_for_att_mask_1)[name = string("transpose_365")];
+            tensor<bool, [1, 56, 56]> pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_417)[name = string("pad_mask_for_att_mask")];
+            tensor<bool, [1, 56, 56]> const_71 = const()[name = string("const_71"), val = tensor<bool, [1, 56, 56]>([[[true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])];
+            tensor<bool, [1, 56, 56]> att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = string("att_mask_9")];
+            tensor<bool, [1, 56, 56]> att_mask = logical_not(x = att_mask_9)[name = string("att_mask")];
+            tensor<bool, [1, 56]> pad_mask_5 = logical_not(x = pad_mask_3)[name = string("pad_mask_5")];
+            tensor<int32, [2]> pad_mask_begin_0 = const()[name = string("pad_mask_begin_0"), val = tensor<int32, [2]>([0, 42])];
+            tensor<int32, [2]> pad_mask_end_0 = const()[name = string("pad_mask_end_0"), val = tensor<int32, [2]>([1, 56])];
+            tensor<bool, [2]> pad_mask_end_mask_0 = const()[name = string("pad_mask_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<bool, [1, 14]> pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = string("pad_mask")];
+            tensor<int32, [3]> mask_9_begin_0 = const()[name = string("mask_9_begin_0"), val = tensor<int32, [3]>([0, 42, 0])];
+            tensor<int32, [3]> mask_9_end_0 = const()[name = string("mask_9_end_0"), val = tensor<int32, [3]>([1, 56, 56])];
+            tensor<bool, [3]> mask_9_end_mask_0 = const()[name = string("mask_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<bool, [1, 14, 56]> mask_9 = slice_by_index(begin = mask_9_begin_0, end = mask_9_end_0, end_mask = mask_9_end_mask_0, x = att_mask)[name = string("mask_9")];
+            tensor<int32, [4]> cache_1_begin_0 = const()[name = string("cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_1_end_0 = const()[name = string("cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 42, 1024])];
+            tensor<bool, [4]> cache_1_end_mask_0 = const()[name = string("cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_1_squeeze_mask_0 = const()[name = string("cache_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 24, 42, 1024]> cache_channel_to_fp16 = cast(dtype = cache_channel_to_fp16_dtype_0, x = cache_channel)[name = string("cast_12")];
+            tensor<fp16, [24, 1, 42, 1024]> value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = cache_channel_to_fp16)[name = string("transpose_364")];
+            tensor<fp16, [1, 42, 1024]> cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_1_cast_fp16")];
+            tensor<int32, [4]> cache_3_begin_0 = const()[name = string("cache_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_3_end_0 = const()[name = string("cache_3_end_0"), val = tensor<int32, [4]>([1, 1, 1024, 8])];
+            tensor<bool, [4]> cache_3_end_mask_0 = const()[name = string("cache_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_3_squeeze_mask_0 = const()[name = string("cache_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 24, 1024, 8]> cache_time_to_fp16 = cast(dtype = cache_time_to_fp16_dtype_0, x = cache_time)[name = string("cast_11")];
+            tensor<fp16, [24, 1, 1024, 8]> value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = cache_time_to_fp16)[name = string("transpose_363")];
+            tensor<fp16, [1, 1024, 8]> cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_3_cast_fp16")];
+            tensor<int32, [1]> input_27_axes_0 = const()[name = string("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606592)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4608704)))];
+            fp16 var_43_to_fp16 = const()[name = string("op_43_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 14, 1024]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_343_cast_fp16)[name = string("input_27_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4610816))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8805184))))[name = string("encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_0_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8813440)))];
+            tensor<fp16, [1, 14, 4096]> linear_1_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = string("input_31_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8821696))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13016064))))[name = string("encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13018176)))];
+            tensor<fp16, [1, 14, 1024]> linear_2_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_2_cast_fp16")];
+            fp16 var_456_to_fp16 = const()[name = string("op_456_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_457_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_456_to_fp16)[name = string("op_457_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_37_cast_fp16 = add(x = var_343_cast_fp16, y = var_457_cast_fp16)[name = string("input_37_cast_fp16")];
+            tensor<int32, [1]> key_1_axes_0 = const()[name = string("key_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13020288)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13022400)))];
+            tensor<fp16, [1, 14, 1024]> key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = string("key_1_cast_fp16")];
+            bool input_39_interleave_0 = const()[name = string("input_39_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_39_cast_fp16 = concat(axis = var_69, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = string("input_39_cast_fp16")];
+            tensor<int32, [3]> var_479_begin_0 = const()[name = string("op_479_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_479_end_0 = const()[name = string("op_479_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_479_end_mask_0 = const()[name = string("op_479_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_479_cast_fp16 = slice_by_index(begin = var_479_begin_0, end = var_479_end_0, end_mask = var_479_end_mask_0, x = cache_1_cast_fp16)[name = string("op_479_cast_fp16")];
+            bool var_485_interleave_0 = const()[name = string("op_485_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_485_cast_fp16 = concat(axis = var_69, interleave = var_485_interleave_0, values = (var_479_cast_fp16, key_1_cast_fp16))[name = string("op_485_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13024512))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14073152))))[name = string("encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14075264)))];
+            tensor<fp16, [1, 14, 1024]> linear_3_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized, x = key_1_cast_fp16)[name = string("linear_3_cast_fp16")];
+            tensor<int32, [4]> var_490 = const()[name = string("op_490"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_1_cast_fp16 = reshape(shape = var_490, x = linear_3_cast_fp16)[name = string("q_1_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14077376))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15126016))))[name = string("encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15128128)))];
+            tensor<fp16, [1, 56, 1024]> linear_4_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_4_cast_fp16")];
+            tensor<int32, [4]> var_495 = const()[name = string("op_495"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_1_cast_fp16 = reshape(shape = var_495, x = linear_4_cast_fp16)[name = string("k_1_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15130240))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16178880))))[name = string("encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16180992)))];
+            tensor<fp16, [1, 56, 1024]> linear_5_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_5_cast_fp16")];
+            tensor<int32, [4]> var_500 = const()[name = string("op_500"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_1_cast_fp16 = reshape(shape = var_500, x = linear_5_cast_fp16)[name = string("v_1_cast_fp16")];
+            tensor<int32, [4]> value_9_perm_0 = const()[name = string("value_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16183104)))];
+            tensor<fp16, [1, 14, 8, 128]> var_513_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = string("op_513_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16185216)))];
+            tensor<fp16, [1, 14, 8, 128]> var_515_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = string("op_515_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_1_perm_0 = const()[name = string("q_with_bias_v_1_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_7_transpose_x_0 = const()[name = string("x_7_transpose_x_0"), val = bool(false)];
+            bool x_7_transpose_y_0 = const()[name = string("x_7_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_517_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16187328))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16301056))))[name = string("op_517_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_515_cast_fp16)[name = string("transpose_362")];
+            tensor<fp16, [1, 8, 14, 111]> x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = op_517_to_fp16_quantized)[name = string("x_7_cast_fp16")];
+            tensor<int32, [8]> x_9_pad_0 = const()[name = string("x_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_9_mode_0 = const()[name = string("x_9_mode_0"), val = string("constant")];
+            fp16 const_79_to_fp16 = const()[name = string("const_79_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_9_cast_fp16 = pad(constant_val = const_79_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = string("x_9_cast_fp16")];
+            tensor<int32, [4]> var_525 = const()[name = string("op_525"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_11_cast_fp16 = reshape(shape = var_525, x = x_9_cast_fp16)[name = string("x_11_cast_fp16")];
+            tensor<int32, [4]> var_529_begin_0 = const()[name = string("op_529_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_529_end_0 = const()[name = string("op_529_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_529_end_mask_0 = const()[name = string("op_529_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_529_cast_fp16 = slice_by_index(begin = var_529_begin_0, end = var_529_end_0, end_mask = var_529_end_mask_0, x = x_11_cast_fp16)[name = string("op_529_cast_fp16")];
+            tensor<int32, [4]> var_530 = const()[name = string("op_530"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_1_cast_fp16 = reshape(shape = var_530, x = var_529_cast_fp16)[name = string("matrix_bd_1_cast_fp16")];
+            bool matrix_ac_1_transpose_x_0 = const()[name = string("matrix_ac_1_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_1_transpose_y_0 = const()[name = string("matrix_ac_1_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1_cast_fp16)[name = string("transpose_360")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_96 = transpose(perm = transpose_96_perm_0, x = var_513_cast_fp16)[name = string("transpose_361")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_96, y = transpose_97)[name = string("matrix_ac_1_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_3_begin_0 = const()[name = string("matrix_bd_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_3_end_0 = const()[name = string("matrix_bd_3_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_3_end_mask_0 = const()[name = string("matrix_bd_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = string("matrix_bd_3_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_539_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = string("op_539_cast_fp16")];
+            fp16 _inversed_scores_1_y_0_to_fp16 = const()[name = string("_inversed_scores_1_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_1_cast_fp16 = mul(x = var_539_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = string("_inversed_scores_1_cast_fp16")];
+            tensor<int32, [1]> mask_11_axes_0 = const()[name = string("mask_11_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 14, 56]> mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = string("mask_11")];
+            fp16 var_46_to_fp16 = const()[name = string("op_46_to_fp16"), val = fp16(-0x1.388p+13)];
+            tensor<fp16, [1, 8, 14, 56]> scores_3_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_11)[name = string("scores_3_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_545_cast_fp16 = softmax(axis = var_60, x = scores_3_cast_fp16)[name = string("op_545_cast_fp16")];
+            fp16 var_45_to_fp16 = const()[name = string("op_45_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 56]> input_41_cast_fp16 = select(a = var_45_to_fp16, b = var_545_cast_fp16, cond = mask_11)[name = string("input_41_cast_fp16")];
+            bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)];
+            bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_1_cast_fp16)[name = string("transpose_359")];
+            tensor<fp16, [1, 8, 14, 128]> x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_9_cast_fp16)[name = string("x_13_cast_fp16")];
+            tensor<int32, [4]> var_549_perm_0 = const()[name = string("op_549_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_550 = const()[name = string("op_550"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_549_cast_fp16 = transpose(perm = var_549_perm_0, x = x_13_cast_fp16)[name = string("transpose_358")];
+            tensor<fp16, [1, 14, 1024]> input_43_cast_fp16 = reshape(shape = var_550, x = var_549_cast_fp16)[name = string("input_43_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16301376))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17350016))))[name = string("encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17352128)))];
+            tensor<fp16, [1, 14, 1024]> linear_7_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_7_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = string("input_47_cast_fp16")];
+            tensor<int32, [1]> x_17_axes_0 = const()[name = string("x_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17354240)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17356352)))];
+            tensor<fp16, [1, 14, 1024]> x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = string("x_17_cast_fp16")];
+            tensor<int32, [3]> input_49_perm_0 = const()[name = string("input_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_51_pad_type_0 = const()[name = string("input_51_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_51_strides_0 = const()[name = string("input_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_51_pad_0 = const()[name = string("input_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_51_dilations_0 = const()[name = string("input_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_51_groups_0 = const()[name = string("input_51_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17358464))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19455680))))[name = string("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = string("transpose_357")];
+            tensor<fp16, [1, 2048, 14]> input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("input_51_cast_fp16")];
+            int32 x_19_split_num_splits_0 = const()[name = string("x_19_split_num_splits_0"), val = int32(2)];
+            int32 x_19_split_axis_0 = const()[name = string("x_19_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_19_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = string("x_19_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = string("x_19_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = string("x_19_cast_fp16")];
+            tensor<int32, [1]> var_576_axes_0 = const()[name = string("op_576_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 14]> var_576 = expand_dims(axes = var_576_axes_0, x = pad_mask)[name = string("op_576")];
+            tensor<fp16, [1, 1024, 14]> input_53_cast_fp16 = select(a = var_45_to_fp16, b = x_19_cast_fp16, cond = var_576)[name = string("input_53_cast_fp16")];
+            bool new_x_3_interleave_0 = const()[name = string("new_x_3_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_3_cast_fp16 = concat(axis = var_60, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = string("new_x_3_cast_fp16")];
+            tensor<int32, [3]> var_589_begin_0 = const()[name = string("op_589_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_589_end_0 = const()[name = string("op_589_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_589_end_mask_0 = const()[name = string("op_589_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_589_cast_fp16 = slice_by_index(begin = var_589_begin_0, end = var_589_end_0, end_mask = var_589_end_mask_0, x = new_x_3_cast_fp16)[name = string("op_589_cast_fp16")];
+            string x_21_pad_type_0 = const()[name = string("x_21_pad_type_0"), val = string("valid")];
+            int32 x_21_groups_0 = const()[name = string("x_21_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_21_strides_0 = const()[name = string("x_21_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_21_pad_0 = const()[name = string("x_21_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_21_dilations_0 = const()[name = string("x_21_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19459840))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19469120))))[name = string("encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_3_cast_fp16)[name = string("x_21_cast_fp16")];
+            tensor<int32, [3]> input_55_perm_0 = const()[name = string("input_55_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_23_axes_0 = const()[name = string("x_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19471232)))];
+            tensor<fp16, [1024]> encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19473344)))];
+            tensor<fp16, [1, 14, 1024]> input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = string("transpose_356")];
+            tensor<fp16, [1, 14, 1024]> x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = string("x_23_cast_fp16")];
+            tensor<int32, [3]> input_57_perm_0 = const()[name = string("input_57_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = string("transpose_355")];
+            tensor<fp16, [1, 1024, 14]> input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = string("input_59_cast_fp16")];
+            string x_25_pad_type_0 = const()[name = string("x_25_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_25_strides_0 = const()[name = string("x_25_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_25_pad_0 = const()[name = string("x_25_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_25_dilations_0 = const()[name = string("x_25_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_25_groups_0 = const()[name = string("x_25_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19475456))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20524096))))[name = string("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("x_25_cast_fp16")];
+            tensor<int32, [3]> input_61_perm_0 = const()[name = string("input_61_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = string("transpose_354")];
+            tensor<fp16, [1, 14, 1024]> input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = string("input_63_cast_fp16")];
+            tensor<int32, [1]> input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20526208)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20528320)))];
+            tensor<fp16, [1, 14, 1024]> input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20530432))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24724800))))[name = string("encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_0_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24733056)))];
+            tensor<fp16, [1, 14, 4096]> linear_8_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_8_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = string("input_69_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24741312))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28935680))))[name = string("encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28937792)))];
+            tensor<fp16, [1, 14, 1024]> linear_9_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_9_cast_fp16")];
+            fp16 var_632_to_fp16 = const()[name = string("op_632_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_633_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_632_to_fp16)[name = string("op_633_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_633_cast_fp16)[name = string("input_75_cast_fp16")];
+            tensor<int32, [1]> input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28939904)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28942016)))];
+            tensor<fp16, [1, 14, 1024]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")];
+            tensor<int32, [4]> cache_5_begin_0 = const()[name = string("cache_5_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_5_end_0 = const()[name = string("cache_5_end_0"), val = tensor<int32, [4]>([2, 1, 42, 1024])];
+            tensor<bool, [4]> cache_5_end_mask_0 = const()[name = string("cache_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_5_squeeze_mask_0 = const()[name = string("cache_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_5_cast_fp16")];
+            tensor<int32, [4]> cache_7_begin_0 = const()[name = string("cache_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_7_end_0 = const()[name = string("cache_7_end_0"), val = tensor<int32, [4]>([2, 1, 1024, 8])];
+            tensor<bool, [4]> cache_7_end_mask_0 = const()[name = string("cache_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_7_squeeze_mask_0 = const()[name = string("cache_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_7_cast_fp16")];
+            tensor<int32, [1]> input_79_axes_0 = const()[name = string("input_79_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28944128)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28946240)))];
+            tensor<fp16, [1, 14, 1024]> input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28948352))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33142720))))[name = string("encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_1_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33150976)))];
+            tensor<fp16, [1, 14, 4096]> linear_10_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_10_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = string("input_83_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33159232))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37353600))))[name = string("encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37355712)))];
+            tensor<fp16, [1, 14, 1024]> linear_11_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_11_cast_fp16")];
+            fp16 var_669_to_fp16 = const()[name = string("op_669_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_670_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_669_to_fp16)[name = string("op_670_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_670_cast_fp16)[name = string("input_89_cast_fp16")];
+            tensor<int32, [1]> key_3_axes_0 = const()[name = string("key_3_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37357824)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37359936)))];
+            tensor<fp16, [1, 14, 1024]> key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = string("key_3_cast_fp16")];
+            bool input_91_interleave_0 = const()[name = string("input_91_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_91_cast_fp16 = concat(axis = var_69, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = string("input_91_cast_fp16")];
+            tensor<int32, [3]> var_692_begin_0 = const()[name = string("op_692_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_692_end_0 = const()[name = string("op_692_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_692_end_mask_0 = const()[name = string("op_692_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_692_cast_fp16 = slice_by_index(begin = var_692_begin_0, end = var_692_end_0, end_mask = var_692_end_mask_0, x = cache_5_cast_fp16)[name = string("op_692_cast_fp16")];
+            bool var_698_interleave_0 = const()[name = string("op_698_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_698_cast_fp16 = concat(axis = var_69, interleave = var_698_interleave_0, values = (var_692_cast_fp16, key_3_cast_fp16))[name = string("op_698_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37362048))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38410688))))[name = string("encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38412800)))];
+            tensor<fp16, [1, 14, 1024]> linear_12_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized, x = key_3_cast_fp16)[name = string("linear_12_cast_fp16")];
+            tensor<int32, [4]> var_703 = const()[name = string("op_703"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_7_cast_fp16 = reshape(shape = var_703, x = linear_12_cast_fp16)[name = string("q_7_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38414912))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39463552))))[name = string("encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39465664)))];
+            tensor<fp16, [1, 56, 1024]> linear_13_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_13_cast_fp16")];
+            tensor<int32, [4]> var_708 = const()[name = string("op_708"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_5_cast_fp16 = reshape(shape = var_708, x = linear_13_cast_fp16)[name = string("k_5_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39467776))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40516416))))[name = string("encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40518528)))];
+            tensor<fp16, [1, 56, 1024]> linear_14_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_14_cast_fp16")];
+            tensor<int32, [4]> var_713 = const()[name = string("op_713"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_3_cast_fp16 = reshape(shape = var_713, x = linear_14_cast_fp16)[name = string("v_3_cast_fp16")];
+            tensor<int32, [4]> value_11_perm_0 = const()[name = string("value_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40520640)))];
+            tensor<fp16, [1, 14, 8, 128]> var_726_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = string("op_726_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40522752)))];
+            tensor<fp16, [1, 14, 8, 128]> var_728_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = string("op_728_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_3_perm_0 = const()[name = string("q_with_bias_v_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_33_transpose_x_0 = const()[name = string("x_33_transpose_x_0"), val = bool(false)];
+            bool x_33_transpose_y_0 = const()[name = string("x_33_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_730_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40524864))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40638592))))[name = string("op_730_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_728_cast_fp16)[name = string("transpose_353")];
+            tensor<fp16, [1, 8, 14, 111]> x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = op_730_to_fp16_quantized)[name = string("x_33_cast_fp16")];
+            tensor<int32, [8]> x_35_pad_0 = const()[name = string("x_35_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_35_mode_0 = const()[name = string("x_35_mode_0"), val = string("constant")];
+            fp16 const_92_to_fp16 = const()[name = string("const_92_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_35_cast_fp16 = pad(constant_val = const_92_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = string("x_35_cast_fp16")];
+            tensor<int32, [4]> var_738 = const()[name = string("op_738"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_37_cast_fp16 = reshape(shape = var_738, x = x_35_cast_fp16)[name = string("x_37_cast_fp16")];
+            tensor<int32, [4]> var_742_begin_0 = const()[name = string("op_742_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_742_end_0 = const()[name = string("op_742_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_742_end_mask_0 = const()[name = string("op_742_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_742_cast_fp16 = slice_by_index(begin = var_742_begin_0, end = var_742_end_0, end_mask = var_742_end_mask_0, x = x_37_cast_fp16)[name = string("op_742_cast_fp16")];
+            tensor<int32, [4]> var_743 = const()[name = string("op_743"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_5_cast_fp16 = reshape(shape = var_743, x = var_742_cast_fp16)[name = string("matrix_bd_5_cast_fp16")];
+            bool matrix_ac_3_transpose_x_0 = const()[name = string("matrix_ac_3_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_3_transpose_y_0 = const()[name = string("matrix_ac_3_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5_cast_fp16)[name = string("transpose_351")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_98 = transpose(perm = transpose_98_perm_0, x = var_726_cast_fp16)[name = string("transpose_352")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_98, y = transpose_99)[name = string("matrix_ac_3_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_7_begin_0 = const()[name = string("matrix_bd_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_7_end_0 = const()[name = string("matrix_bd_7_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_7_end_mask_0 = const()[name = string("matrix_bd_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = string("matrix_bd_7_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_752_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = string("op_752_cast_fp16")];
+            fp16 _inversed_scores_5_y_0_to_fp16 = const()[name = string("_inversed_scores_5_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_5_cast_fp16 = mul(x = var_752_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = string("_inversed_scores_5_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_7_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_11)[name = string("scores_7_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_758_cast_fp16 = softmax(axis = var_60, x = scores_7_cast_fp16)[name = string("op_758_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_93_cast_fp16 = select(a = var_45_to_fp16, b = var_758_cast_fp16, cond = mask_11)[name = string("input_93_cast_fp16")];
+            bool x_39_transpose_x_0 = const()[name = string("x_39_transpose_x_0"), val = bool(false)];
+            bool x_39_transpose_y_0 = const()[name = string("x_39_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_3_cast_fp16)[name = string("transpose_350")];
+            tensor<fp16, [1, 8, 14, 128]> x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_11_cast_fp16)[name = string("x_39_cast_fp16")];
+            tensor<int32, [4]> var_762_perm_0 = const()[name = string("op_762_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_763 = const()[name = string("op_763"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_762_cast_fp16 = transpose(perm = var_762_perm_0, x = x_39_cast_fp16)[name = string("transpose_349")];
+            tensor<fp16, [1, 14, 1024]> input_95_cast_fp16 = reshape(shape = var_763, x = var_762_cast_fp16)[name = string("input_95_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40638912))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41687552))))[name = string("encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41689664)))];
+            tensor<fp16, [1, 14, 1024]> linear_16_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_16_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = string("input_99_cast_fp16")];
+            tensor<int32, [1]> x_43_axes_0 = const()[name = string("x_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41691776)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41693888)))];
+            tensor<fp16, [1, 14, 1024]> x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = string("x_43_cast_fp16")];
+            tensor<int32, [3]> input_101_perm_0 = const()[name = string("input_101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_103_pad_type_0 = const()[name = string("input_103_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_103_strides_0 = const()[name = string("input_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_103_pad_0 = const()[name = string("input_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_103_dilations_0 = const()[name = string("input_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_103_groups_0 = const()[name = string("input_103_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41696000))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43793216))))[name = string("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = string("transpose_348")];
+            tensor<fp16, [1, 2048, 14]> input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")];
+            int32 x_45_split_num_splits_0 = const()[name = string("x_45_split_num_splits_0"), val = int32(2)];
+            int32 x_45_split_axis_0 = const()[name = string("x_45_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_45_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = string("x_45_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = string("x_45_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = string("x_45_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_105_cast_fp16 = select(a = var_45_to_fp16, b = x_45_cast_fp16, cond = var_576)[name = string("input_105_cast_fp16")];
+            bool new_x_7_interleave_0 = const()[name = string("new_x_7_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_7_cast_fp16 = concat(axis = var_60, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = string("new_x_7_cast_fp16")];
+            tensor<int32, [3]> var_802_begin_0 = const()[name = string("op_802_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_802_end_0 = const()[name = string("op_802_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_802_end_mask_0 = const()[name = string("op_802_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_802_cast_fp16 = slice_by_index(begin = var_802_begin_0, end = var_802_end_0, end_mask = var_802_end_mask_0, x = new_x_7_cast_fp16)[name = string("op_802_cast_fp16")];
+            string x_47_pad_type_0 = const()[name = string("x_47_pad_type_0"), val = string("valid")];
+            int32 x_47_groups_0 = const()[name = string("x_47_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_47_strides_0 = const()[name = string("x_47_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_47_pad_0 = const()[name = string("x_47_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_47_dilations_0 = const()[name = string("x_47_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43797376))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43806656))))[name = string("encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_7_cast_fp16)[name = string("x_47_cast_fp16")];
+            tensor<int32, [3]> input_107_perm_0 = const()[name = string("input_107_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_49_axes_0 = const()[name = string("x_49_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43808768)))];
+            tensor<fp16, [1024]> encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43810880)))];
+            tensor<fp16, [1, 14, 1024]> input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = string("transpose_347")];
+            tensor<fp16, [1, 14, 1024]> x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = string("x_49_cast_fp16")];
+            tensor<int32, [3]> input_109_perm_0 = const()[name = string("input_109_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = string("transpose_346")];
+            tensor<fp16, [1, 1024, 14]> input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = string("input_111_cast_fp16")];
+            string x_51_pad_type_0 = const()[name = string("x_51_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_51_strides_0 = const()[name = string("x_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_51_pad_0 = const()[name = string("x_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_51_dilations_0 = const()[name = string("x_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_51_groups_0 = const()[name = string("x_51_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43812992))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44861632))))[name = string("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("x_51_cast_fp16")];
+            tensor<int32, [3]> input_113_perm_0 = const()[name = string("input_113_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = string("transpose_345")];
+            tensor<fp16, [1, 14, 1024]> input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = string("input_115_cast_fp16")];
+            tensor<int32, [1]> input_117_axes_0 = const()[name = string("input_117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44863744)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44865856)))];
+            tensor<fp16, [1, 14, 1024]> input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = string("input_117_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44867968))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49062336))))[name = string("encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_1_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49070592)))];
+            tensor<fp16, [1, 14, 4096]> linear_17_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_17_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = string("input_121_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49078848))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53273216))))[name = string("encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53275328)))];
+            tensor<fp16, [1, 14, 1024]> linear_18_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_18_cast_fp16")];
+            fp16 var_845_to_fp16 = const()[name = string("op_845_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_846_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_845_to_fp16)[name = string("op_846_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_846_cast_fp16)[name = string("input_127_cast_fp16")];
+            tensor<int32, [1]> input_129_axes_0 = const()[name = string("input_129_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53277440)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53279552)))];
+            tensor<fp16, [1, 14, 1024]> input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = string("input_129_cast_fp16")];
+            tensor<int32, [4]> cache_9_begin_0 = const()[name = string("cache_9_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_9_end_0 = const()[name = string("cache_9_end_0"), val = tensor<int32, [4]>([3, 1, 42, 1024])];
+            tensor<bool, [4]> cache_9_end_mask_0 = const()[name = string("cache_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_9_squeeze_mask_0 = const()[name = string("cache_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_9_cast_fp16")];
+            tensor<int32, [4]> cache_11_begin_0 = const()[name = string("cache_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_11_end_0 = const()[name = string("cache_11_end_0"), val = tensor<int32, [4]>([3, 1, 1024, 8])];
+            tensor<bool, [4]> cache_11_end_mask_0 = const()[name = string("cache_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_11_squeeze_mask_0 = const()[name = string("cache_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_11_cast_fp16")];
+            tensor<int32, [1]> input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53281664)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53283776)))];
+            tensor<fp16, [1, 14, 1024]> input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53285888))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57480256))))[name = string("encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_2_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57488512)))];
+            tensor<fp16, [1, 14, 4096]> linear_19_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_19_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = string("input_135_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57496768))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61691136))))[name = string("encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61693248)))];
+            tensor<fp16, [1, 14, 1024]> linear_20_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_20_cast_fp16")];
+            fp16 var_882_to_fp16 = const()[name = string("op_882_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_883_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_882_to_fp16)[name = string("op_883_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_883_cast_fp16)[name = string("input_141_cast_fp16")];
+            tensor<int32, [1]> key_5_axes_0 = const()[name = string("key_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61695360)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61697472)))];
+            tensor<fp16, [1, 14, 1024]> key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = string("key_5_cast_fp16")];
+            bool input_143_interleave_0 = const()[name = string("input_143_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_143_cast_fp16 = concat(axis = var_69, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = string("input_143_cast_fp16")];
+            tensor<int32, [3]> var_905_begin_0 = const()[name = string("op_905_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_905_end_0 = const()[name = string("op_905_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_905_end_mask_0 = const()[name = string("op_905_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_905_cast_fp16 = slice_by_index(begin = var_905_begin_0, end = var_905_end_0, end_mask = var_905_end_mask_0, x = cache_9_cast_fp16)[name = string("op_905_cast_fp16")];
+            bool var_911_interleave_0 = const()[name = string("op_911_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_911_cast_fp16 = concat(axis = var_69, interleave = var_911_interleave_0, values = (var_905_cast_fp16, key_5_cast_fp16))[name = string("op_911_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61699584))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62748224))))[name = string("encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62750336)))];
+            tensor<fp16, [1, 14, 1024]> linear_21_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized, x = key_5_cast_fp16)[name = string("linear_21_cast_fp16")];
+            tensor<int32, [4]> var_916 = const()[name = string("op_916"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_13_cast_fp16 = reshape(shape = var_916, x = linear_21_cast_fp16)[name = string("q_13_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62752448))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63801088))))[name = string("encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63803200)))];
+            tensor<fp16, [1, 56, 1024]> linear_22_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_22_cast_fp16")];
+            tensor<int32, [4]> var_921 = const()[name = string("op_921"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_9_cast_fp16 = reshape(shape = var_921, x = linear_22_cast_fp16)[name = string("k_9_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63805312))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64853952))))[name = string("encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64856064)))];
+            tensor<fp16, [1, 56, 1024]> linear_23_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_23_cast_fp16")];
+            tensor<int32, [4]> var_926 = const()[name = string("op_926"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_5_cast_fp16 = reshape(shape = var_926, x = linear_23_cast_fp16)[name = string("v_5_cast_fp16")];
+            tensor<int32, [4]> value_13_perm_0 = const()[name = string("value_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64858176)))];
+            tensor<fp16, [1, 14, 8, 128]> var_939_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = string("op_939_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64860288)))];
+            tensor<fp16, [1, 14, 8, 128]> var_941_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = string("op_941_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_5_perm_0 = const()[name = string("q_with_bias_v_5_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_59_transpose_x_0 = const()[name = string("x_59_transpose_x_0"), val = bool(false)];
+            bool x_59_transpose_y_0 = const()[name = string("x_59_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_943_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64862400))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64976128))))[name = string("op_943_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_941_cast_fp16)[name = string("transpose_344")];
+            tensor<fp16, [1, 8, 14, 111]> x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = op_943_to_fp16_quantized)[name = string("x_59_cast_fp16")];
+            tensor<int32, [8]> x_61_pad_0 = const()[name = string("x_61_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_61_mode_0 = const()[name = string("x_61_mode_0"), val = string("constant")];
+            fp16 const_105_to_fp16 = const()[name = string("const_105_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_61_cast_fp16 = pad(constant_val = const_105_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = string("x_61_cast_fp16")];
+            tensor<int32, [4]> var_951 = const()[name = string("op_951"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_63_cast_fp16 = reshape(shape = var_951, x = x_61_cast_fp16)[name = string("x_63_cast_fp16")];
+            tensor<int32, [4]> var_955_begin_0 = const()[name = string("op_955_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_955_end_0 = const()[name = string("op_955_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_955_end_mask_0 = const()[name = string("op_955_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_955_cast_fp16 = slice_by_index(begin = var_955_begin_0, end = var_955_end_0, end_mask = var_955_end_mask_0, x = x_63_cast_fp16)[name = string("op_955_cast_fp16")];
+            tensor<int32, [4]> var_956 = const()[name = string("op_956"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_9_cast_fp16 = reshape(shape = var_956, x = var_955_cast_fp16)[name = string("matrix_bd_9_cast_fp16")];
+            bool matrix_ac_5_transpose_x_0 = const()[name = string("matrix_ac_5_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_5_transpose_y_0 = const()[name = string("matrix_ac_5_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9_cast_fp16)[name = string("transpose_342")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_100 = transpose(perm = transpose_100_perm_0, x = var_939_cast_fp16)[name = string("transpose_343")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_100, y = transpose_101)[name = string("matrix_ac_5_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_11_begin_0 = const()[name = string("matrix_bd_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_11_end_0 = const()[name = string("matrix_bd_11_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_11_end_mask_0 = const()[name = string("matrix_bd_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = string("matrix_bd_11_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_965_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = string("op_965_cast_fp16")];
+            fp16 _inversed_scores_9_y_0_to_fp16 = const()[name = string("_inversed_scores_9_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_9_cast_fp16 = mul(x = var_965_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = string("_inversed_scores_9_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_11_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_11)[name = string("scores_11_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_971_cast_fp16 = softmax(axis = var_60, x = scores_11_cast_fp16)[name = string("op_971_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_145_cast_fp16 = select(a = var_45_to_fp16, b = var_971_cast_fp16, cond = mask_11)[name = string("input_145_cast_fp16")];
+            bool x_65_transpose_x_0 = const()[name = string("x_65_transpose_x_0"), val = bool(false)];
+            bool x_65_transpose_y_0 = const()[name = string("x_65_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_5_cast_fp16)[name = string("transpose_341")];
+            tensor<fp16, [1, 8, 14, 128]> x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_13_cast_fp16)[name = string("x_65_cast_fp16")];
+            tensor<int32, [4]> var_975_perm_0 = const()[name = string("op_975_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_976 = const()[name = string("op_976"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_975_cast_fp16 = transpose(perm = var_975_perm_0, x = x_65_cast_fp16)[name = string("transpose_340")];
+            tensor<fp16, [1, 14, 1024]> input_147_cast_fp16 = reshape(shape = var_976, x = var_975_cast_fp16)[name = string("input_147_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64976448))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65762944))))[name = string("encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65763136)))];
+            tensor<fp16, [1, 14, 1024]> linear_25_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = string("linear_25_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = string("input_151_cast_fp16")];
+            tensor<int32, [1]> x_69_axes_0 = const()[name = string("x_69_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65765248)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65767360)))];
+            tensor<fp16, [1, 14, 1024]> x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = string("x_69_cast_fp16")];
+            tensor<int32, [3]> input_153_perm_0 = const()[name = string("input_153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_155_pad_type_0 = const()[name = string("input_155_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_155_strides_0 = const()[name = string("input_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_155_pad_0 = const()[name = string("input_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_155_dilations_0 = const()[name = string("input_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_155_groups_0 = const()[name = string("input_155_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65769472))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67866688))))[name = string("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = string("transpose_339")];
+            tensor<fp16, [1, 2048, 14]> input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")];
+            int32 x_71_split_num_splits_0 = const()[name = string("x_71_split_num_splits_0"), val = int32(2)];
+            int32 x_71_split_axis_0 = const()[name = string("x_71_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_71_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = string("x_71_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = string("x_71_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = string("x_71_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_157_cast_fp16 = select(a = var_45_to_fp16, b = x_71_cast_fp16, cond = var_576)[name = string("input_157_cast_fp16")];
+            bool new_x_11_interleave_0 = const()[name = string("new_x_11_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_11_cast_fp16 = concat(axis = var_60, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = string("new_x_11_cast_fp16")];
+            tensor<int32, [3]> var_1015_begin_0 = const()[name = string("op_1015_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_1015_end_0 = const()[name = string("op_1015_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_1015_end_mask_0 = const()[name = string("op_1015_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1015_cast_fp16 = slice_by_index(begin = var_1015_begin_0, end = var_1015_end_0, end_mask = var_1015_end_mask_0, x = new_x_11_cast_fp16)[name = string("op_1015_cast_fp16")];
+            string x_73_pad_type_0 = const()[name = string("x_73_pad_type_0"), val = string("valid")];
+            int32 x_73_groups_0 = const()[name = string("x_73_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_73_strides_0 = const()[name = string("x_73_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_73_pad_0 = const()[name = string("x_73_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_73_dilations_0 = const()[name = string("x_73_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67870848))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67880128))))[name = string("encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_11_cast_fp16)[name = string("x_73_cast_fp16")];
+            tensor<int32, [3]> input_159_perm_0 = const()[name = string("input_159_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_75_axes_0 = const()[name = string("x_75_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67882240)))];
+            tensor<fp16, [1024]> encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67884352)))];
+            tensor<fp16, [1, 14, 1024]> input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = string("transpose_338")];
+            tensor<fp16, [1, 14, 1024]> x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = string("x_75_cast_fp16")];
+            tensor<int32, [3]> input_161_perm_0 = const()[name = string("input_161_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = string("transpose_337")];
+            tensor<fp16, [1, 1024, 14]> input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = string("input_163_cast_fp16")];
+            string x_77_pad_type_0 = const()[name = string("x_77_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_77_strides_0 = const()[name = string("x_77_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_77_pad_0 = const()[name = string("x_77_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_77_dilations_0 = const()[name = string("x_77_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_77_groups_0 = const()[name = string("x_77_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67886464))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68935104))))[name = string("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("x_77_cast_fp16")];
+            tensor<int32, [3]> input_165_perm_0 = const()[name = string("input_165_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = string("transpose_336")];
+            tensor<fp16, [1, 14, 1024]> input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = string("input_167_cast_fp16")];
+            tensor<int32, [1]> input_169_axes_0 = const()[name = string("input_169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68937216)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68939328)))];
+            tensor<fp16, [1, 14, 1024]> input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = string("input_169_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68941440))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72087232))))[name = string("encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_2_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72087424)))];
+            tensor<fp16, [1, 14, 4096]> linear_26_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = string("linear_26_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = string("input_173_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72095680))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75241472))))[name = string("encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_2_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75241664)))];
+            tensor<fp16, [1, 14, 1024]> linear_27_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized, x = input_173_cast_fp16)[name = string("linear_27_cast_fp16")];
+            fp16 var_1058_to_fp16 = const()[name = string("op_1058_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1059_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1058_to_fp16)[name = string("op_1059_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_1059_cast_fp16)[name = string("input_179_cast_fp16")];
+            tensor<int32, [1]> input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75243776)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75245888)))];
+            tensor<fp16, [1, 14, 1024]> input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = string("input_181_cast_fp16")];
+            tensor<int32, [4]> cache_13_begin_0 = const()[name = string("cache_13_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_13_end_0 = const()[name = string("cache_13_end_0"), val = tensor<int32, [4]>([4, 1, 42, 1024])];
+            tensor<bool, [4]> cache_13_end_mask_0 = const()[name = string("cache_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_13_squeeze_mask_0 = const()[name = string("cache_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_13_cast_fp16")];
+            tensor<int32, [4]> cache_15_begin_0 = const()[name = string("cache_15_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_15_end_0 = const()[name = string("cache_15_end_0"), val = tensor<int32, [4]>([4, 1, 1024, 8])];
+            tensor<bool, [4]> cache_15_end_mask_0 = const()[name = string("cache_15_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_15_squeeze_mask_0 = const()[name = string("cache_15_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_15_cast_fp16")];
+            tensor<int32, [1]> input_183_axes_0 = const()[name = string("input_183_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75248000)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75250112)))];
+            tensor<fp16, [1, 14, 1024]> input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = string("input_183_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75252224))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78398016))))[name = string("encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_3_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78398208)))];
+            tensor<fp16, [1, 14, 4096]> linear_28_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = string("linear_28_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = string("input_187_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78406464))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81552256))))[name = string("encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81552448)))];
+            tensor<fp16, [1, 14, 1024]> linear_29_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = string("linear_29_cast_fp16")];
+            fp16 var_1095_to_fp16 = const()[name = string("op_1095_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1096_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1095_to_fp16)[name = string("op_1096_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_1096_cast_fp16)[name = string("input_193_cast_fp16")];
+            tensor<int32, [1]> key_7_axes_0 = const()[name = string("key_7_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81554560)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81556672)))];
+            tensor<fp16, [1, 14, 1024]> key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = string("key_7_cast_fp16")];
+            bool input_195_interleave_0 = const()[name = string("input_195_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_195_cast_fp16 = concat(axis = var_69, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = string("input_195_cast_fp16")];
+            tensor<int32, [3]> var_1118_begin_0 = const()[name = string("op_1118_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_1118_end_0 = const()[name = string("op_1118_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1118_end_mask_0 = const()[name = string("op_1118_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_1118_cast_fp16 = slice_by_index(begin = var_1118_begin_0, end = var_1118_end_0, end_mask = var_1118_end_mask_0, x = cache_13_cast_fp16)[name = string("op_1118_cast_fp16")];
+            bool var_1124_interleave_0 = const()[name = string("op_1124_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1124_cast_fp16 = concat(axis = var_69, interleave = var_1124_interleave_0, values = (var_1118_cast_fp16, key_7_cast_fp16))[name = string("op_1124_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81558784))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82345280))))[name = string("encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82345472)))];
+            tensor<fp16, [1, 14, 1024]> linear_30_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized, x = key_7_cast_fp16)[name = string("linear_30_cast_fp16")];
+            tensor<int32, [4]> var_1129 = const()[name = string("op_1129"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_19_cast_fp16 = reshape(shape = var_1129, x = linear_30_cast_fp16)[name = string("q_19_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82347584))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83134080))))[name = string("encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83134272)))];
+            tensor<fp16, [1, 56, 1024]> linear_31_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_31_cast_fp16")];
+            tensor<int32, [4]> var_1134 = const()[name = string("op_1134"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_13_cast_fp16 = reshape(shape = var_1134, x = linear_31_cast_fp16)[name = string("k_13_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83136384))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83922880))))[name = string("encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83923072)))];
+            tensor<fp16, [1, 56, 1024]> linear_32_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_32_cast_fp16")];
+            tensor<int32, [4]> var_1139 = const()[name = string("op_1139"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_7_cast_fp16 = reshape(shape = var_1139, x = linear_32_cast_fp16)[name = string("v_7_cast_fp16")];
+            tensor<int32, [4]> value_15_perm_0 = const()[name = string("value_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83925184)))];
+            tensor<fp16, [1, 14, 8, 128]> var_1152_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = string("op_1152_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83927296)))];
+            tensor<fp16, [1, 14, 8, 128]> var_1154_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = string("op_1154_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_7_perm_0 = const()[name = string("q_with_bias_v_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_85_transpose_x_0 = const()[name = string("x_85_transpose_x_0"), val = bool(false)];
+            bool x_85_transpose_y_0 = const()[name = string("x_85_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_1156_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83929408))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84043136))))[name = string("op_1156_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1154_cast_fp16)[name = string("transpose_335")];
+            tensor<fp16, [1, 8, 14, 111]> x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = op_1156_to_fp16_quantized)[name = string("x_85_cast_fp16")];
+            tensor<int32, [8]> x_87_pad_0 = const()[name = string("x_87_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("constant")];
+            fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_87_cast_fp16 = pad(constant_val = const_118_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = string("x_87_cast_fp16")];
+            tensor<int32, [4]> var_1164 = const()[name = string("op_1164"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_89_cast_fp16 = reshape(shape = var_1164, x = x_87_cast_fp16)[name = string("x_89_cast_fp16")];
+            tensor<int32, [4]> var_1168_begin_0 = const()[name = string("op_1168_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1168_end_0 = const()[name = string("op_1168_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_1168_end_mask_0 = const()[name = string("op_1168_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_1168_cast_fp16 = slice_by_index(begin = var_1168_begin_0, end = var_1168_end_0, end_mask = var_1168_end_mask_0, x = x_89_cast_fp16)[name = string("op_1168_cast_fp16")];
+            tensor<int32, [4]> var_1169 = const()[name = string("op_1169"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_13_cast_fp16 = reshape(shape = var_1169, x = var_1168_cast_fp16)[name = string("matrix_bd_13_cast_fp16")];
+            bool matrix_ac_7_transpose_x_0 = const()[name = string("matrix_ac_7_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_7_transpose_y_0 = const()[name = string("matrix_ac_7_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13_cast_fp16)[name = string("transpose_333")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1152_cast_fp16)[name = string("transpose_334")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_102, y = transpose_103)[name = string("matrix_ac_7_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_15_begin_0 = const()[name = string("matrix_bd_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_15_end_0 = const()[name = string("matrix_bd_15_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_15_end_mask_0 = const()[name = string("matrix_bd_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = string("matrix_bd_15_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_1178_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = string("op_1178_cast_fp16")];
+            fp16 _inversed_scores_13_y_0_to_fp16 = const()[name = string("_inversed_scores_13_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_13_cast_fp16 = mul(x = var_1178_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = string("_inversed_scores_13_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_15_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_11)[name = string("scores_15_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_1184_cast_fp16 = softmax(axis = var_60, x = scores_15_cast_fp16)[name = string("op_1184_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_197_cast_fp16 = select(a = var_45_to_fp16, b = var_1184_cast_fp16, cond = mask_11)[name = string("input_197_cast_fp16")];
+            bool x_91_transpose_x_0 = const()[name = string("x_91_transpose_x_0"), val = bool(false)];
+            bool x_91_transpose_y_0 = const()[name = string("x_91_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_7_cast_fp16)[name = string("transpose_332")];
+            tensor<fp16, [1, 8, 14, 128]> x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_15_cast_fp16)[name = string("x_91_cast_fp16")];
+            tensor<int32, [4]> var_1188_perm_0 = const()[name = string("op_1188_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1189 = const()[name = string("op_1189"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_1188_cast_fp16 = transpose(perm = var_1188_perm_0, x = x_91_cast_fp16)[name = string("transpose_331")];
+            tensor<fp16, [1, 14, 1024]> input_199_cast_fp16 = reshape(shape = var_1189, x = var_1188_cast_fp16)[name = string("input_199_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84043456))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84829952))))[name = string("encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84830144)))];
+            tensor<fp16, [1, 14, 1024]> linear_34_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = string("linear_34_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = string("input_203_cast_fp16")];
+            tensor<int32, [1]> x_95_axes_0 = const()[name = string("x_95_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84832256)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84834368)))];
+            tensor<fp16, [1, 14, 1024]> x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = string("x_95_cast_fp16")];
+            tensor<int32, [3]> input_205_perm_0 = const()[name = string("input_205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_207_pad_type_0 = const()[name = string("input_207_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_207_strides_0 = const()[name = string("input_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_207_pad_0 = const()[name = string("input_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_207_dilations_0 = const()[name = string("input_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_207_groups_0 = const()[name = string("input_207_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84836480))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86933696))))[name = string("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = string("transpose_330")];
+            tensor<fp16, [1, 2048, 14]> input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")];
+            int32 x_97_split_num_splits_0 = const()[name = string("x_97_split_num_splits_0"), val = int32(2)];
+            int32 x_97_split_axis_0 = const()[name = string("x_97_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_97_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = string("x_97_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = string("x_97_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = string("x_97_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_209_cast_fp16 = select(a = var_45_to_fp16, b = x_97_cast_fp16, cond = var_576)[name = string("input_209_cast_fp16")];
+            bool new_x_15_interleave_0 = const()[name = string("new_x_15_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_15_cast_fp16 = concat(axis = var_60, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = string("new_x_15_cast_fp16")];
+            tensor<int32, [3]> var_1228_begin_0 = const()[name = string("op_1228_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_1228_end_0 = const()[name = string("op_1228_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_1228_end_mask_0 = const()[name = string("op_1228_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1228_cast_fp16 = slice_by_index(begin = var_1228_begin_0, end = var_1228_end_0, end_mask = var_1228_end_mask_0, x = new_x_15_cast_fp16)[name = string("op_1228_cast_fp16")];
+            string x_99_pad_type_0 = const()[name = string("x_99_pad_type_0"), val = string("valid")];
+            int32 x_99_groups_0 = const()[name = string("x_99_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_99_strides_0 = const()[name = string("x_99_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_99_pad_0 = const()[name = string("x_99_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_99_dilations_0 = const()[name = string("x_99_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86937856))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86947136))))[name = string("encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_15_cast_fp16)[name = string("x_99_cast_fp16")];
+            tensor<int32, [3]> input_211_perm_0 = const()[name = string("input_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_101_axes_0 = const()[name = string("x_101_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86949248)))];
+            tensor<fp16, [1024]> encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86951360)))];
+            tensor<fp16, [1, 14, 1024]> input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = string("transpose_329")];
+            tensor<fp16, [1, 14, 1024]> x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = string("x_101_cast_fp16")];
+            tensor<int32, [3]> input_213_perm_0 = const()[name = string("input_213_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = string("transpose_328")];
+            tensor<fp16, [1, 1024, 14]> input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = string("input_215_cast_fp16")];
+            string x_103_pad_type_0 = const()[name = string("x_103_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_103_strides_0 = const()[name = string("x_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_103_pad_0 = const()[name = string("x_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_103_dilations_0 = const()[name = string("x_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_103_groups_0 = const()[name = string("x_103_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86953472))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88002112))))[name = string("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("x_103_cast_fp16")];
+            tensor<int32, [3]> input_217_perm_0 = const()[name = string("input_217_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = string("transpose_327")];
+            tensor<fp16, [1, 14, 1024]> input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = string("input_219_cast_fp16")];
+            tensor<int32, [1]> input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88004224)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88006336)))];
+            tensor<fp16, [1, 14, 1024]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88008448))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91154240))))[name = string("encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_3_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91154432)))];
+            tensor<fp16, [1, 14, 4096]> linear_35_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = string("linear_35_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = string("input_225_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91162688))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94308480))))[name = string("encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94308672)))];
+            tensor<fp16, [1, 14, 1024]> linear_36_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = string("linear_36_cast_fp16")];
+            fp16 var_1271_to_fp16 = const()[name = string("op_1271_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1272_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1271_to_fp16)[name = string("op_1272_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1272_cast_fp16)[name = string("input_231_cast_fp16")];
+            tensor<int32, [1]> input_233_axes_0 = const()[name = string("input_233_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94310784)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94312896)))];
+            tensor<fp16, [1, 14, 1024]> input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = string("input_233_cast_fp16")];
+            tensor<int32, [4]> cache_17_begin_0 = const()[name = string("cache_17_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_17_end_0 = const()[name = string("cache_17_end_0"), val = tensor<int32, [4]>([5, 1, 42, 1024])];
+            tensor<bool, [4]> cache_17_end_mask_0 = const()[name = string("cache_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_17_squeeze_mask_0 = const()[name = string("cache_17_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_17_cast_fp16")];
+            tensor<int32, [4]> cache_19_begin_0 = const()[name = string("cache_19_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_19_end_0 = const()[name = string("cache_19_end_0"), val = tensor<int32, [4]>([5, 1, 1024, 8])];
+            tensor<bool, [4]> cache_19_end_mask_0 = const()[name = string("cache_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_19_squeeze_mask_0 = const()[name = string("cache_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_19_cast_fp16")];
+            tensor<int32, [1]> input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94315008)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94317120)))];
+            tensor<fp16, [1, 14, 1024]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94319232))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97465024))))[name = string("encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_4_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97465216)))];
+            tensor<fp16, [1, 14, 4096]> linear_37_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized, x = input_235_cast_fp16)[name = string("linear_37_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = string("input_239_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97473472))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100619264))))[name = string("encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100619456)))];
+            tensor<fp16, [1, 14, 1024]> linear_38_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = string("linear_38_cast_fp16")];
+            fp16 var_1308_to_fp16 = const()[name = string("op_1308_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1309_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1308_to_fp16)[name = string("op_1309_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1309_cast_fp16)[name = string("input_245_cast_fp16")];
+            tensor<int32, [1]> key_9_axes_0 = const()[name = string("key_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100621568)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100623680)))];
+            tensor<fp16, [1, 14, 1024]> key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = string("key_9_cast_fp16")];
+            bool input_247_interleave_0 = const()[name = string("input_247_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_247_cast_fp16 = concat(axis = var_69, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = string("input_247_cast_fp16")];
+            tensor<int32, [3]> var_1331_begin_0 = const()[name = string("op_1331_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_1331_end_0 = const()[name = string("op_1331_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1331_end_mask_0 = const()[name = string("op_1331_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_1331_cast_fp16 = slice_by_index(begin = var_1331_begin_0, end = var_1331_end_0, end_mask = var_1331_end_mask_0, x = cache_17_cast_fp16)[name = string("op_1331_cast_fp16")];
+            bool var_1337_interleave_0 = const()[name = string("op_1337_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1337_cast_fp16 = concat(axis = var_69, interleave = var_1337_interleave_0, values = (var_1331_cast_fp16, key_9_cast_fp16))[name = string("op_1337_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100625792))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101412288))))[name = string("encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101412480)))];
+            tensor<fp16, [1, 14, 1024]> linear_39_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized, x = key_9_cast_fp16)[name = string("linear_39_cast_fp16")];
+            tensor<int32, [4]> var_1342 = const()[name = string("op_1342"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_25_cast_fp16 = reshape(shape = var_1342, x = linear_39_cast_fp16)[name = string("q_25_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101414592))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102201088))))[name = string("encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102201280)))];
+            tensor<fp16, [1, 56, 1024]> linear_40_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_40_cast_fp16")];
+            tensor<int32, [4]> var_1347 = const()[name = string("op_1347"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_17_cast_fp16 = reshape(shape = var_1347, x = linear_40_cast_fp16)[name = string("k_17_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102203392))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102989888))))[name = string("encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102990080)))];
+            tensor<fp16, [1, 56, 1024]> linear_41_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_41_cast_fp16")];
+            tensor<int32, [4]> var_1352 = const()[name = string("op_1352"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_9_cast_fp16 = reshape(shape = var_1352, x = linear_41_cast_fp16)[name = string("v_9_cast_fp16")];
+            tensor<int32, [4]> value_17_perm_0 = const()[name = string("value_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102992192)))];
+            tensor<fp16, [1, 14, 8, 128]> var_1365_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = string("op_1365_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102994304)))];
+            tensor<fp16, [1, 14, 8, 128]> var_1367_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = string("op_1367_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_9_perm_0 = const()[name = string("q_with_bias_v_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_111_transpose_x_0 = const()[name = string("x_111_transpose_x_0"), val = bool(false)];
+            bool x_111_transpose_y_0 = const()[name = string("x_111_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_1369_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102996416))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103110144))))[name = string("op_1369_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1367_cast_fp16)[name = string("transpose_326")];
+            tensor<fp16, [1, 8, 14, 111]> x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = op_1369_to_fp16_quantized)[name = string("x_111_cast_fp16")];
+            tensor<int32, [8]> x_113_pad_0 = const()[name = string("x_113_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_113_mode_0 = const()[name = string("x_113_mode_0"), val = string("constant")];
+            fp16 const_131_to_fp16 = const()[name = string("const_131_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_113_cast_fp16 = pad(constant_val = const_131_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = string("x_113_cast_fp16")];
+            tensor<int32, [4]> var_1377 = const()[name = string("op_1377"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_115_cast_fp16 = reshape(shape = var_1377, x = x_113_cast_fp16)[name = string("x_115_cast_fp16")];
+            tensor<int32, [4]> var_1381_begin_0 = const()[name = string("op_1381_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1381_end_0 = const()[name = string("op_1381_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_1381_end_mask_0 = const()[name = string("op_1381_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_1381_cast_fp16 = slice_by_index(begin = var_1381_begin_0, end = var_1381_end_0, end_mask = var_1381_end_mask_0, x = x_115_cast_fp16)[name = string("op_1381_cast_fp16")];
+            tensor<int32, [4]> var_1382 = const()[name = string("op_1382"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_17_cast_fp16 = reshape(shape = var_1382, x = var_1381_cast_fp16)[name = string("matrix_bd_17_cast_fp16")];
+            bool matrix_ac_9_transpose_x_0 = const()[name = string("matrix_ac_9_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_9_transpose_y_0 = const()[name = string("matrix_ac_9_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17_cast_fp16)[name = string("transpose_324")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1365_cast_fp16)[name = string("transpose_325")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_104, y = transpose_105)[name = string("matrix_ac_9_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_19_begin_0 = const()[name = string("matrix_bd_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_19_end_0 = const()[name = string("matrix_bd_19_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_19_end_mask_0 = const()[name = string("matrix_bd_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = string("matrix_bd_19_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_1391_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = string("op_1391_cast_fp16")];
+            fp16 _inversed_scores_17_y_0_to_fp16 = const()[name = string("_inversed_scores_17_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_17_cast_fp16 = mul(x = var_1391_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = string("_inversed_scores_17_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_19_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_11)[name = string("scores_19_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_1397_cast_fp16 = softmax(axis = var_60, x = scores_19_cast_fp16)[name = string("op_1397_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_249_cast_fp16 = select(a = var_45_to_fp16, b = var_1397_cast_fp16, cond = mask_11)[name = string("input_249_cast_fp16")];
+            bool x_117_transpose_x_0 = const()[name = string("x_117_transpose_x_0"), val = bool(false)];
+            bool x_117_transpose_y_0 = const()[name = string("x_117_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_9_cast_fp16)[name = string("transpose_323")];
+            tensor<fp16, [1, 8, 14, 128]> x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_17_cast_fp16)[name = string("x_117_cast_fp16")];
+            tensor<int32, [4]> var_1401_perm_0 = const()[name = string("op_1401_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1402 = const()[name = string("op_1402"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_1401_cast_fp16 = transpose(perm = var_1401_perm_0, x = x_117_cast_fp16)[name = string("transpose_322")];
+            tensor<fp16, [1, 14, 1024]> input_251_cast_fp16 = reshape(shape = var_1402, x = var_1401_cast_fp16)[name = string("input_251_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103110464))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103896960))))[name = string("encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103897152)))];
+            tensor<fp16, [1, 14, 1024]> linear_43_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = string("linear_43_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = string("input_255_cast_fp16")];
+            tensor<int32, [1]> x_121_axes_0 = const()[name = string("x_121_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103899264)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103901376)))];
+            tensor<fp16, [1, 14, 1024]> x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = string("x_121_cast_fp16")];
+            tensor<int32, [3]> input_257_perm_0 = const()[name = string("input_257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_259_pad_type_0 = const()[name = string("input_259_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_259_strides_0 = const()[name = string("input_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_259_pad_0 = const()[name = string("input_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_259_dilations_0 = const()[name = string("input_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_259_groups_0 = const()[name = string("input_259_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103903488))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106000704))))[name = string("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = string("transpose_321")];
+            tensor<fp16, [1, 2048, 14]> input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")];
+            int32 x_123_split_num_splits_0 = const()[name = string("x_123_split_num_splits_0"), val = int32(2)];
+            int32 x_123_split_axis_0 = const()[name = string("x_123_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_123_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = string("x_123_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = string("x_123_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = string("x_123_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_261_cast_fp16 = select(a = var_45_to_fp16, b = x_123_cast_fp16, cond = var_576)[name = string("input_261_cast_fp16")];
+            bool new_x_19_interleave_0 = const()[name = string("new_x_19_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_19_cast_fp16 = concat(axis = var_60, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = string("new_x_19_cast_fp16")];
+            tensor<int32, [3]> var_1441_begin_0 = const()[name = string("op_1441_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_1441_end_0 = const()[name = string("op_1441_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_1441_end_mask_0 = const()[name = string("op_1441_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1441_cast_fp16 = slice_by_index(begin = var_1441_begin_0, end = var_1441_end_0, end_mask = var_1441_end_mask_0, x = new_x_19_cast_fp16)[name = string("op_1441_cast_fp16")];
+            string x_125_pad_type_0 = const()[name = string("x_125_pad_type_0"), val = string("valid")];
+            int32 x_125_groups_0 = const()[name = string("x_125_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_125_strides_0 = const()[name = string("x_125_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_125_pad_0 = const()[name = string("x_125_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_125_dilations_0 = const()[name = string("x_125_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106004864))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106014144))))[name = string("encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_19_cast_fp16)[name = string("x_125_cast_fp16")];
+            tensor<int32, [3]> input_263_perm_0 = const()[name = string("input_263_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_127_axes_0 = const()[name = string("x_127_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106016256)))];
+            tensor<fp16, [1024]> encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106018368)))];
+            tensor<fp16, [1, 14, 1024]> input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = string("transpose_320")];
+            tensor<fp16, [1, 14, 1024]> x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = string("x_127_cast_fp16")];
+            tensor<int32, [3]> input_265_perm_0 = const()[name = string("input_265_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = string("transpose_319")];
+            tensor<fp16, [1, 1024, 14]> input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = string("input_267_cast_fp16")];
+            string x_129_pad_type_0 = const()[name = string("x_129_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_129_strides_0 = const()[name = string("x_129_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_129_pad_0 = const()[name = string("x_129_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_129_dilations_0 = const()[name = string("x_129_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_129_groups_0 = const()[name = string("x_129_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106020480))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107069120))))[name = string("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("x_129_cast_fp16")];
+            tensor<int32, [3]> input_269_perm_0 = const()[name = string("input_269_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = string("transpose_318")];
+            tensor<fp16, [1, 14, 1024]> input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = string("input_271_cast_fp16")];
+            tensor<int32, [1]> input_273_axes_0 = const()[name = string("input_273_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107071232)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107073344)))];
+            tensor<fp16, [1, 14, 1024]> input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = string("input_273_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107075456))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110221248))))[name = string("encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_4_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110221440)))];
+            tensor<fp16, [1, 14, 4096]> linear_44_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized, x = input_273_cast_fp16)[name = string("linear_44_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = string("input_277_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110229696))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113375488))))[name = string("encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113375680)))];
+            tensor<fp16, [1, 14, 1024]> linear_45_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized, x = input_277_cast_fp16)[name = string("linear_45_cast_fp16")];
+            fp16 var_1484_to_fp16 = const()[name = string("op_1484_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1485_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1484_to_fp16)[name = string("op_1485_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1485_cast_fp16)[name = string("input_283_cast_fp16")];
+            tensor<int32, [1]> input_285_axes_0 = const()[name = string("input_285_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113377792)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113379904)))];
+            tensor<fp16, [1, 14, 1024]> input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = string("input_285_cast_fp16")];
+            tensor<int32, [4]> cache_21_begin_0 = const()[name = string("cache_21_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_21_end_0 = const()[name = string("cache_21_end_0"), val = tensor<int32, [4]>([6, 1, 42, 1024])];
+            tensor<bool, [4]> cache_21_end_mask_0 = const()[name = string("cache_21_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_21_squeeze_mask_0 = const()[name = string("cache_21_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_21_cast_fp16")];
+            tensor<int32, [4]> cache_23_begin_0 = const()[name = string("cache_23_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_23_end_0 = const()[name = string("cache_23_end_0"), val = tensor<int32, [4]>([6, 1, 1024, 8])];
+            tensor<bool, [4]> cache_23_end_mask_0 = const()[name = string("cache_23_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_23_squeeze_mask_0 = const()[name = string("cache_23_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_23_cast_fp16")];
+            tensor<int32, [1]> input_287_axes_0 = const()[name = string("input_287_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113382016)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113384128)))];
+            tensor<fp16, [1, 14, 1024]> input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = string("input_287_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113386240))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116532032))))[name = string("encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_5_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116532224)))];
+            tensor<fp16, [1, 14, 4096]> linear_46_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = string("linear_46_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("input_291_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116540480))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119686272))))[name = string("encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119686464)))];
+            tensor<fp16, [1, 14, 1024]> linear_47_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = string("linear_47_cast_fp16")];
+            fp16 var_1521_to_fp16 = const()[name = string("op_1521_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1522_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1521_to_fp16)[name = string("op_1522_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1522_cast_fp16)[name = string("input_297_cast_fp16")];
+            tensor<int32, [1]> key_11_axes_0 = const()[name = string("key_11_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119688576)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119690688)))];
+            tensor<fp16, [1, 14, 1024]> key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = string("key_11_cast_fp16")];
+            bool input_299_interleave_0 = const()[name = string("input_299_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_299_cast_fp16 = concat(axis = var_69, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = string("input_299_cast_fp16")];
+            tensor<int32, [3]> var_1544_begin_0 = const()[name = string("op_1544_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_1544_end_0 = const()[name = string("op_1544_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1544_end_mask_0 = const()[name = string("op_1544_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_1544_cast_fp16 = slice_by_index(begin = var_1544_begin_0, end = var_1544_end_0, end_mask = var_1544_end_mask_0, x = cache_21_cast_fp16)[name = string("op_1544_cast_fp16")];
+            bool var_1550_interleave_0 = const()[name = string("op_1550_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1550_cast_fp16 = concat(axis = var_69, interleave = var_1550_interleave_0, values = (var_1544_cast_fp16, key_11_cast_fp16))[name = string("op_1550_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119692800))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120479296))))[name = string("encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120479488)))];
+            tensor<fp16, [1, 14, 1024]> linear_48_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized, x = key_11_cast_fp16)[name = string("linear_48_cast_fp16")];
+            tensor<int32, [4]> var_1555 = const()[name = string("op_1555"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_31_cast_fp16 = reshape(shape = var_1555, x = linear_48_cast_fp16)[name = string("q_31_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120481600))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121268096))))[name = string("encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121268288)))];
+            tensor<fp16, [1, 56, 1024]> linear_49_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_49_cast_fp16")];
+            tensor<int32, [4]> var_1560 = const()[name = string("op_1560"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_21_cast_fp16 = reshape(shape = var_1560, x = linear_49_cast_fp16)[name = string("k_21_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121270400))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122056896))))[name = string("encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122057088)))];
+            tensor<fp16, [1, 56, 1024]> linear_50_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_50_cast_fp16")];
+            tensor<int32, [4]> var_1565 = const()[name = string("op_1565"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_11_cast_fp16 = reshape(shape = var_1565, x = linear_50_cast_fp16)[name = string("v_11_cast_fp16")];
+            tensor<int32, [4]> value_19_perm_0 = const()[name = string("value_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122059200)))];
+            tensor<fp16, [1, 14, 8, 128]> var_1578_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = string("op_1578_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122061312)))];
+            tensor<fp16, [1, 14, 8, 128]> var_1580_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = string("op_1580_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_11_perm_0 = const()[name = string("q_with_bias_v_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_137_transpose_x_0 = const()[name = string("x_137_transpose_x_0"), val = bool(false)];
+            bool x_137_transpose_y_0 = const()[name = string("x_137_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_1582_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122063424))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122177152))))[name = string("op_1582_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1580_cast_fp16)[name = string("transpose_317")];
+            tensor<fp16, [1, 8, 14, 111]> x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = op_1582_to_fp16_quantized)[name = string("x_137_cast_fp16")];
+            tensor<int32, [8]> x_139_pad_0 = const()[name = string("x_139_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_139_mode_0 = const()[name = string("x_139_mode_0"), val = string("constant")];
+            fp16 const_144_to_fp16 = const()[name = string("const_144_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_139_cast_fp16 = pad(constant_val = const_144_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = string("x_139_cast_fp16")];
+            tensor<int32, [4]> var_1590 = const()[name = string("op_1590"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_141_cast_fp16 = reshape(shape = var_1590, x = x_139_cast_fp16)[name = string("x_141_cast_fp16")];
+            tensor<int32, [4]> var_1594_begin_0 = const()[name = string("op_1594_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1594_end_0 = const()[name = string("op_1594_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_1594_end_mask_0 = const()[name = string("op_1594_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_1594_cast_fp16 = slice_by_index(begin = var_1594_begin_0, end = var_1594_end_0, end_mask = var_1594_end_mask_0, x = x_141_cast_fp16)[name = string("op_1594_cast_fp16")];
+            tensor<int32, [4]> var_1595 = const()[name = string("op_1595"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_21_cast_fp16 = reshape(shape = var_1595, x = var_1594_cast_fp16)[name = string("matrix_bd_21_cast_fp16")];
+            bool matrix_ac_11_transpose_x_0 = const()[name = string("matrix_ac_11_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_11_transpose_y_0 = const()[name = string("matrix_ac_11_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21_cast_fp16)[name = string("transpose_315")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1578_cast_fp16)[name = string("transpose_316")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_106, y = transpose_107)[name = string("matrix_ac_11_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_23_begin_0 = const()[name = string("matrix_bd_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_23_end_0 = const()[name = string("matrix_bd_23_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_23_end_mask_0 = const()[name = string("matrix_bd_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = string("matrix_bd_23_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_1604_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = string("op_1604_cast_fp16")];
+            fp16 _inversed_scores_21_y_0_to_fp16 = const()[name = string("_inversed_scores_21_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_21_cast_fp16 = mul(x = var_1604_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = string("_inversed_scores_21_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_23_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_11)[name = string("scores_23_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_1610_cast_fp16 = softmax(axis = var_60, x = scores_23_cast_fp16)[name = string("op_1610_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_301_cast_fp16 = select(a = var_45_to_fp16, b = var_1610_cast_fp16, cond = mask_11)[name = string("input_301_cast_fp16")];
+            bool x_143_transpose_x_0 = const()[name = string("x_143_transpose_x_0"), val = bool(false)];
+            bool x_143_transpose_y_0 = const()[name = string("x_143_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_11_cast_fp16)[name = string("transpose_314")];
+            tensor<fp16, [1, 8, 14, 128]> x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_19_cast_fp16)[name = string("x_143_cast_fp16")];
+            tensor<int32, [4]> var_1614_perm_0 = const()[name = string("op_1614_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1615 = const()[name = string("op_1615"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_1614_cast_fp16 = transpose(perm = var_1614_perm_0, x = x_143_cast_fp16)[name = string("transpose_313")];
+            tensor<fp16, [1, 14, 1024]> input_303_cast_fp16 = reshape(shape = var_1615, x = var_1614_cast_fp16)[name = string("input_303_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122177472))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122963968))))[name = string("encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122964160)))];
+            tensor<fp16, [1, 14, 1024]> linear_52_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized, x = input_303_cast_fp16)[name = string("linear_52_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = string("input_307_cast_fp16")];
+            tensor<int32, [1]> x_147_axes_0 = const()[name = string("x_147_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122966272)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122968384)))];
+            tensor<fp16, [1, 14, 1024]> x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = string("x_147_cast_fp16")];
+            tensor<int32, [3]> input_309_perm_0 = const()[name = string("input_309_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_311_pad_type_0 = const()[name = string("input_311_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_311_strides_0 = const()[name = string("input_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_311_pad_0 = const()[name = string("input_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_311_dilations_0 = const()[name = string("input_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_311_groups_0 = const()[name = string("input_311_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122970496))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125067712))))[name = string("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = string("transpose_312")];
+            tensor<fp16, [1, 2048, 14]> input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_309_cast_fp16)[name = string("input_311_cast_fp16")];
+            int32 x_149_split_num_splits_0 = const()[name = string("x_149_split_num_splits_0"), val = int32(2)];
+            int32 x_149_split_axis_0 = const()[name = string("x_149_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_149_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = string("x_149_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = string("x_149_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = string("x_149_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_313_cast_fp16 = select(a = var_45_to_fp16, b = x_149_cast_fp16, cond = var_576)[name = string("input_313_cast_fp16")];
+            bool new_x_23_interleave_0 = const()[name = string("new_x_23_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_23_cast_fp16 = concat(axis = var_60, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = string("new_x_23_cast_fp16")];
+            tensor<int32, [3]> var_1654_begin_0 = const()[name = string("op_1654_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_1654_end_0 = const()[name = string("op_1654_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_1654_end_mask_0 = const()[name = string("op_1654_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1654_cast_fp16 = slice_by_index(begin = var_1654_begin_0, end = var_1654_end_0, end_mask = var_1654_end_mask_0, x = new_x_23_cast_fp16)[name = string("op_1654_cast_fp16")];
+            string x_151_pad_type_0 = const()[name = string("x_151_pad_type_0"), val = string("valid")];
+            int32 x_151_groups_0 = const()[name = string("x_151_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_151_strides_0 = const()[name = string("x_151_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_151_pad_0 = const()[name = string("x_151_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_151_dilations_0 = const()[name = string("x_151_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125071872))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125081152))))[name = string("encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_23_cast_fp16)[name = string("x_151_cast_fp16")];
+            tensor<int32, [3]> input_315_perm_0 = const()[name = string("input_315_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_153_axes_0 = const()[name = string("x_153_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125083264)))];
+            tensor<fp16, [1024]> encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125085376)))];
+            tensor<fp16, [1, 14, 1024]> input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = string("transpose_311")];
+            tensor<fp16, [1, 14, 1024]> x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = string("x_153_cast_fp16")];
+            tensor<int32, [3]> input_317_perm_0 = const()[name = string("input_317_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = string("transpose_310")];
+            tensor<fp16, [1, 1024, 14]> input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = string("input_319_cast_fp16")];
+            string x_155_pad_type_0 = const()[name = string("x_155_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_155_strides_0 = const()[name = string("x_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_155_pad_0 = const()[name = string("x_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_155_dilations_0 = const()[name = string("x_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_155_groups_0 = const()[name = string("x_155_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125087488))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126136128))))[name = string("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_319_cast_fp16)[name = string("x_155_cast_fp16")];
+            tensor<int32, [3]> input_321_perm_0 = const()[name = string("input_321_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = string("transpose_309")];
+            tensor<fp16, [1, 14, 1024]> input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = string("input_323_cast_fp16")];
+            tensor<int32, [1]> input_325_axes_0 = const()[name = string("input_325_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126138240)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126140352)))];
+            tensor<fp16, [1, 14, 1024]> input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = string("input_325_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126142464))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129288256))))[name = string("encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_5_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129288448)))];
+            tensor<fp16, [1, 14, 4096]> linear_53_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized, x = input_325_cast_fp16)[name = string("linear_53_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("input_329_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129296704))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132442496))))[name = string("encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132442688)))];
+            tensor<fp16, [1, 14, 1024]> linear_54_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = string("linear_54_cast_fp16")];
+            fp16 var_1697_to_fp16 = const()[name = string("op_1697_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1698_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1697_to_fp16)[name = string("op_1698_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1698_cast_fp16)[name = string("input_335_cast_fp16")];
+            tensor<int32, [1]> input_337_axes_0 = const()[name = string("input_337_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132444800)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132446912)))];
+            tensor<fp16, [1, 14, 1024]> input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = string("input_337_cast_fp16")];
+            tensor<int32, [4]> cache_25_begin_0 = const()[name = string("cache_25_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_25_end_0 = const()[name = string("cache_25_end_0"), val = tensor<int32, [4]>([7, 1, 42, 1024])];
+            tensor<bool, [4]> cache_25_end_mask_0 = const()[name = string("cache_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_25_squeeze_mask_0 = const()[name = string("cache_25_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_25_cast_fp16")];
+            tensor<int32, [4]> cache_27_begin_0 = const()[name = string("cache_27_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_27_end_0 = const()[name = string("cache_27_end_0"), val = tensor<int32, [4]>([7, 1, 1024, 8])];
+            tensor<bool, [4]> cache_27_end_mask_0 = const()[name = string("cache_27_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_27_squeeze_mask_0 = const()[name = string("cache_27_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_27_cast_fp16")];
+            tensor<int32, [1]> input_339_axes_0 = const()[name = string("input_339_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132449024)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132451136)))];
+            tensor<fp16, [1, 14, 1024]> input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = string("input_339_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132453248))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135599040))))[name = string("encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_6_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135599232)))];
+            tensor<fp16, [1, 14, 4096]> linear_55_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized, x = input_339_cast_fp16)[name = string("linear_55_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = string("input_343_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135607488))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138753280))))[name = string("encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138753472)))];
+            tensor<fp16, [1, 14, 1024]> linear_56_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized, x = input_343_cast_fp16)[name = string("linear_56_cast_fp16")];
+            fp16 var_1734_to_fp16 = const()[name = string("op_1734_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1735_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1734_to_fp16)[name = string("op_1735_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1735_cast_fp16)[name = string("input_349_cast_fp16")];
+            tensor<int32, [1]> key_13_axes_0 = const()[name = string("key_13_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138755584)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138757696)))];
+            tensor<fp16, [1, 14, 1024]> key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = string("key_13_cast_fp16")];
+            bool input_351_interleave_0 = const()[name = string("input_351_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_351_cast_fp16 = concat(axis = var_69, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = string("input_351_cast_fp16")];
+            tensor<int32, [3]> var_1757_begin_0 = const()[name = string("op_1757_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_1757_end_0 = const()[name = string("op_1757_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1757_end_mask_0 = const()[name = string("op_1757_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_1757_cast_fp16 = slice_by_index(begin = var_1757_begin_0, end = var_1757_end_0, end_mask = var_1757_end_mask_0, x = cache_25_cast_fp16)[name = string("op_1757_cast_fp16")];
+            bool var_1763_interleave_0 = const()[name = string("op_1763_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1763_cast_fp16 = concat(axis = var_69, interleave = var_1763_interleave_0, values = (var_1757_cast_fp16, key_13_cast_fp16))[name = string("op_1763_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138759808))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139546304))))[name = string("encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139546496)))];
+            tensor<fp16, [1, 14, 1024]> linear_57_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized, x = key_13_cast_fp16)[name = string("linear_57_cast_fp16")];
+            tensor<int32, [4]> var_1768 = const()[name = string("op_1768"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_37_cast_fp16 = reshape(shape = var_1768, x = linear_57_cast_fp16)[name = string("q_37_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139548608))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140335104))))[name = string("encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140335296)))];
+            tensor<fp16, [1, 56, 1024]> linear_58_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_58_cast_fp16")];
+            tensor<int32, [4]> var_1773 = const()[name = string("op_1773"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_25_cast_fp16 = reshape(shape = var_1773, x = linear_58_cast_fp16)[name = string("k_25_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140337408))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141123904))))[name = string("encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141124096)))];
+            tensor<fp16, [1, 56, 1024]> linear_59_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_59_cast_fp16")];
+            tensor<int32, [4]> var_1778 = const()[name = string("op_1778"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_13_cast_fp16 = reshape(shape = var_1778, x = linear_59_cast_fp16)[name = string("v_13_cast_fp16")];
+            tensor<int32, [4]> value_21_perm_0 = const()[name = string("value_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141126208)))];
+            tensor<fp16, [1, 14, 8, 128]> var_1791_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = string("op_1791_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141128320)))];
+            tensor<fp16, [1, 14, 8, 128]> var_1793_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = string("op_1793_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_13_perm_0 = const()[name = string("q_with_bias_v_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_163_transpose_x_0 = const()[name = string("x_163_transpose_x_0"), val = bool(false)];
+            bool x_163_transpose_y_0 = const()[name = string("x_163_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_1795_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141130432))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141244160))))[name = string("op_1795_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1793_cast_fp16)[name = string("transpose_308")];
+            tensor<fp16, [1, 8, 14, 111]> x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = op_1795_to_fp16_quantized)[name = string("x_163_cast_fp16")];
+            tensor<int32, [8]> x_165_pad_0 = const()[name = string("x_165_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_165_mode_0 = const()[name = string("x_165_mode_0"), val = string("constant")];
+            fp16 const_157_to_fp16 = const()[name = string("const_157_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_165_cast_fp16 = pad(constant_val = const_157_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = string("x_165_cast_fp16")];
+            tensor<int32, [4]> var_1803 = const()[name = string("op_1803"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_167_cast_fp16 = reshape(shape = var_1803, x = x_165_cast_fp16)[name = string("x_167_cast_fp16")];
+            tensor<int32, [4]> var_1807_begin_0 = const()[name = string("op_1807_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1807_end_0 = const()[name = string("op_1807_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_1807_end_mask_0 = const()[name = string("op_1807_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_1807_cast_fp16 = slice_by_index(begin = var_1807_begin_0, end = var_1807_end_0, end_mask = var_1807_end_mask_0, x = x_167_cast_fp16)[name = string("op_1807_cast_fp16")];
+            tensor<int32, [4]> var_1808 = const()[name = string("op_1808"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_25_cast_fp16 = reshape(shape = var_1808, x = var_1807_cast_fp16)[name = string("matrix_bd_25_cast_fp16")];
+            bool matrix_ac_13_transpose_x_0 = const()[name = string("matrix_ac_13_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_13_transpose_y_0 = const()[name = string("matrix_ac_13_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25_cast_fp16)[name = string("transpose_306")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1791_cast_fp16)[name = string("transpose_307")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_108, y = transpose_109)[name = string("matrix_ac_13_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_27_begin_0 = const()[name = string("matrix_bd_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_27_end_0 = const()[name = string("matrix_bd_27_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_27_end_mask_0 = const()[name = string("matrix_bd_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = string("matrix_bd_27_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_1817_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = string("op_1817_cast_fp16")];
+            fp16 _inversed_scores_25_y_0_to_fp16 = const()[name = string("_inversed_scores_25_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_25_cast_fp16 = mul(x = var_1817_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = string("_inversed_scores_25_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_27_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_11)[name = string("scores_27_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_1823_cast_fp16 = softmax(axis = var_60, x = scores_27_cast_fp16)[name = string("op_1823_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_353_cast_fp16 = select(a = var_45_to_fp16, b = var_1823_cast_fp16, cond = mask_11)[name = string("input_353_cast_fp16")];
+            bool x_169_transpose_x_0 = const()[name = string("x_169_transpose_x_0"), val = bool(false)];
+            bool x_169_transpose_y_0 = const()[name = string("x_169_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_13_cast_fp16)[name = string("transpose_305")];
+            tensor<fp16, [1, 8, 14, 128]> x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_21_cast_fp16)[name = string("x_169_cast_fp16")];
+            tensor<int32, [4]> var_1827_perm_0 = const()[name = string("op_1827_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1828 = const()[name = string("op_1828"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_1827_cast_fp16 = transpose(perm = var_1827_perm_0, x = x_169_cast_fp16)[name = string("transpose_304")];
+            tensor<fp16, [1, 14, 1024]> input_355_cast_fp16 = reshape(shape = var_1828, x = var_1827_cast_fp16)[name = string("input_355_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141244480))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142030976))))[name = string("encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142031168)))];
+            tensor<fp16, [1, 14, 1024]> linear_61_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized, x = input_355_cast_fp16)[name = string("linear_61_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = string("input_359_cast_fp16")];
+            tensor<int32, [1]> x_173_axes_0 = const()[name = string("x_173_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142033280)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142035392)))];
+            tensor<fp16, [1, 14, 1024]> x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = string("x_173_cast_fp16")];
+            tensor<int32, [3]> input_361_perm_0 = const()[name = string("input_361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_363_pad_type_0 = const()[name = string("input_363_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_363_strides_0 = const()[name = string("input_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_363_pad_0 = const()[name = string("input_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_363_dilations_0 = const()[name = string("input_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_363_groups_0 = const()[name = string("input_363_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142037504))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144134720))))[name = string("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = string("transpose_303")];
+            tensor<fp16, [1, 2048, 14]> input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_361_cast_fp16)[name = string("input_363_cast_fp16")];
+            int32 x_175_split_num_splits_0 = const()[name = string("x_175_split_num_splits_0"), val = int32(2)];
+            int32 x_175_split_axis_0 = const()[name = string("x_175_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_175_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = string("x_175_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = string("x_175_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = string("x_175_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_365_cast_fp16 = select(a = var_45_to_fp16, b = x_175_cast_fp16, cond = var_576)[name = string("input_365_cast_fp16")];
+            bool new_x_27_interleave_0 = const()[name = string("new_x_27_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_27_cast_fp16 = concat(axis = var_60, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = string("new_x_27_cast_fp16")];
+            tensor<int32, [3]> var_1867_begin_0 = const()[name = string("op_1867_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_1867_end_0 = const()[name = string("op_1867_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_1867_end_mask_0 = const()[name = string("op_1867_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1867_cast_fp16 = slice_by_index(begin = var_1867_begin_0, end = var_1867_end_0, end_mask = var_1867_end_mask_0, x = new_x_27_cast_fp16)[name = string("op_1867_cast_fp16")];
+            string x_177_pad_type_0 = const()[name = string("x_177_pad_type_0"), val = string("valid")];
+            int32 x_177_groups_0 = const()[name = string("x_177_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_177_strides_0 = const()[name = string("x_177_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_177_pad_0 = const()[name = string("x_177_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_177_dilations_0 = const()[name = string("x_177_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144138880))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144148160))))[name = string("encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_27_cast_fp16)[name = string("x_177_cast_fp16")];
+            tensor<int32, [3]> input_367_perm_0 = const()[name = string("input_367_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_179_axes_0 = const()[name = string("x_179_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144150272)))];
+            tensor<fp16, [1024]> encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144152384)))];
+            tensor<fp16, [1, 14, 1024]> input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = string("transpose_302")];
+            tensor<fp16, [1, 14, 1024]> x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = string("x_179_cast_fp16")];
+            tensor<int32, [3]> input_369_perm_0 = const()[name = string("input_369_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = string("transpose_301")];
+            tensor<fp16, [1, 1024, 14]> input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = string("input_371_cast_fp16")];
+            string x_181_pad_type_0 = const()[name = string("x_181_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_181_strides_0 = const()[name = string("x_181_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_181_pad_0 = const()[name = string("x_181_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_181_dilations_0 = const()[name = string("x_181_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_181_groups_0 = const()[name = string("x_181_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144154496))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145203136))))[name = string("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_371_cast_fp16)[name = string("x_181_cast_fp16")];
+            tensor<int32, [3]> input_373_perm_0 = const()[name = string("input_373_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = string("transpose_300")];
+            tensor<fp16, [1, 14, 1024]> input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = string("input_375_cast_fp16")];
+            tensor<int32, [1]> input_377_axes_0 = const()[name = string("input_377_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145205248)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145207360)))];
+            tensor<fp16, [1, 14, 1024]> input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = string("input_377_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145209472))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148355264))))[name = string("encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_6_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148355456)))];
+            tensor<fp16, [1, 14, 4096]> linear_62_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized, x = input_377_cast_fp16)[name = string("linear_62_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = string("input_381_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148363712))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151509504))))[name = string("encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151509696)))];
+            tensor<fp16, [1, 14, 1024]> linear_63_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized, x = input_381_cast_fp16)[name = string("linear_63_cast_fp16")];
+            fp16 var_1910_to_fp16 = const()[name = string("op_1910_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1911_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1910_to_fp16)[name = string("op_1911_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1911_cast_fp16)[name = string("input_387_cast_fp16")];
+            tensor<int32, [1]> input_389_axes_0 = const()[name = string("input_389_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151511808)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151513920)))];
+            tensor<fp16, [1, 14, 1024]> input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")];
+            tensor<int32, [4]> cache_29_begin_0 = const()[name = string("cache_29_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_29_end_0 = const()[name = string("cache_29_end_0"), val = tensor<int32, [4]>([8, 1, 42, 1024])];
+            tensor<bool, [4]> cache_29_end_mask_0 = const()[name = string("cache_29_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_29_squeeze_mask_0 = const()[name = string("cache_29_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_29_cast_fp16")];
+            tensor<int32, [4]> cache_31_begin_0 = const()[name = string("cache_31_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_31_end_0 = const()[name = string("cache_31_end_0"), val = tensor<int32, [4]>([8, 1, 1024, 8])];
+            tensor<bool, [4]> cache_31_end_mask_0 = const()[name = string("cache_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_31_squeeze_mask_0 = const()[name = string("cache_31_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_31_cast_fp16")];
+            tensor<int32, [1]> input_391_axes_0 = const()[name = string("input_391_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151516032)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151518144)))];
+            tensor<fp16, [1, 14, 1024]> input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = string("input_391_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151520256))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154666048))))[name = string("encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_7_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154666240)))];
+            tensor<fp16, [1, 14, 4096]> linear_64_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized, x = input_391_cast_fp16)[name = string("linear_64_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = string("input_395_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154674496))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157820288))))[name = string("encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157820480)))];
+            tensor<fp16, [1, 14, 1024]> linear_65_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized, x = input_395_cast_fp16)[name = string("linear_65_cast_fp16")];
+            fp16 var_1947_to_fp16 = const()[name = string("op_1947_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1948_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1947_to_fp16)[name = string("op_1948_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1948_cast_fp16)[name = string("input_401_cast_fp16")];
+            tensor<int32, [1]> key_15_axes_0 = const()[name = string("key_15_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157822592)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157824704)))];
+            tensor<fp16, [1, 14, 1024]> key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = string("key_15_cast_fp16")];
+            bool input_403_interleave_0 = const()[name = string("input_403_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_403_cast_fp16 = concat(axis = var_69, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = string("input_403_cast_fp16")];
+            tensor<int32, [3]> var_1970_begin_0 = const()[name = string("op_1970_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_1970_end_0 = const()[name = string("op_1970_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1970_end_mask_0 = const()[name = string("op_1970_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_1970_cast_fp16 = slice_by_index(begin = var_1970_begin_0, end = var_1970_end_0, end_mask = var_1970_end_mask_0, x = cache_29_cast_fp16)[name = string("op_1970_cast_fp16")];
+            bool var_1976_interleave_0 = const()[name = string("op_1976_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1976_cast_fp16 = concat(axis = var_69, interleave = var_1976_interleave_0, values = (var_1970_cast_fp16, key_15_cast_fp16))[name = string("op_1976_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157826816))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158613312))))[name = string("encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158613504)))];
+            tensor<fp16, [1, 14, 1024]> linear_66_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized, x = key_15_cast_fp16)[name = string("linear_66_cast_fp16")];
+            tensor<int32, [4]> var_1981 = const()[name = string("op_1981"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_43_cast_fp16 = reshape(shape = var_1981, x = linear_66_cast_fp16)[name = string("q_43_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158615616))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159402112))))[name = string("encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159402304)))];
+            tensor<fp16, [1, 56, 1024]> linear_67_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_67_cast_fp16")];
+            tensor<int32, [4]> var_1986 = const()[name = string("op_1986"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_29_cast_fp16 = reshape(shape = var_1986, x = linear_67_cast_fp16)[name = string("k_29_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159404416))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160190912))))[name = string("encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160191104)))];
+            tensor<fp16, [1, 56, 1024]> linear_68_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_68_cast_fp16")];
+            tensor<int32, [4]> var_1991 = const()[name = string("op_1991"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_15_cast_fp16 = reshape(shape = var_1991, x = linear_68_cast_fp16)[name = string("v_15_cast_fp16")];
+            tensor<int32, [4]> value_23_perm_0 = const()[name = string("value_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160193216)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2004_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = string("op_2004_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160195328)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2006_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = string("op_2006_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_15_perm_0 = const()[name = string("q_with_bias_v_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_189_transpose_x_0 = const()[name = string("x_189_transpose_x_0"), val = bool(false)];
+            bool x_189_transpose_y_0 = const()[name = string("x_189_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_2008_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160197440))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160311168))))[name = string("op_2008_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_2006_cast_fp16)[name = string("transpose_299")];
+            tensor<fp16, [1, 8, 14, 111]> x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = op_2008_to_fp16_quantized)[name = string("x_189_cast_fp16")];
+            tensor<int32, [8]> x_191_pad_0 = const()[name = string("x_191_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_191_mode_0 = const()[name = string("x_191_mode_0"), val = string("constant")];
+            fp16 const_170_to_fp16 = const()[name = string("const_170_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_191_cast_fp16 = pad(constant_val = const_170_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = string("x_191_cast_fp16")];
+            tensor<int32, [4]> var_2016 = const()[name = string("op_2016"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_193_cast_fp16 = reshape(shape = var_2016, x = x_191_cast_fp16)[name = string("x_193_cast_fp16")];
+            tensor<int32, [4]> var_2020_begin_0 = const()[name = string("op_2020_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2020_end_0 = const()[name = string("op_2020_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_2020_end_mask_0 = const()[name = string("op_2020_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_2020_cast_fp16 = slice_by_index(begin = var_2020_begin_0, end = var_2020_end_0, end_mask = var_2020_end_mask_0, x = x_193_cast_fp16)[name = string("op_2020_cast_fp16")];
+            tensor<int32, [4]> var_2021 = const()[name = string("op_2021"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_29_cast_fp16 = reshape(shape = var_2021, x = var_2020_cast_fp16)[name = string("matrix_bd_29_cast_fp16")];
+            bool matrix_ac_15_transpose_x_0 = const()[name = string("matrix_ac_15_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_15_transpose_y_0 = const()[name = string("matrix_ac_15_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29_cast_fp16)[name = string("transpose_297")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_110 = transpose(perm = transpose_110_perm_0, x = var_2004_cast_fp16)[name = string("transpose_298")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_110, y = transpose_111)[name = string("matrix_ac_15_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_31_begin_0 = const()[name = string("matrix_bd_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_31_end_0 = const()[name = string("matrix_bd_31_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_31_end_mask_0 = const()[name = string("matrix_bd_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = string("matrix_bd_31_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2030_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = string("op_2030_cast_fp16")];
+            fp16 _inversed_scores_29_y_0_to_fp16 = const()[name = string("_inversed_scores_29_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_29_cast_fp16 = mul(x = var_2030_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = string("_inversed_scores_29_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_31_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_11)[name = string("scores_31_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2036_cast_fp16 = softmax(axis = var_60, x = scores_31_cast_fp16)[name = string("op_2036_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_405_cast_fp16 = select(a = var_45_to_fp16, b = var_2036_cast_fp16, cond = mask_11)[name = string("input_405_cast_fp16")];
+            bool x_195_transpose_x_0 = const()[name = string("x_195_transpose_x_0"), val = bool(false)];
+            bool x_195_transpose_y_0 = const()[name = string("x_195_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_15_cast_fp16)[name = string("transpose_296")];
+            tensor<fp16, [1, 8, 14, 128]> x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_23_cast_fp16)[name = string("x_195_cast_fp16")];
+            tensor<int32, [4]> var_2040_perm_0 = const()[name = string("op_2040_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2041 = const()[name = string("op_2041"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_2040_cast_fp16 = transpose(perm = var_2040_perm_0, x = x_195_cast_fp16)[name = string("transpose_295")];
+            tensor<fp16, [1, 14, 1024]> input_407_cast_fp16 = reshape(shape = var_2041, x = var_2040_cast_fp16)[name = string("input_407_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160311488))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161097984))))[name = string("encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161098176)))];
+            tensor<fp16, [1, 14, 1024]> linear_70_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = string("linear_70_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = string("input_411_cast_fp16")];
+            tensor<int32, [1]> x_199_axes_0 = const()[name = string("x_199_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161100288)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161102400)))];
+            tensor<fp16, [1, 14, 1024]> x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = string("x_199_cast_fp16")];
+            tensor<int32, [3]> input_413_perm_0 = const()[name = string("input_413_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_415_pad_type_0 = const()[name = string("input_415_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_415_strides_0 = const()[name = string("input_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_415_pad_0 = const()[name = string("input_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_415_dilations_0 = const()[name = string("input_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_415_groups_0 = const()[name = string("input_415_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161104512))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163201728))))[name = string("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = string("transpose_294")];
+            tensor<fp16, [1, 2048, 14]> input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_413_cast_fp16)[name = string("input_415_cast_fp16")];
+            int32 x_201_split_num_splits_0 = const()[name = string("x_201_split_num_splits_0"), val = int32(2)];
+            int32 x_201_split_axis_0 = const()[name = string("x_201_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_201_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = string("x_201_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = string("x_201_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = string("x_201_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_417_cast_fp16 = select(a = var_45_to_fp16, b = x_201_cast_fp16, cond = var_576)[name = string("input_417_cast_fp16")];
+            bool new_x_31_interleave_0 = const()[name = string("new_x_31_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_31_cast_fp16 = concat(axis = var_60, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = string("new_x_31_cast_fp16")];
+            tensor<int32, [3]> var_2080_begin_0 = const()[name = string("op_2080_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_2080_end_0 = const()[name = string("op_2080_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_2080_end_mask_0 = const()[name = string("op_2080_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2080_cast_fp16 = slice_by_index(begin = var_2080_begin_0, end = var_2080_end_0, end_mask = var_2080_end_mask_0, x = new_x_31_cast_fp16)[name = string("op_2080_cast_fp16")];
+            string x_203_pad_type_0 = const()[name = string("x_203_pad_type_0"), val = string("valid")];
+            int32 x_203_groups_0 = const()[name = string("x_203_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_203_strides_0 = const()[name = string("x_203_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_203_pad_0 = const()[name = string("x_203_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_203_dilations_0 = const()[name = string("x_203_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163205888))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163215168))))[name = string("encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_31_cast_fp16)[name = string("x_203_cast_fp16")];
+            tensor<int32, [3]> input_419_perm_0 = const()[name = string("input_419_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_205_axes_0 = const()[name = string("x_205_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163217280)))];
+            tensor<fp16, [1024]> encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163219392)))];
+            tensor<fp16, [1, 14, 1024]> input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = string("transpose_293")];
+            tensor<fp16, [1, 14, 1024]> x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = string("x_205_cast_fp16")];
+            tensor<int32, [3]> input_421_perm_0 = const()[name = string("input_421_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = string("transpose_292")];
+            tensor<fp16, [1, 1024, 14]> input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = string("input_423_cast_fp16")];
+            string x_207_pad_type_0 = const()[name = string("x_207_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_207_strides_0 = const()[name = string("x_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_207_pad_0 = const()[name = string("x_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_207_dilations_0 = const()[name = string("x_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_207_groups_0 = const()[name = string("x_207_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163221504))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164270144))))[name = string("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_423_cast_fp16)[name = string("x_207_cast_fp16")];
+            tensor<int32, [3]> input_425_perm_0 = const()[name = string("input_425_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = string("transpose_291")];
+            tensor<fp16, [1, 14, 1024]> input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = string("input_427_cast_fp16")];
+            tensor<int32, [1]> input_429_axes_0 = const()[name = string("input_429_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164272256)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164274368)))];
+            tensor<fp16, [1, 14, 1024]> input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = string("input_429_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164276480))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167422272))))[name = string("encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_7_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167422464)))];
+            tensor<fp16, [1, 14, 4096]> linear_71_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized, x = input_429_cast_fp16)[name = string("linear_71_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = string("input_433_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167430720))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170576512))))[name = string("encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170576704)))];
+            tensor<fp16, [1, 14, 1024]> linear_72_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized, x = input_433_cast_fp16)[name = string("linear_72_cast_fp16")];
+            fp16 var_2123_to_fp16 = const()[name = string("op_2123_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2124_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2123_to_fp16)[name = string("op_2124_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_2124_cast_fp16)[name = string("input_439_cast_fp16")];
+            tensor<int32, [1]> input_441_axes_0 = const()[name = string("input_441_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170578816)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170580928)))];
+            tensor<fp16, [1, 14, 1024]> input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = string("input_441_cast_fp16")];
+            tensor<int32, [4]> cache_33_begin_0 = const()[name = string("cache_33_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_33_end_0 = const()[name = string("cache_33_end_0"), val = tensor<int32, [4]>([9, 1, 42, 1024])];
+            tensor<bool, [4]> cache_33_end_mask_0 = const()[name = string("cache_33_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_33_squeeze_mask_0 = const()[name = string("cache_33_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_33_cast_fp16")];
+            tensor<int32, [4]> cache_35_begin_0 = const()[name = string("cache_35_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_35_end_0 = const()[name = string("cache_35_end_0"), val = tensor<int32, [4]>([9, 1, 1024, 8])];
+            tensor<bool, [4]> cache_35_end_mask_0 = const()[name = string("cache_35_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_35_squeeze_mask_0 = const()[name = string("cache_35_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_35_cast_fp16")];
+            tensor<int32, [1]> input_443_axes_0 = const()[name = string("input_443_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170583040)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170585152)))];
+            tensor<fp16, [1, 14, 1024]> input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = string("input_443_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170587264))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173733056))))[name = string("encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_8_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173733248)))];
+            tensor<fp16, [1, 14, 4096]> linear_73_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("linear_73_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = string("input_447_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173741504))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176887296))))[name = string("encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176887488)))];
+            tensor<fp16, [1, 14, 1024]> linear_74_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = string("linear_74_cast_fp16")];
+            fp16 var_2160_to_fp16 = const()[name = string("op_2160_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2161_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2160_to_fp16)[name = string("op_2161_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_2161_cast_fp16)[name = string("input_453_cast_fp16")];
+            tensor<int32, [1]> key_17_axes_0 = const()[name = string("key_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176889600)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176891712)))];
+            tensor<fp16, [1, 14, 1024]> key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = string("key_17_cast_fp16")];
+            bool input_455_interleave_0 = const()[name = string("input_455_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_455_cast_fp16 = concat(axis = var_69, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = string("input_455_cast_fp16")];
+            tensor<int32, [3]> var_2183_begin_0 = const()[name = string("op_2183_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_2183_end_0 = const()[name = string("op_2183_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2183_end_mask_0 = const()[name = string("op_2183_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_2183_cast_fp16 = slice_by_index(begin = var_2183_begin_0, end = var_2183_end_0, end_mask = var_2183_end_mask_0, x = cache_33_cast_fp16)[name = string("op_2183_cast_fp16")];
+            bool var_2189_interleave_0 = const()[name = string("op_2189_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_2189_cast_fp16 = concat(axis = var_69, interleave = var_2189_interleave_0, values = (var_2183_cast_fp16, key_17_cast_fp16))[name = string("op_2189_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176893824))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177680320))))[name = string("encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177680512)))];
+            tensor<fp16, [1, 14, 1024]> linear_75_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized, x = key_17_cast_fp16)[name = string("linear_75_cast_fp16")];
+            tensor<int32, [4]> var_2194 = const()[name = string("op_2194"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_49_cast_fp16 = reshape(shape = var_2194, x = linear_75_cast_fp16)[name = string("q_49_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177682624))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178469120))))[name = string("encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178469312)))];
+            tensor<fp16, [1, 56, 1024]> linear_76_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_76_cast_fp16")];
+            tensor<int32, [4]> var_2199 = const()[name = string("op_2199"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_33_cast_fp16 = reshape(shape = var_2199, x = linear_76_cast_fp16)[name = string("k_33_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178471424))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179257920))))[name = string("encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179258112)))];
+            tensor<fp16, [1, 56, 1024]> linear_77_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_77_cast_fp16")];
+            tensor<int32, [4]> var_2204 = const()[name = string("op_2204"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_17_cast_fp16 = reshape(shape = var_2204, x = linear_77_cast_fp16)[name = string("v_17_cast_fp16")];
+            tensor<int32, [4]> value_25_perm_0 = const()[name = string("value_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179260224)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2217_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = string("op_2217_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179262336)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2219_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = string("op_2219_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_17_perm_0 = const()[name = string("q_with_bias_v_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_215_transpose_x_0 = const()[name = string("x_215_transpose_x_0"), val = bool(false)];
+            bool x_215_transpose_y_0 = const()[name = string("x_215_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_2221_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179264448))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179378176))))[name = string("op_2221_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2219_cast_fp16)[name = string("transpose_290")];
+            tensor<fp16, [1, 8, 14, 111]> x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = op_2221_to_fp16_quantized)[name = string("x_215_cast_fp16")];
+            tensor<int32, [8]> x_217_pad_0 = const()[name = string("x_217_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_217_mode_0 = const()[name = string("x_217_mode_0"), val = string("constant")];
+            fp16 const_183_to_fp16 = const()[name = string("const_183_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_217_cast_fp16 = pad(constant_val = const_183_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = string("x_217_cast_fp16")];
+            tensor<int32, [4]> var_2229 = const()[name = string("op_2229"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_219_cast_fp16 = reshape(shape = var_2229, x = x_217_cast_fp16)[name = string("x_219_cast_fp16")];
+            tensor<int32, [4]> var_2233_begin_0 = const()[name = string("op_2233_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2233_end_0 = const()[name = string("op_2233_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_2233_end_mask_0 = const()[name = string("op_2233_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_2233_cast_fp16 = slice_by_index(begin = var_2233_begin_0, end = var_2233_end_0, end_mask = var_2233_end_mask_0, x = x_219_cast_fp16)[name = string("op_2233_cast_fp16")];
+            tensor<int32, [4]> var_2234 = const()[name = string("op_2234"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_33_cast_fp16 = reshape(shape = var_2234, x = var_2233_cast_fp16)[name = string("matrix_bd_33_cast_fp16")];
+            bool matrix_ac_17_transpose_x_0 = const()[name = string("matrix_ac_17_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_17_transpose_y_0 = const()[name = string("matrix_ac_17_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33_cast_fp16)[name = string("transpose_288")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_112 = transpose(perm = transpose_112_perm_0, x = var_2217_cast_fp16)[name = string("transpose_289")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_112, y = transpose_113)[name = string("matrix_ac_17_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_35_begin_0 = const()[name = string("matrix_bd_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_35_end_0 = const()[name = string("matrix_bd_35_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_35_end_mask_0 = const()[name = string("matrix_bd_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = string("matrix_bd_35_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2243_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = string("op_2243_cast_fp16")];
+            fp16 _inversed_scores_33_y_0_to_fp16 = const()[name = string("_inversed_scores_33_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_33_cast_fp16 = mul(x = var_2243_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = string("_inversed_scores_33_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_35_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_11)[name = string("scores_35_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2249_cast_fp16 = softmax(axis = var_60, x = scores_35_cast_fp16)[name = string("op_2249_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_457_cast_fp16 = select(a = var_45_to_fp16, b = var_2249_cast_fp16, cond = mask_11)[name = string("input_457_cast_fp16")];
+            bool x_221_transpose_x_0 = const()[name = string("x_221_transpose_x_0"), val = bool(false)];
+            bool x_221_transpose_y_0 = const()[name = string("x_221_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_17_cast_fp16)[name = string("transpose_287")];
+            tensor<fp16, [1, 8, 14, 128]> x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_25_cast_fp16)[name = string("x_221_cast_fp16")];
+            tensor<int32, [4]> var_2253_perm_0 = const()[name = string("op_2253_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2254 = const()[name = string("op_2254"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_2253_cast_fp16 = transpose(perm = var_2253_perm_0, x = x_221_cast_fp16)[name = string("transpose_286")];
+            tensor<fp16, [1, 14, 1024]> input_459_cast_fp16 = reshape(shape = var_2254, x = var_2253_cast_fp16)[name = string("input_459_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179378496))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180164992))))[name = string("encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180165184)))];
+            tensor<fp16, [1, 14, 1024]> linear_79_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized, x = input_459_cast_fp16)[name = string("linear_79_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = string("input_463_cast_fp16")];
+            tensor<int32, [1]> x_225_axes_0 = const()[name = string("x_225_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180167296)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180169408)))];
+            tensor<fp16, [1, 14, 1024]> x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = string("x_225_cast_fp16")];
+            tensor<int32, [3]> input_465_perm_0 = const()[name = string("input_465_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_467_pad_type_0 = const()[name = string("input_467_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_467_strides_0 = const()[name = string("input_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_467_pad_0 = const()[name = string("input_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_467_dilations_0 = const()[name = string("input_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_467_groups_0 = const()[name = string("input_467_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180171520))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182268736))))[name = string("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = string("transpose_285")];
+            tensor<fp16, [1, 2048, 14]> input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_465_cast_fp16)[name = string("input_467_cast_fp16")];
+            int32 x_227_split_num_splits_0 = const()[name = string("x_227_split_num_splits_0"), val = int32(2)];
+            int32 x_227_split_axis_0 = const()[name = string("x_227_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_227_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = string("x_227_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = string("x_227_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = string("x_227_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_469_cast_fp16 = select(a = var_45_to_fp16, b = x_227_cast_fp16, cond = var_576)[name = string("input_469_cast_fp16")];
+            bool new_x_35_interleave_0 = const()[name = string("new_x_35_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_35_cast_fp16 = concat(axis = var_60, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = string("new_x_35_cast_fp16")];
+            tensor<int32, [3]> var_2293_begin_0 = const()[name = string("op_2293_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_2293_end_0 = const()[name = string("op_2293_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_2293_end_mask_0 = const()[name = string("op_2293_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2293_cast_fp16 = slice_by_index(begin = var_2293_begin_0, end = var_2293_end_0, end_mask = var_2293_end_mask_0, x = new_x_35_cast_fp16)[name = string("op_2293_cast_fp16")];
+            string x_229_pad_type_0 = const()[name = string("x_229_pad_type_0"), val = string("valid")];
+            int32 x_229_groups_0 = const()[name = string("x_229_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_229_strides_0 = const()[name = string("x_229_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_229_pad_0 = const()[name = string("x_229_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_229_dilations_0 = const()[name = string("x_229_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182272896))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182282176))))[name = string("encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_35_cast_fp16)[name = string("x_229_cast_fp16")];
+            tensor<int32, [3]> input_471_perm_0 = const()[name = string("input_471_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_231_axes_0 = const()[name = string("x_231_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182284288)))];
+            tensor<fp16, [1024]> encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182286400)))];
+            tensor<fp16, [1, 14, 1024]> input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = string("transpose_284")];
+            tensor<fp16, [1, 14, 1024]> x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = string("x_231_cast_fp16")];
+            tensor<int32, [3]> input_473_perm_0 = const()[name = string("input_473_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = string("transpose_283")];
+            tensor<fp16, [1, 1024, 14]> input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = string("input_475_cast_fp16")];
+            string x_233_pad_type_0 = const()[name = string("x_233_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_233_strides_0 = const()[name = string("x_233_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_233_pad_0 = const()[name = string("x_233_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_233_dilations_0 = const()[name = string("x_233_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_233_groups_0 = const()[name = string("x_233_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182288512))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183337152))))[name = string("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_475_cast_fp16)[name = string("x_233_cast_fp16")];
+            tensor<int32, [3]> input_477_perm_0 = const()[name = string("input_477_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = string("transpose_282")];
+            tensor<fp16, [1, 14, 1024]> input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = string("input_479_cast_fp16")];
+            tensor<int32, [1]> input_481_axes_0 = const()[name = string("input_481_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183339264)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183341376)))];
+            tensor<fp16, [1, 14, 1024]> input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = string("input_481_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183343488))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186489280))))[name = string("encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_8_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186489472)))];
+            tensor<fp16, [1, 14, 4096]> linear_80_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized, x = input_481_cast_fp16)[name = string("linear_80_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = string("input_485_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186497728))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189643520))))[name = string("encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189643712)))];
+            tensor<fp16, [1, 14, 1024]> linear_81_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = string("linear_81_cast_fp16")];
+            fp16 var_2336_to_fp16 = const()[name = string("op_2336_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2337_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2336_to_fp16)[name = string("op_2337_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2337_cast_fp16)[name = string("input_491_cast_fp16")];
+            tensor<int32, [1]> input_493_axes_0 = const()[name = string("input_493_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189645824)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189647936)))];
+            tensor<fp16, [1, 14, 1024]> input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = string("input_493_cast_fp16")];
+            tensor<int32, [4]> cache_37_begin_0 = const()[name = string("cache_37_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_37_end_0 = const()[name = string("cache_37_end_0"), val = tensor<int32, [4]>([10, 1, 42, 1024])];
+            tensor<bool, [4]> cache_37_end_mask_0 = const()[name = string("cache_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_37_squeeze_mask_0 = const()[name = string("cache_37_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_37_cast_fp16")];
+            tensor<int32, [4]> cache_39_begin_0 = const()[name = string("cache_39_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_39_end_0 = const()[name = string("cache_39_end_0"), val = tensor<int32, [4]>([10, 1, 1024, 8])];
+            tensor<bool, [4]> cache_39_end_mask_0 = const()[name = string("cache_39_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_39_squeeze_mask_0 = const()[name = string("cache_39_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_39_cast_fp16")];
+            tensor<int32, [1]> input_495_axes_0 = const()[name = string("input_495_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189650048)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189652160)))];
+            tensor<fp16, [1, 14, 1024]> input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = string("input_495_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189654272))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192800064))))[name = string("encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_9_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192800256)))];
+            tensor<fp16, [1, 14, 4096]> linear_82_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized, x = input_495_cast_fp16)[name = string("linear_82_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = string("input_499_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192808512))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195954304))))[name = string("encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195954496)))];
+            tensor<fp16, [1, 14, 1024]> linear_83_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized, x = input_499_cast_fp16)[name = string("linear_83_cast_fp16")];
+            fp16 var_2373_to_fp16 = const()[name = string("op_2373_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2374_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2373_to_fp16)[name = string("op_2374_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2374_cast_fp16)[name = string("input_505_cast_fp16")];
+            tensor<int32, [1]> key_19_axes_0 = const()[name = string("key_19_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195956608)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195958720)))];
+            tensor<fp16, [1, 14, 1024]> key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = string("key_19_cast_fp16")];
+            bool input_507_interleave_0 = const()[name = string("input_507_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_507_cast_fp16 = concat(axis = var_69, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = string("input_507_cast_fp16")];
+            tensor<int32, [3]> var_2396_begin_0 = const()[name = string("op_2396_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_2396_end_0 = const()[name = string("op_2396_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2396_end_mask_0 = const()[name = string("op_2396_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_2396_cast_fp16 = slice_by_index(begin = var_2396_begin_0, end = var_2396_end_0, end_mask = var_2396_end_mask_0, x = cache_37_cast_fp16)[name = string("op_2396_cast_fp16")];
+            bool var_2402_interleave_0 = const()[name = string("op_2402_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_2402_cast_fp16 = concat(axis = var_69, interleave = var_2402_interleave_0, values = (var_2396_cast_fp16, key_19_cast_fp16))[name = string("op_2402_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195960832))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196747328))))[name = string("encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196747520)))];
+            tensor<fp16, [1, 14, 1024]> linear_84_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized, x = key_19_cast_fp16)[name = string("linear_84_cast_fp16")];
+            tensor<int32, [4]> var_2407 = const()[name = string("op_2407"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_55_cast_fp16 = reshape(shape = var_2407, x = linear_84_cast_fp16)[name = string("q_55_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196749632))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197536128))))[name = string("encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197536320)))];
+            tensor<fp16, [1, 56, 1024]> linear_85_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_85_cast_fp16")];
+            tensor<int32, [4]> var_2412 = const()[name = string("op_2412"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_37_cast_fp16 = reshape(shape = var_2412, x = linear_85_cast_fp16)[name = string("k_37_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197538432))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198324928))))[name = string("encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198325120)))];
+            tensor<fp16, [1, 56, 1024]> linear_86_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_86_cast_fp16")];
+            tensor<int32, [4]> var_2417 = const()[name = string("op_2417"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_19_cast_fp16 = reshape(shape = var_2417, x = linear_86_cast_fp16)[name = string("v_19_cast_fp16")];
+            tensor<int32, [4]> value_27_perm_0 = const()[name = string("value_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198327232)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2430_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = string("op_2430_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198329344)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2432_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = string("op_2432_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_19_perm_0 = const()[name = string("q_with_bias_v_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_241_transpose_x_0 = const()[name = string("x_241_transpose_x_0"), val = bool(false)];
+            bool x_241_transpose_y_0 = const()[name = string("x_241_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_2434_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198331456))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198445184))))[name = string("op_2434_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2432_cast_fp16)[name = string("transpose_281")];
+            tensor<fp16, [1, 8, 14, 111]> x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = op_2434_to_fp16_quantized)[name = string("x_241_cast_fp16")];
+            tensor<int32, [8]> x_243_pad_0 = const()[name = string("x_243_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_243_mode_0 = const()[name = string("x_243_mode_0"), val = string("constant")];
+            fp16 const_196_to_fp16 = const()[name = string("const_196_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_243_cast_fp16 = pad(constant_val = const_196_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = string("x_243_cast_fp16")];
+            tensor<int32, [4]> var_2442 = const()[name = string("op_2442"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_245_cast_fp16 = reshape(shape = var_2442, x = x_243_cast_fp16)[name = string("x_245_cast_fp16")];
+            tensor<int32, [4]> var_2446_begin_0 = const()[name = string("op_2446_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2446_end_0 = const()[name = string("op_2446_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_2446_end_mask_0 = const()[name = string("op_2446_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_2446_cast_fp16 = slice_by_index(begin = var_2446_begin_0, end = var_2446_end_0, end_mask = var_2446_end_mask_0, x = x_245_cast_fp16)[name = string("op_2446_cast_fp16")];
+            tensor<int32, [4]> var_2447 = const()[name = string("op_2447"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_37_cast_fp16 = reshape(shape = var_2447, x = var_2446_cast_fp16)[name = string("matrix_bd_37_cast_fp16")];
+            bool matrix_ac_19_transpose_x_0 = const()[name = string("matrix_ac_19_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_19_transpose_y_0 = const()[name = string("matrix_ac_19_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37_cast_fp16)[name = string("transpose_279")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_114 = transpose(perm = transpose_114_perm_0, x = var_2430_cast_fp16)[name = string("transpose_280")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_114, y = transpose_115)[name = string("matrix_ac_19_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_39_begin_0 = const()[name = string("matrix_bd_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_39_end_0 = const()[name = string("matrix_bd_39_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_39_end_mask_0 = const()[name = string("matrix_bd_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = string("matrix_bd_39_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2456_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = string("op_2456_cast_fp16")];
+            fp16 _inversed_scores_37_y_0_to_fp16 = const()[name = string("_inversed_scores_37_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_37_cast_fp16 = mul(x = var_2456_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = string("_inversed_scores_37_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_39_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_11)[name = string("scores_39_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2462_cast_fp16 = softmax(axis = var_60, x = scores_39_cast_fp16)[name = string("op_2462_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_509_cast_fp16 = select(a = var_45_to_fp16, b = var_2462_cast_fp16, cond = mask_11)[name = string("input_509_cast_fp16")];
+            bool x_247_transpose_x_0 = const()[name = string("x_247_transpose_x_0"), val = bool(false)];
+            bool x_247_transpose_y_0 = const()[name = string("x_247_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_19_cast_fp16)[name = string("transpose_278")];
+            tensor<fp16, [1, 8, 14, 128]> x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_27_cast_fp16)[name = string("x_247_cast_fp16")];
+            tensor<int32, [4]> var_2466_perm_0 = const()[name = string("op_2466_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2467 = const()[name = string("op_2467"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_2466_cast_fp16 = transpose(perm = var_2466_perm_0, x = x_247_cast_fp16)[name = string("transpose_277")];
+            tensor<fp16, [1, 14, 1024]> input_511_cast_fp16 = reshape(shape = var_2467, x = var_2466_cast_fp16)[name = string("input_511_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198445504))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199232000))))[name = string("encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199232192)))];
+            tensor<fp16, [1, 14, 1024]> linear_88_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized, x = input_511_cast_fp16)[name = string("linear_88_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = string("input_515_cast_fp16")];
+            tensor<int32, [1]> x_251_axes_0 = const()[name = string("x_251_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199234304)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199236416)))];
+            tensor<fp16, [1, 14, 1024]> x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = string("x_251_cast_fp16")];
+            tensor<int32, [3]> input_517_perm_0 = const()[name = string("input_517_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_519_pad_type_0 = const()[name = string("input_519_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_519_strides_0 = const()[name = string("input_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_519_pad_0 = const()[name = string("input_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_519_dilations_0 = const()[name = string("input_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_519_groups_0 = const()[name = string("input_519_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199238528))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201335744))))[name = string("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = string("transpose_276")];
+            tensor<fp16, [1, 2048, 14]> input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_517_cast_fp16)[name = string("input_519_cast_fp16")];
+            int32 x_253_split_num_splits_0 = const()[name = string("x_253_split_num_splits_0"), val = int32(2)];
+            int32 x_253_split_axis_0 = const()[name = string("x_253_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_253_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = string("x_253_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = string("x_253_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = string("x_253_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_521_cast_fp16 = select(a = var_45_to_fp16, b = x_253_cast_fp16, cond = var_576)[name = string("input_521_cast_fp16")];
+            bool new_x_39_interleave_0 = const()[name = string("new_x_39_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_39_cast_fp16 = concat(axis = var_60, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = string("new_x_39_cast_fp16")];
+            tensor<int32, [3]> var_2506_begin_0 = const()[name = string("op_2506_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_2506_end_0 = const()[name = string("op_2506_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_2506_end_mask_0 = const()[name = string("op_2506_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2506_cast_fp16 = slice_by_index(begin = var_2506_begin_0, end = var_2506_end_0, end_mask = var_2506_end_mask_0, x = new_x_39_cast_fp16)[name = string("op_2506_cast_fp16")];
+            string x_255_pad_type_0 = const()[name = string("x_255_pad_type_0"), val = string("valid")];
+            int32 x_255_groups_0 = const()[name = string("x_255_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_255_strides_0 = const()[name = string("x_255_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_255_pad_0 = const()[name = string("x_255_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_255_dilations_0 = const()[name = string("x_255_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201339904))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201349184))))[name = string("encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_39_cast_fp16)[name = string("x_255_cast_fp16")];
+            tensor<int32, [3]> input_523_perm_0 = const()[name = string("input_523_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_257_axes_0 = const()[name = string("x_257_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201351296)))];
+            tensor<fp16, [1024]> encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201353408)))];
+            tensor<fp16, [1, 14, 1024]> input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = string("transpose_275")];
+            tensor<fp16, [1, 14, 1024]> x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = string("x_257_cast_fp16")];
+            tensor<int32, [3]> input_525_perm_0 = const()[name = string("input_525_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = string("transpose_274")];
+            tensor<fp16, [1, 1024, 14]> input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = string("input_527_cast_fp16")];
+            string x_259_pad_type_0 = const()[name = string("x_259_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_259_strides_0 = const()[name = string("x_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_259_pad_0 = const()[name = string("x_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_259_dilations_0 = const()[name = string("x_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_259_groups_0 = const()[name = string("x_259_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201355520))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202404160))))[name = string("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_527_cast_fp16)[name = string("x_259_cast_fp16")];
+            tensor<int32, [3]> input_529_perm_0 = const()[name = string("input_529_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = string("transpose_273")];
+            tensor<fp16, [1, 14, 1024]> input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = string("input_531_cast_fp16")];
+            tensor<int32, [1]> input_533_axes_0 = const()[name = string("input_533_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202406272)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202408384)))];
+            tensor<fp16, [1, 14, 1024]> input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = string("input_533_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202410496))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205556288))))[name = string("encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_9_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205556480)))];
+            tensor<fp16, [1, 14, 4096]> linear_89_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized, x = input_533_cast_fp16)[name = string("linear_89_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = string("input_537_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205564736))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208710528))))[name = string("encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208710720)))];
+            tensor<fp16, [1, 14, 1024]> linear_90_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = string("linear_90_cast_fp16")];
+            fp16 var_2549_to_fp16 = const()[name = string("op_2549_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2550_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2549_to_fp16)[name = string("op_2550_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2550_cast_fp16)[name = string("input_543_cast_fp16")];
+            tensor<int32, [1]> input_545_axes_0 = const()[name = string("input_545_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208712832)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208714944)))];
+            tensor<fp16, [1, 14, 1024]> input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = string("input_545_cast_fp16")];
+            tensor<int32, [4]> cache_41_begin_0 = const()[name = string("cache_41_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_41_end_0 = const()[name = string("cache_41_end_0"), val = tensor<int32, [4]>([11, 1, 42, 1024])];
+            tensor<bool, [4]> cache_41_end_mask_0 = const()[name = string("cache_41_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_41_squeeze_mask_0 = const()[name = string("cache_41_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_41_cast_fp16")];
+            tensor<int32, [4]> cache_43_begin_0 = const()[name = string("cache_43_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_43_end_0 = const()[name = string("cache_43_end_0"), val = tensor<int32, [4]>([11, 1, 1024, 8])];
+            tensor<bool, [4]> cache_43_end_mask_0 = const()[name = string("cache_43_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_43_squeeze_mask_0 = const()[name = string("cache_43_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_43_cast_fp16")];
+            tensor<int32, [1]> input_547_axes_0 = const()[name = string("input_547_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208717056)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208719168)))];
+            tensor<fp16, [1, 14, 1024]> input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = string("input_547_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208721280))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211867072))))[name = string("encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_10_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211867264)))];
+            tensor<fp16, [1, 14, 4096]> linear_91_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = string("linear_91_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = string("input_551_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211875520))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215021312))))[name = string("encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215021504)))];
+            tensor<fp16, [1, 14, 1024]> linear_92_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized, x = input_551_cast_fp16)[name = string("linear_92_cast_fp16")];
+            fp16 var_2586_to_fp16 = const()[name = string("op_2586_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2587_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2586_to_fp16)[name = string("op_2587_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2587_cast_fp16)[name = string("input_557_cast_fp16")];
+            tensor<int32, [1]> key_21_axes_0 = const()[name = string("key_21_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215023616)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215025728)))];
+            tensor<fp16, [1, 14, 1024]> key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = string("key_21_cast_fp16")];
+            bool input_559_interleave_0 = const()[name = string("input_559_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_559_cast_fp16 = concat(axis = var_69, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = string("input_559_cast_fp16")];
+            tensor<int32, [3]> var_2609_begin_0 = const()[name = string("op_2609_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_2609_end_0 = const()[name = string("op_2609_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2609_end_mask_0 = const()[name = string("op_2609_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_2609_cast_fp16 = slice_by_index(begin = var_2609_begin_0, end = var_2609_end_0, end_mask = var_2609_end_mask_0, x = cache_41_cast_fp16)[name = string("op_2609_cast_fp16")];
+            bool var_2615_interleave_0 = const()[name = string("op_2615_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_2615_cast_fp16 = concat(axis = var_69, interleave = var_2615_interleave_0, values = (var_2609_cast_fp16, key_21_cast_fp16))[name = string("op_2615_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215027840))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215814336))))[name = string("encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215814528)))];
+            tensor<fp16, [1, 14, 1024]> linear_93_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized, x = key_21_cast_fp16)[name = string("linear_93_cast_fp16")];
+            tensor<int32, [4]> var_2620 = const()[name = string("op_2620"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_61_cast_fp16 = reshape(shape = var_2620, x = linear_93_cast_fp16)[name = string("q_61_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215816640))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216603136))))[name = string("encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216603328)))];
+            tensor<fp16, [1, 56, 1024]> linear_94_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_94_cast_fp16")];
+            tensor<int32, [4]> var_2625 = const()[name = string("op_2625"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_41_cast_fp16 = reshape(shape = var_2625, x = linear_94_cast_fp16)[name = string("k_41_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216605440))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217391936))))[name = string("encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217392128)))];
+            tensor<fp16, [1, 56, 1024]> linear_95_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_95_cast_fp16")];
+            tensor<int32, [4]> var_2630 = const()[name = string("op_2630"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_21_cast_fp16 = reshape(shape = var_2630, x = linear_95_cast_fp16)[name = string("v_21_cast_fp16")];
+            tensor<int32, [4]> value_29_perm_0 = const()[name = string("value_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217394240)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2643_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = string("op_2643_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217396352)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2645_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = string("op_2645_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_21_perm_0 = const()[name = string("q_with_bias_v_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_267_transpose_x_0 = const()[name = string("x_267_transpose_x_0"), val = bool(false)];
+            bool x_267_transpose_y_0 = const()[name = string("x_267_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_2647_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217398464))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217512192))))[name = string("op_2647_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2645_cast_fp16)[name = string("transpose_272")];
+            tensor<fp16, [1, 8, 14, 111]> x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = op_2647_to_fp16_quantized)[name = string("x_267_cast_fp16")];
+            tensor<int32, [8]> x_269_pad_0 = const()[name = string("x_269_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_269_mode_0 = const()[name = string("x_269_mode_0"), val = string("constant")];
+            fp16 const_209_to_fp16 = const()[name = string("const_209_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_269_cast_fp16 = pad(constant_val = const_209_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = string("x_269_cast_fp16")];
+            tensor<int32, [4]> var_2655 = const()[name = string("op_2655"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_271_cast_fp16 = reshape(shape = var_2655, x = x_269_cast_fp16)[name = string("x_271_cast_fp16")];
+            tensor<int32, [4]> var_2659_begin_0 = const()[name = string("op_2659_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2659_end_0 = const()[name = string("op_2659_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_2659_end_mask_0 = const()[name = string("op_2659_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_2659_cast_fp16 = slice_by_index(begin = var_2659_begin_0, end = var_2659_end_0, end_mask = var_2659_end_mask_0, x = x_271_cast_fp16)[name = string("op_2659_cast_fp16")];
+            tensor<int32, [4]> var_2660 = const()[name = string("op_2660"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_41_cast_fp16 = reshape(shape = var_2660, x = var_2659_cast_fp16)[name = string("matrix_bd_41_cast_fp16")];
+            bool matrix_ac_21_transpose_x_0 = const()[name = string("matrix_ac_21_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_21_transpose_y_0 = const()[name = string("matrix_ac_21_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41_cast_fp16)[name = string("transpose_270")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_116 = transpose(perm = transpose_116_perm_0, x = var_2643_cast_fp16)[name = string("transpose_271")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_116, y = transpose_117)[name = string("matrix_ac_21_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_43_begin_0 = const()[name = string("matrix_bd_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_43_end_0 = const()[name = string("matrix_bd_43_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_43_end_mask_0 = const()[name = string("matrix_bd_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = string("matrix_bd_43_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2669_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = string("op_2669_cast_fp16")];
+            fp16 _inversed_scores_41_y_0_to_fp16 = const()[name = string("_inversed_scores_41_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_41_cast_fp16 = mul(x = var_2669_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = string("_inversed_scores_41_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_43_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_11)[name = string("scores_43_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2675_cast_fp16 = softmax(axis = var_60, x = scores_43_cast_fp16)[name = string("op_2675_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_561_cast_fp16 = select(a = var_45_to_fp16, b = var_2675_cast_fp16, cond = mask_11)[name = string("input_561_cast_fp16")];
+            bool x_273_transpose_x_0 = const()[name = string("x_273_transpose_x_0"), val = bool(false)];
+            bool x_273_transpose_y_0 = const()[name = string("x_273_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_21_cast_fp16)[name = string("transpose_269")];
+            tensor<fp16, [1, 8, 14, 128]> x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_29_cast_fp16)[name = string("x_273_cast_fp16")];
+            tensor<int32, [4]> var_2679_perm_0 = const()[name = string("op_2679_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2680 = const()[name = string("op_2680"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_2679_cast_fp16 = transpose(perm = var_2679_perm_0, x = x_273_cast_fp16)[name = string("transpose_268")];
+            tensor<fp16, [1, 14, 1024]> input_563_cast_fp16 = reshape(shape = var_2680, x = var_2679_cast_fp16)[name = string("input_563_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217512512))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218299008))))[name = string("encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218299200)))];
+            tensor<fp16, [1, 14, 1024]> linear_97_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized, x = input_563_cast_fp16)[name = string("linear_97_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = string("input_567_cast_fp16")];
+            tensor<int32, [1]> x_277_axes_0 = const()[name = string("x_277_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218301312)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218303424)))];
+            tensor<fp16, [1, 14, 1024]> x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = string("x_277_cast_fp16")];
+            tensor<int32, [3]> input_569_perm_0 = const()[name = string("input_569_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_571_pad_type_0 = const()[name = string("input_571_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_571_strides_0 = const()[name = string("input_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_571_pad_0 = const()[name = string("input_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_571_dilations_0 = const()[name = string("input_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_571_groups_0 = const()[name = string("input_571_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218305536))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220402752))))[name = string("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = string("transpose_267")];
+            tensor<fp16, [1, 2048, 14]> input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")];
+            int32 x_279_split_num_splits_0 = const()[name = string("x_279_split_num_splits_0"), val = int32(2)];
+            int32 x_279_split_axis_0 = const()[name = string("x_279_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_279_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = string("x_279_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = string("x_279_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = string("x_279_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_573_cast_fp16 = select(a = var_45_to_fp16, b = x_279_cast_fp16, cond = var_576)[name = string("input_573_cast_fp16")];
+            bool new_x_43_interleave_0 = const()[name = string("new_x_43_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_43_cast_fp16 = concat(axis = var_60, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = string("new_x_43_cast_fp16")];
+            tensor<int32, [3]> var_2719_begin_0 = const()[name = string("op_2719_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_2719_end_0 = const()[name = string("op_2719_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_2719_end_mask_0 = const()[name = string("op_2719_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2719_cast_fp16 = slice_by_index(begin = var_2719_begin_0, end = var_2719_end_0, end_mask = var_2719_end_mask_0, x = new_x_43_cast_fp16)[name = string("op_2719_cast_fp16")];
+            string x_281_pad_type_0 = const()[name = string("x_281_pad_type_0"), val = string("valid")];
+            int32 x_281_groups_0 = const()[name = string("x_281_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_281_strides_0 = const()[name = string("x_281_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_281_pad_0 = const()[name = string("x_281_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_281_dilations_0 = const()[name = string("x_281_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220406912))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220416192))))[name = string("encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_43_cast_fp16)[name = string("x_281_cast_fp16")];
+            tensor<int32, [3]> input_575_perm_0 = const()[name = string("input_575_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_283_axes_0 = const()[name = string("x_283_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220418304)))];
+            tensor<fp16, [1024]> encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220420416)))];
+            tensor<fp16, [1, 14, 1024]> input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = string("transpose_266")];
+            tensor<fp16, [1, 14, 1024]> x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = string("x_283_cast_fp16")];
+            tensor<int32, [3]> input_577_perm_0 = const()[name = string("input_577_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = string("transpose_265")];
+            tensor<fp16, [1, 1024, 14]> input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = string("input_579_cast_fp16")];
+            string x_285_pad_type_0 = const()[name = string("x_285_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_285_strides_0 = const()[name = string("x_285_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_285_pad_0 = const()[name = string("x_285_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_285_dilations_0 = const()[name = string("x_285_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_285_groups_0 = const()[name = string("x_285_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220422528))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221471168))))[name = string("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_579_cast_fp16)[name = string("x_285_cast_fp16")];
+            tensor<int32, [3]> input_581_perm_0 = const()[name = string("input_581_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = string("transpose_264")];
+            tensor<fp16, [1, 14, 1024]> input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = string("input_583_cast_fp16")];
+            tensor<int32, [1]> input_585_axes_0 = const()[name = string("input_585_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221473280)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221475392)))];
+            tensor<fp16, [1, 14, 1024]> input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = string("input_585_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221477504))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224623296))))[name = string("encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_10_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224623488)))];
+            tensor<fp16, [1, 14, 4096]> linear_98_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized, x = input_585_cast_fp16)[name = string("linear_98_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = string("input_589_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224631744))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227777536))))[name = string("encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227777728)))];
+            tensor<fp16, [1, 14, 1024]> linear_99_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized, x = input_589_cast_fp16)[name = string("linear_99_cast_fp16")];
+            fp16 var_2762_to_fp16 = const()[name = string("op_2762_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2763_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2762_to_fp16)[name = string("op_2763_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2763_cast_fp16)[name = string("input_595_cast_fp16")];
+            tensor<int32, [1]> input_597_axes_0 = const()[name = string("input_597_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227779840)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227781952)))];
+            tensor<fp16, [1, 14, 1024]> input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = string("input_597_cast_fp16")];
+            tensor<int32, [4]> cache_45_begin_0 = const()[name = string("cache_45_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_45_end_0 = const()[name = string("cache_45_end_0"), val = tensor<int32, [4]>([12, 1, 42, 1024])];
+            tensor<bool, [4]> cache_45_end_mask_0 = const()[name = string("cache_45_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_45_squeeze_mask_0 = const()[name = string("cache_45_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_45_cast_fp16")];
+            tensor<int32, [4]> cache_47_begin_0 = const()[name = string("cache_47_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_47_end_0 = const()[name = string("cache_47_end_0"), val = tensor<int32, [4]>([12, 1, 1024, 8])];
+            tensor<bool, [4]> cache_47_end_mask_0 = const()[name = string("cache_47_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_47_squeeze_mask_0 = const()[name = string("cache_47_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_47_cast_fp16")];
+            tensor<int32, [1]> input_599_axes_0 = const()[name = string("input_599_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227784064)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227786176)))];
+            tensor<fp16, [1, 14, 1024]> input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = string("input_599_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227788288))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230934080))))[name = string("encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_11_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230934272)))];
+            tensor<fp16, [1, 14, 4096]> linear_100_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized, x = input_599_cast_fp16)[name = string("linear_100_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = string("input_603_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230942528))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234088320))))[name = string("encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234088512)))];
+            tensor<fp16, [1, 14, 1024]> linear_101_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized, x = input_603_cast_fp16)[name = string("linear_101_cast_fp16")];
+            fp16 var_2799_to_fp16 = const()[name = string("op_2799_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2800_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2799_to_fp16)[name = string("op_2800_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2800_cast_fp16)[name = string("input_609_cast_fp16")];
+            tensor<int32, [1]> key_23_axes_0 = const()[name = string("key_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234090624)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234092736)))];
+            tensor<fp16, [1, 14, 1024]> key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = string("key_23_cast_fp16")];
+            bool input_611_interleave_0 = const()[name = string("input_611_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_611_cast_fp16 = concat(axis = var_69, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = string("input_611_cast_fp16")];
+            tensor<int32, [3]> var_2822_begin_0 = const()[name = string("op_2822_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_2822_end_0 = const()[name = string("op_2822_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2822_end_mask_0 = const()[name = string("op_2822_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_2822_cast_fp16 = slice_by_index(begin = var_2822_begin_0, end = var_2822_end_0, end_mask = var_2822_end_mask_0, x = cache_45_cast_fp16)[name = string("op_2822_cast_fp16")];
+            bool var_2828_interleave_0 = const()[name = string("op_2828_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_2828_cast_fp16 = concat(axis = var_69, interleave = var_2828_interleave_0, values = (var_2822_cast_fp16, key_23_cast_fp16))[name = string("op_2828_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234094848))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234881344))))[name = string("encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234881536)))];
+            tensor<fp16, [1, 14, 1024]> linear_102_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized, x = key_23_cast_fp16)[name = string("linear_102_cast_fp16")];
+            tensor<int32, [4]> var_2833 = const()[name = string("op_2833"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_67_cast_fp16 = reshape(shape = var_2833, x = linear_102_cast_fp16)[name = string("q_67_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234883648))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235670144))))[name = string("encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235670336)))];
+            tensor<fp16, [1, 56, 1024]> linear_103_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_103_cast_fp16")];
+            tensor<int32, [4]> var_2838 = const()[name = string("op_2838"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_45_cast_fp16 = reshape(shape = var_2838, x = linear_103_cast_fp16)[name = string("k_45_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235672448))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236458944))))[name = string("encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236459136)))];
+            tensor<fp16, [1, 56, 1024]> linear_104_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_104_cast_fp16")];
+            tensor<int32, [4]> var_2843 = const()[name = string("op_2843"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_23_cast_fp16 = reshape(shape = var_2843, x = linear_104_cast_fp16)[name = string("v_23_cast_fp16")];
+            tensor<int32, [4]> value_31_perm_0 = const()[name = string("value_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236461248)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2856_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = string("op_2856_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236463360)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2858_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = string("op_2858_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_23_perm_0 = const()[name = string("q_with_bias_v_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_293_transpose_x_0 = const()[name = string("x_293_transpose_x_0"), val = bool(false)];
+            bool x_293_transpose_y_0 = const()[name = string("x_293_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_2860_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236465472))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236579200))))[name = string("op_2860_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2858_cast_fp16)[name = string("transpose_263")];
+            tensor<fp16, [1, 8, 14, 111]> x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = op_2860_to_fp16_quantized)[name = string("x_293_cast_fp16")];
+            tensor<int32, [8]> x_295_pad_0 = const()[name = string("x_295_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_295_mode_0 = const()[name = string("x_295_mode_0"), val = string("constant")];
+            fp16 const_222_to_fp16 = const()[name = string("const_222_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_295_cast_fp16 = pad(constant_val = const_222_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = string("x_295_cast_fp16")];
+            tensor<int32, [4]> var_2868 = const()[name = string("op_2868"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_297_cast_fp16 = reshape(shape = var_2868, x = x_295_cast_fp16)[name = string("x_297_cast_fp16")];
+            tensor<int32, [4]> var_2872_begin_0 = const()[name = string("op_2872_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2872_end_0 = const()[name = string("op_2872_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_2872_end_mask_0 = const()[name = string("op_2872_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_2872_cast_fp16 = slice_by_index(begin = var_2872_begin_0, end = var_2872_end_0, end_mask = var_2872_end_mask_0, x = x_297_cast_fp16)[name = string("op_2872_cast_fp16")];
+            tensor<int32, [4]> var_2873 = const()[name = string("op_2873"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_45_cast_fp16 = reshape(shape = var_2873, x = var_2872_cast_fp16)[name = string("matrix_bd_45_cast_fp16")];
+            bool matrix_ac_23_transpose_x_0 = const()[name = string("matrix_ac_23_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_23_transpose_y_0 = const()[name = string("matrix_ac_23_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45_cast_fp16)[name = string("transpose_261")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_118 = transpose(perm = transpose_118_perm_0, x = var_2856_cast_fp16)[name = string("transpose_262")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_118, y = transpose_119)[name = string("matrix_ac_23_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_47_begin_0 = const()[name = string("matrix_bd_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_47_end_0 = const()[name = string("matrix_bd_47_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_47_end_mask_0 = const()[name = string("matrix_bd_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = string("matrix_bd_47_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2882_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = string("op_2882_cast_fp16")];
+            fp16 _inversed_scores_45_y_0_to_fp16 = const()[name = string("_inversed_scores_45_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_45_cast_fp16 = mul(x = var_2882_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = string("_inversed_scores_45_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_47_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_11)[name = string("scores_47_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2888_cast_fp16 = softmax(axis = var_60, x = scores_47_cast_fp16)[name = string("op_2888_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_613_cast_fp16 = select(a = var_45_to_fp16, b = var_2888_cast_fp16, cond = mask_11)[name = string("input_613_cast_fp16")];
+            bool x_299_transpose_x_0 = const()[name = string("x_299_transpose_x_0"), val = bool(false)];
+            bool x_299_transpose_y_0 = const()[name = string("x_299_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_23_cast_fp16)[name = string("transpose_260")];
+            tensor<fp16, [1, 8, 14, 128]> x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_31_cast_fp16)[name = string("x_299_cast_fp16")];
+            tensor<int32, [4]> var_2892_perm_0 = const()[name = string("op_2892_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2893 = const()[name = string("op_2893"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_2892_cast_fp16 = transpose(perm = var_2892_perm_0, x = x_299_cast_fp16)[name = string("transpose_259")];
+            tensor<fp16, [1, 14, 1024]> input_615_cast_fp16 = reshape(shape = var_2893, x = var_2892_cast_fp16)[name = string("input_615_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236579520))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237366016))))[name = string("encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237366208)))];
+            tensor<fp16, [1, 14, 1024]> linear_106_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized, x = input_615_cast_fp16)[name = string("linear_106_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = string("input_619_cast_fp16")];
+            tensor<int32, [1]> x_303_axes_0 = const()[name = string("x_303_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237368320)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237370432)))];
+            tensor<fp16, [1, 14, 1024]> x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = string("x_303_cast_fp16")];
+            tensor<int32, [3]> input_621_perm_0 = const()[name = string("input_621_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_623_pad_type_0 = const()[name = string("input_623_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_623_strides_0 = const()[name = string("input_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_623_pad_0 = const()[name = string("input_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_623_dilations_0 = const()[name = string("input_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_623_groups_0 = const()[name = string("input_623_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237372544))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239469760))))[name = string("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = string("transpose_258")];
+            tensor<fp16, [1, 2048, 14]> input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_621_cast_fp16)[name = string("input_623_cast_fp16")];
+            int32 x_305_split_num_splits_0 = const()[name = string("x_305_split_num_splits_0"), val = int32(2)];
+            int32 x_305_split_axis_0 = const()[name = string("x_305_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_305_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = string("x_305_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = string("x_305_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = string("x_305_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_625_cast_fp16 = select(a = var_45_to_fp16, b = x_305_cast_fp16, cond = var_576)[name = string("input_625_cast_fp16")];
+            bool new_x_47_interleave_0 = const()[name = string("new_x_47_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_47_cast_fp16 = concat(axis = var_60, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = string("new_x_47_cast_fp16")];
+            tensor<int32, [3]> var_2932_begin_0 = const()[name = string("op_2932_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_2932_end_0 = const()[name = string("op_2932_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_2932_end_mask_0 = const()[name = string("op_2932_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2932_cast_fp16 = slice_by_index(begin = var_2932_begin_0, end = var_2932_end_0, end_mask = var_2932_end_mask_0, x = new_x_47_cast_fp16)[name = string("op_2932_cast_fp16")];
+            string x_307_pad_type_0 = const()[name = string("x_307_pad_type_0"), val = string("valid")];
+            int32 x_307_groups_0 = const()[name = string("x_307_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_307_strides_0 = const()[name = string("x_307_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_307_pad_0 = const()[name = string("x_307_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_307_dilations_0 = const()[name = string("x_307_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239473920))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239483200))))[name = string("encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_47_cast_fp16)[name = string("x_307_cast_fp16")];
+            tensor<int32, [3]> input_627_perm_0 = const()[name = string("input_627_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_309_axes_0 = const()[name = string("x_309_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239485312)))];
+            tensor<fp16, [1024]> encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239487424)))];
+            tensor<fp16, [1, 14, 1024]> input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = string("transpose_257")];
+            tensor<fp16, [1, 14, 1024]> x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = string("x_309_cast_fp16")];
+            tensor<int32, [3]> input_629_perm_0 = const()[name = string("input_629_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = string("transpose_256")];
+            tensor<fp16, [1, 1024, 14]> input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = string("input_631_cast_fp16")];
+            string x_311_pad_type_0 = const()[name = string("x_311_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_311_strides_0 = const()[name = string("x_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_311_pad_0 = const()[name = string("x_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_311_dilations_0 = const()[name = string("x_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_311_groups_0 = const()[name = string("x_311_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239489536))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240538176))))[name = string("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_631_cast_fp16)[name = string("x_311_cast_fp16")];
+            tensor<int32, [3]> input_633_perm_0 = const()[name = string("input_633_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = string("transpose_255")];
+            tensor<fp16, [1, 14, 1024]> input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = string("input_635_cast_fp16")];
+            tensor<int32, [1]> input_637_axes_0 = const()[name = string("input_637_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240540288)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240542400)))];
+            tensor<fp16, [1, 14, 1024]> input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = string("input_637_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240544512))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243690304))))[name = string("encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_11_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243690496)))];
+            tensor<fp16, [1, 14, 4096]> linear_107_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized, x = input_637_cast_fp16)[name = string("linear_107_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = string("input_641_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243698752))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246844544))))[name = string("encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246844736)))];
+            tensor<fp16, [1, 14, 1024]> linear_108_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized, x = input_641_cast_fp16)[name = string("linear_108_cast_fp16")];
+            fp16 var_2975_to_fp16 = const()[name = string("op_2975_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2976_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2975_to_fp16)[name = string("op_2976_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2976_cast_fp16)[name = string("input_647_cast_fp16")];
+            tensor<int32, [1]> input_649_axes_0 = const()[name = string("input_649_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246846848)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246848960)))];
+            tensor<fp16, [1, 14, 1024]> input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = string("input_649_cast_fp16")];
+            tensor<int32, [4]> cache_49_begin_0 = const()[name = string("cache_49_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_49_end_0 = const()[name = string("cache_49_end_0"), val = tensor<int32, [4]>([13, 1, 42, 1024])];
+            tensor<bool, [4]> cache_49_end_mask_0 = const()[name = string("cache_49_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_49_squeeze_mask_0 = const()[name = string("cache_49_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_49_cast_fp16")];
+            tensor<int32, [4]> cache_51_begin_0 = const()[name = string("cache_51_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_51_end_0 = const()[name = string("cache_51_end_0"), val = tensor<int32, [4]>([13, 1, 1024, 8])];
+            tensor<bool, [4]> cache_51_end_mask_0 = const()[name = string("cache_51_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_51_squeeze_mask_0 = const()[name = string("cache_51_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_51_cast_fp16")];
+            tensor<int32, [1]> input_651_axes_0 = const()[name = string("input_651_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246851072)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246853184)))];
+            tensor<fp16, [1, 14, 1024]> input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = string("input_651_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246855296))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250001088))))[name = string("encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_12_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250001280)))];
+            tensor<fp16, [1, 14, 4096]> linear_109_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized, x = input_651_cast_fp16)[name = string("linear_109_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("input_655_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250009536))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253155328))))[name = string("encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253155520)))];
+            tensor<fp16, [1, 14, 1024]> linear_110_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized, x = input_655_cast_fp16)[name = string("linear_110_cast_fp16")];
+            fp16 var_3012_to_fp16 = const()[name = string("op_3012_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3013_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_3012_to_fp16)[name = string("op_3013_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_3013_cast_fp16)[name = string("input_661_cast_fp16")];
+            tensor<int32, [1]> key_25_axes_0 = const()[name = string("key_25_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253157632)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253159744)))];
+            tensor<fp16, [1, 14, 1024]> key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = string("key_25_cast_fp16")];
+            bool input_663_interleave_0 = const()[name = string("input_663_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_663_cast_fp16 = concat(axis = var_69, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = string("input_663_cast_fp16")];
+            tensor<int32, [3]> var_3035_begin_0 = const()[name = string("op_3035_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_3035_end_0 = const()[name = string("op_3035_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3035_end_mask_0 = const()[name = string("op_3035_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_3035_cast_fp16 = slice_by_index(begin = var_3035_begin_0, end = var_3035_end_0, end_mask = var_3035_end_mask_0, x = cache_49_cast_fp16)[name = string("op_3035_cast_fp16")];
+            bool var_3041_interleave_0 = const()[name = string("op_3041_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3041_cast_fp16 = concat(axis = var_69, interleave = var_3041_interleave_0, values = (var_3035_cast_fp16, key_25_cast_fp16))[name = string("op_3041_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253161856))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253948352))))[name = string("encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253948544)))];
+            tensor<fp16, [1, 14, 1024]> linear_111_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized, x = key_25_cast_fp16)[name = string("linear_111_cast_fp16")];
+            tensor<int32, [4]> var_3046 = const()[name = string("op_3046"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_73_cast_fp16 = reshape(shape = var_3046, x = linear_111_cast_fp16)[name = string("q_73_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253950656))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254737152))))[name = string("encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254737344)))];
+            tensor<fp16, [1, 56, 1024]> linear_112_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_112_cast_fp16")];
+            tensor<int32, [4]> var_3051 = const()[name = string("op_3051"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_49_cast_fp16 = reshape(shape = var_3051, x = linear_112_cast_fp16)[name = string("k_49_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254739456))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255525952))))[name = string("encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255526144)))];
+            tensor<fp16, [1, 56, 1024]> linear_113_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_113_cast_fp16")];
+            tensor<int32, [4]> var_3056 = const()[name = string("op_3056"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_25_cast_fp16 = reshape(shape = var_3056, x = linear_113_cast_fp16)[name = string("v_25_cast_fp16")];
+            tensor<int32, [4]> value_33_perm_0 = const()[name = string("value_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255528256)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3069_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = string("op_3069_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255530368)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3071_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = string("op_3071_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_25_perm_0 = const()[name = string("q_with_bias_v_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_319_transpose_x_0 = const()[name = string("x_319_transpose_x_0"), val = bool(false)];
+            bool x_319_transpose_y_0 = const()[name = string("x_319_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_3073_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255532480))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255646208))))[name = string("op_3073_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_3071_cast_fp16)[name = string("transpose_254")];
+            tensor<fp16, [1, 8, 14, 111]> x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = op_3073_to_fp16_quantized)[name = string("x_319_cast_fp16")];
+            tensor<int32, [8]> x_321_pad_0 = const()[name = string("x_321_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("constant")];
+            fp16 const_235_to_fp16 = const()[name = string("const_235_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_321_cast_fp16 = pad(constant_val = const_235_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = string("x_321_cast_fp16")];
+            tensor<int32, [4]> var_3081 = const()[name = string("op_3081"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_323_cast_fp16 = reshape(shape = var_3081, x = x_321_cast_fp16)[name = string("x_323_cast_fp16")];
+            tensor<int32, [4]> var_3085_begin_0 = const()[name = string("op_3085_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3085_end_0 = const()[name = string("op_3085_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_3085_end_mask_0 = const()[name = string("op_3085_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_3085_cast_fp16 = slice_by_index(begin = var_3085_begin_0, end = var_3085_end_0, end_mask = var_3085_end_mask_0, x = x_323_cast_fp16)[name = string("op_3085_cast_fp16")];
+            tensor<int32, [4]> var_3086 = const()[name = string("op_3086"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_49_cast_fp16 = reshape(shape = var_3086, x = var_3085_cast_fp16)[name = string("matrix_bd_49_cast_fp16")];
+            bool matrix_ac_25_transpose_x_0 = const()[name = string("matrix_ac_25_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_25_transpose_y_0 = const()[name = string("matrix_ac_25_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_120_perm_0 = const()[name = string("transpose_120_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_121_perm_0 = const()[name = string("transpose_121_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49_cast_fp16)[name = string("transpose_252")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_120 = transpose(perm = transpose_120_perm_0, x = var_3069_cast_fp16)[name = string("transpose_253")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_120, y = transpose_121)[name = string("matrix_ac_25_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_51_begin_0 = const()[name = string("matrix_bd_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_51_end_0 = const()[name = string("matrix_bd_51_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_51_end_mask_0 = const()[name = string("matrix_bd_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = string("matrix_bd_51_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3095_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = string("op_3095_cast_fp16")];
+            fp16 _inversed_scores_49_y_0_to_fp16 = const()[name = string("_inversed_scores_49_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_49_cast_fp16 = mul(x = var_3095_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = string("_inversed_scores_49_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_51_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_11)[name = string("scores_51_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3101_cast_fp16 = softmax(axis = var_60, x = scores_51_cast_fp16)[name = string("op_3101_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_665_cast_fp16 = select(a = var_45_to_fp16, b = var_3101_cast_fp16, cond = mask_11)[name = string("input_665_cast_fp16")];
+            bool x_325_transpose_x_0 = const()[name = string("x_325_transpose_x_0"), val = bool(false)];
+            bool x_325_transpose_y_0 = const()[name = string("x_325_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_25_cast_fp16)[name = string("transpose_251")];
+            tensor<fp16, [1, 8, 14, 128]> x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_33_cast_fp16)[name = string("x_325_cast_fp16")];
+            tensor<int32, [4]> var_3105_perm_0 = const()[name = string("op_3105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3106 = const()[name = string("op_3106"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_3105_cast_fp16 = transpose(perm = var_3105_perm_0, x = x_325_cast_fp16)[name = string("transpose_250")];
+            tensor<fp16, [1, 14, 1024]> input_667_cast_fp16 = reshape(shape = var_3106, x = var_3105_cast_fp16)[name = string("input_667_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255646528))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256433024))))[name = string("encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256433216)))];
+            tensor<fp16, [1, 14, 1024]> linear_115_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized, x = input_667_cast_fp16)[name = string("linear_115_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = string("input_671_cast_fp16")];
+            tensor<int32, [1]> x_329_axes_0 = const()[name = string("x_329_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256435328)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256437440)))];
+            tensor<fp16, [1, 14, 1024]> x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = string("x_329_cast_fp16")];
+            tensor<int32, [3]> input_673_perm_0 = const()[name = string("input_673_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_675_pad_type_0 = const()[name = string("input_675_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_675_strides_0 = const()[name = string("input_675_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_675_pad_0 = const()[name = string("input_675_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_675_dilations_0 = const()[name = string("input_675_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_675_groups_0 = const()[name = string("input_675_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256439552))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258536768))))[name = string("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = string("transpose_249")];
+            tensor<fp16, [1, 2048, 14]> input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_673_cast_fp16)[name = string("input_675_cast_fp16")];
+            int32 x_331_split_num_splits_0 = const()[name = string("x_331_split_num_splits_0"), val = int32(2)];
+            int32 x_331_split_axis_0 = const()[name = string("x_331_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_331_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = string("x_331_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = string("x_331_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = string("x_331_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_677_cast_fp16 = select(a = var_45_to_fp16, b = x_331_cast_fp16, cond = var_576)[name = string("input_677_cast_fp16")];
+            bool new_x_51_interleave_0 = const()[name = string("new_x_51_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_51_cast_fp16 = concat(axis = var_60, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = string("new_x_51_cast_fp16")];
+            tensor<int32, [3]> var_3145_begin_0 = const()[name = string("op_3145_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_3145_end_0 = const()[name = string("op_3145_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_3145_end_mask_0 = const()[name = string("op_3145_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3145_cast_fp16 = slice_by_index(begin = var_3145_begin_0, end = var_3145_end_0, end_mask = var_3145_end_mask_0, x = new_x_51_cast_fp16)[name = string("op_3145_cast_fp16")];
+            string x_333_pad_type_0 = const()[name = string("x_333_pad_type_0"), val = string("valid")];
+            int32 x_333_groups_0 = const()[name = string("x_333_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_333_strides_0 = const()[name = string("x_333_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_333_pad_0 = const()[name = string("x_333_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_333_dilations_0 = const()[name = string("x_333_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258540928))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258550208))))[name = string("encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_51_cast_fp16)[name = string("x_333_cast_fp16")];
+            tensor<int32, [3]> input_679_perm_0 = const()[name = string("input_679_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_335_axes_0 = const()[name = string("x_335_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258552320)))];
+            tensor<fp16, [1024]> encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258554432)))];
+            tensor<fp16, [1, 14, 1024]> input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = string("transpose_248")];
+            tensor<fp16, [1, 14, 1024]> x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = string("x_335_cast_fp16")];
+            tensor<int32, [3]> input_681_perm_0 = const()[name = string("input_681_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = string("transpose_247")];
+            tensor<fp16, [1, 1024, 14]> input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = string("input_683_cast_fp16")];
+            string x_337_pad_type_0 = const()[name = string("x_337_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_337_strides_0 = const()[name = string("x_337_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_337_pad_0 = const()[name = string("x_337_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_337_dilations_0 = const()[name = string("x_337_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_337_groups_0 = const()[name = string("x_337_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258556544))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259605184))))[name = string("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_683_cast_fp16)[name = string("x_337_cast_fp16")];
+            tensor<int32, [3]> input_685_perm_0 = const()[name = string("input_685_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = string("transpose_246")];
+            tensor<fp16, [1, 14, 1024]> input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = string("input_687_cast_fp16")];
+            tensor<int32, [1]> input_689_axes_0 = const()[name = string("input_689_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259607296)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259609408)))];
+            tensor<fp16, [1, 14, 1024]> input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = string("input_689_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259611520))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262757312))))[name = string("encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_12_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262757504)))];
+            tensor<fp16, [1, 14, 4096]> linear_116_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized, x = input_689_cast_fp16)[name = string("linear_116_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("input_693_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262765760))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265911552))))[name = string("encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265911744)))];
+            tensor<fp16, [1, 14, 1024]> linear_117_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized, x = input_693_cast_fp16)[name = string("linear_117_cast_fp16")];
+            fp16 var_3188_to_fp16 = const()[name = string("op_3188_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3189_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3188_to_fp16)[name = string("op_3189_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_3189_cast_fp16)[name = string("input_699_cast_fp16")];
+            tensor<int32, [1]> input_701_axes_0 = const()[name = string("input_701_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265913856)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265915968)))];
+            tensor<fp16, [1, 14, 1024]> input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = string("input_701_cast_fp16")];
+            tensor<int32, [4]> cache_53_begin_0 = const()[name = string("cache_53_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_53_end_0 = const()[name = string("cache_53_end_0"), val = tensor<int32, [4]>([14, 1, 42, 1024])];
+            tensor<bool, [4]> cache_53_end_mask_0 = const()[name = string("cache_53_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_53_squeeze_mask_0 = const()[name = string("cache_53_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_53_cast_fp16")];
+            tensor<int32, [4]> cache_55_begin_0 = const()[name = string("cache_55_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_55_end_0 = const()[name = string("cache_55_end_0"), val = tensor<int32, [4]>([14, 1, 1024, 8])];
+            tensor<bool, [4]> cache_55_end_mask_0 = const()[name = string("cache_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_55_squeeze_mask_0 = const()[name = string("cache_55_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_55_cast_fp16")];
+            tensor<int32, [1]> input_703_axes_0 = const()[name = string("input_703_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265918080)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265920192)))];
+            tensor<fp16, [1, 14, 1024]> input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = string("input_703_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265922304))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269068096))))[name = string("encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_13_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269068288)))];
+            tensor<fp16, [1, 14, 4096]> linear_118_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized, x = input_703_cast_fp16)[name = string("linear_118_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = string("input_707_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269076544))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272222336))))[name = string("encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272222528)))];
+            tensor<fp16, [1, 14, 1024]> linear_119_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized, x = input_707_cast_fp16)[name = string("linear_119_cast_fp16")];
+            fp16 var_3225_to_fp16 = const()[name = string("op_3225_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3226_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3225_to_fp16)[name = string("op_3226_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_3226_cast_fp16)[name = string("input_713_cast_fp16")];
+            tensor<int32, [1]> key_27_axes_0 = const()[name = string("key_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272224640)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272226752)))];
+            tensor<fp16, [1, 14, 1024]> key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = string("key_27_cast_fp16")];
+            bool input_715_interleave_0 = const()[name = string("input_715_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_715_cast_fp16 = concat(axis = var_69, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = string("input_715_cast_fp16")];
+            tensor<int32, [3]> var_3248_begin_0 = const()[name = string("op_3248_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_3248_end_0 = const()[name = string("op_3248_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3248_end_mask_0 = const()[name = string("op_3248_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_3248_cast_fp16 = slice_by_index(begin = var_3248_begin_0, end = var_3248_end_0, end_mask = var_3248_end_mask_0, x = cache_53_cast_fp16)[name = string("op_3248_cast_fp16")];
+            bool var_3254_interleave_0 = const()[name = string("op_3254_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3254_cast_fp16 = concat(axis = var_69, interleave = var_3254_interleave_0, values = (var_3248_cast_fp16, key_27_cast_fp16))[name = string("op_3254_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272228864))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273015360))))[name = string("encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273015552)))];
+            tensor<fp16, [1, 14, 1024]> linear_120_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized, x = key_27_cast_fp16)[name = string("linear_120_cast_fp16")];
+            tensor<int32, [4]> var_3259 = const()[name = string("op_3259"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_79_cast_fp16 = reshape(shape = var_3259, x = linear_120_cast_fp16)[name = string("q_79_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273017664))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273804160))))[name = string("encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273804352)))];
+            tensor<fp16, [1, 56, 1024]> linear_121_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_121_cast_fp16")];
+            tensor<int32, [4]> var_3264 = const()[name = string("op_3264"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_53_cast_fp16 = reshape(shape = var_3264, x = linear_121_cast_fp16)[name = string("k_53_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273806464))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274592960))))[name = string("encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274593152)))];
+            tensor<fp16, [1, 56, 1024]> linear_122_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_122_cast_fp16")];
+            tensor<int32, [4]> var_3269 = const()[name = string("op_3269"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_27_cast_fp16 = reshape(shape = var_3269, x = linear_122_cast_fp16)[name = string("v_27_cast_fp16")];
+            tensor<int32, [4]> value_35_perm_0 = const()[name = string("value_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274595264)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3282_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = string("op_3282_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274597376)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3284_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = string("op_3284_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_27_perm_0 = const()[name = string("q_with_bias_v_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_345_transpose_x_0 = const()[name = string("x_345_transpose_x_0"), val = bool(false)];
+            bool x_345_transpose_y_0 = const()[name = string("x_345_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_3286_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274599488))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274713216))))[name = string("op_3286_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3284_cast_fp16)[name = string("transpose_245")];
+            tensor<fp16, [1, 8, 14, 111]> x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = op_3286_to_fp16_quantized)[name = string("x_345_cast_fp16")];
+            tensor<int32, [8]> x_347_pad_0 = const()[name = string("x_347_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_347_mode_0 = const()[name = string("x_347_mode_0"), val = string("constant")];
+            fp16 const_248_to_fp16 = const()[name = string("const_248_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_347_cast_fp16 = pad(constant_val = const_248_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = string("x_347_cast_fp16")];
+            tensor<int32, [4]> var_3294 = const()[name = string("op_3294"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_349_cast_fp16 = reshape(shape = var_3294, x = x_347_cast_fp16)[name = string("x_349_cast_fp16")];
+            tensor<int32, [4]> var_3298_begin_0 = const()[name = string("op_3298_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3298_end_0 = const()[name = string("op_3298_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_3298_end_mask_0 = const()[name = string("op_3298_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_3298_cast_fp16 = slice_by_index(begin = var_3298_begin_0, end = var_3298_end_0, end_mask = var_3298_end_mask_0, x = x_349_cast_fp16)[name = string("op_3298_cast_fp16")];
+            tensor<int32, [4]> var_3299 = const()[name = string("op_3299"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_53_cast_fp16 = reshape(shape = var_3299, x = var_3298_cast_fp16)[name = string("matrix_bd_53_cast_fp16")];
+            bool matrix_ac_27_transpose_x_0 = const()[name = string("matrix_ac_27_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_27_transpose_y_0 = const()[name = string("matrix_ac_27_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_122_perm_0 = const()[name = string("transpose_122_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_123_perm_0 = const()[name = string("transpose_123_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53_cast_fp16)[name = string("transpose_243")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_122 = transpose(perm = transpose_122_perm_0, x = var_3282_cast_fp16)[name = string("transpose_244")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_122, y = transpose_123)[name = string("matrix_ac_27_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_55_begin_0 = const()[name = string("matrix_bd_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_55_end_0 = const()[name = string("matrix_bd_55_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_55_end_mask_0 = const()[name = string("matrix_bd_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = string("matrix_bd_55_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3308_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = string("op_3308_cast_fp16")];
+            fp16 _inversed_scores_53_y_0_to_fp16 = const()[name = string("_inversed_scores_53_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_53_cast_fp16 = mul(x = var_3308_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = string("_inversed_scores_53_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_55_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_11)[name = string("scores_55_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3314_cast_fp16 = softmax(axis = var_60, x = scores_55_cast_fp16)[name = string("op_3314_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_717_cast_fp16 = select(a = var_45_to_fp16, b = var_3314_cast_fp16, cond = mask_11)[name = string("input_717_cast_fp16")];
+            bool x_351_transpose_x_0 = const()[name = string("x_351_transpose_x_0"), val = bool(false)];
+            bool x_351_transpose_y_0 = const()[name = string("x_351_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = v_27_cast_fp16)[name = string("transpose_242")];
+            tensor<fp16, [1, 8, 14, 128]> x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_35_cast_fp16)[name = string("x_351_cast_fp16")];
+            tensor<int32, [4]> var_3318_perm_0 = const()[name = string("op_3318_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3319 = const()[name = string("op_3319"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_3318_cast_fp16 = transpose(perm = var_3318_perm_0, x = x_351_cast_fp16)[name = string("transpose_241")];
+            tensor<fp16, [1, 14, 1024]> input_719_cast_fp16 = reshape(shape = var_3319, x = var_3318_cast_fp16)[name = string("input_719_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274713536))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275500032))))[name = string("encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275500224)))];
+            tensor<fp16, [1, 14, 1024]> linear_124_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("linear_124_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = string("input_723_cast_fp16")];
+            tensor<int32, [1]> x_355_axes_0 = const()[name = string("x_355_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275502336)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275504448)))];
+            tensor<fp16, [1, 14, 1024]> x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = string("x_355_cast_fp16")];
+            tensor<int32, [3]> input_725_perm_0 = const()[name = string("input_725_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_727_pad_type_0 = const()[name = string("input_727_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_727_strides_0 = const()[name = string("input_727_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_727_pad_0 = const()[name = string("input_727_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_727_dilations_0 = const()[name = string("input_727_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_727_groups_0 = const()[name = string("input_727_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275506560))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277603776))))[name = string("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = string("transpose_240")];
+            tensor<fp16, [1, 2048, 14]> input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_725_cast_fp16)[name = string("input_727_cast_fp16")];
+            int32 x_357_split_num_splits_0 = const()[name = string("x_357_split_num_splits_0"), val = int32(2)];
+            int32 x_357_split_axis_0 = const()[name = string("x_357_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_357_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = string("x_357_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = string("x_357_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = string("x_357_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_729_cast_fp16 = select(a = var_45_to_fp16, b = x_357_cast_fp16, cond = var_576)[name = string("input_729_cast_fp16")];
+            bool new_x_55_interleave_0 = const()[name = string("new_x_55_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_55_cast_fp16 = concat(axis = var_60, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = string("new_x_55_cast_fp16")];
+            tensor<int32, [3]> var_3358_begin_0 = const()[name = string("op_3358_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_3358_end_0 = const()[name = string("op_3358_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_3358_end_mask_0 = const()[name = string("op_3358_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3358_cast_fp16 = slice_by_index(begin = var_3358_begin_0, end = var_3358_end_0, end_mask = var_3358_end_mask_0, x = new_x_55_cast_fp16)[name = string("op_3358_cast_fp16")];
+            string x_359_pad_type_0 = const()[name = string("x_359_pad_type_0"), val = string("valid")];
+            int32 x_359_groups_0 = const()[name = string("x_359_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_359_strides_0 = const()[name = string("x_359_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_359_pad_0 = const()[name = string("x_359_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_359_dilations_0 = const()[name = string("x_359_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277607936))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277617216))))[name = string("encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_55_cast_fp16)[name = string("x_359_cast_fp16")];
+            tensor<int32, [3]> input_731_perm_0 = const()[name = string("input_731_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_361_axes_0 = const()[name = string("x_361_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277619328)))];
+            tensor<fp16, [1024]> encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277621440)))];
+            tensor<fp16, [1, 14, 1024]> input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = string("transpose_239")];
+            tensor<fp16, [1, 14, 1024]> x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = string("x_361_cast_fp16")];
+            tensor<int32, [3]> input_733_perm_0 = const()[name = string("input_733_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = string("transpose_238")];
+            tensor<fp16, [1, 1024, 14]> input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = string("input_735_cast_fp16")];
+            string x_363_pad_type_0 = const()[name = string("x_363_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_363_strides_0 = const()[name = string("x_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_363_pad_0 = const()[name = string("x_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_363_dilations_0 = const()[name = string("x_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_363_groups_0 = const()[name = string("x_363_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277623552))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278672192))))[name = string("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_735_cast_fp16)[name = string("x_363_cast_fp16")];
+            tensor<int32, [3]> input_737_perm_0 = const()[name = string("input_737_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = string("transpose_237")];
+            tensor<fp16, [1, 14, 1024]> input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = string("input_739_cast_fp16")];
+            tensor<int32, [1]> input_741_axes_0 = const()[name = string("input_741_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278674304)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278676416)))];
+            tensor<fp16, [1, 14, 1024]> input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = string("input_741_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278678528))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281824320))))[name = string("encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_13_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281824512)))];
+            tensor<fp16, [1, 14, 4096]> linear_125_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized, x = input_741_cast_fp16)[name = string("linear_125_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = string("input_745_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281832768))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284978560))))[name = string("encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284978752)))];
+            tensor<fp16, [1, 14, 1024]> linear_126_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized, x = input_745_cast_fp16)[name = string("linear_126_cast_fp16")];
+            fp16 var_3401_to_fp16 = const()[name = string("op_3401_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3402_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3401_to_fp16)[name = string("op_3402_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3402_cast_fp16)[name = string("input_751_cast_fp16")];
+            tensor<int32, [1]> input_753_axes_0 = const()[name = string("input_753_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284980864)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284982976)))];
+            tensor<fp16, [1, 14, 1024]> input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = string("input_753_cast_fp16")];
+            tensor<int32, [4]> cache_57_begin_0 = const()[name = string("cache_57_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_57_end_0 = const()[name = string("cache_57_end_0"), val = tensor<int32, [4]>([15, 1, 42, 1024])];
+            tensor<bool, [4]> cache_57_end_mask_0 = const()[name = string("cache_57_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_57_squeeze_mask_0 = const()[name = string("cache_57_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_57_cast_fp16")];
+            tensor<int32, [4]> cache_59_begin_0 = const()[name = string("cache_59_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_59_end_0 = const()[name = string("cache_59_end_0"), val = tensor<int32, [4]>([15, 1, 1024, 8])];
+            tensor<bool, [4]> cache_59_end_mask_0 = const()[name = string("cache_59_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_59_squeeze_mask_0 = const()[name = string("cache_59_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_59_cast_fp16")];
+            tensor<int32, [1]> input_755_axes_0 = const()[name = string("input_755_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284985088)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284987200)))];
+            tensor<fp16, [1, 14, 1024]> input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = string("input_755_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284989312))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288135104))))[name = string("encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_14_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288135296)))];
+            tensor<fp16, [1, 14, 4096]> linear_127_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized, x = input_755_cast_fp16)[name = string("linear_127_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = string("input_759_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288143552))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291289344))))[name = string("encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291289536)))];
+            tensor<fp16, [1, 14, 1024]> linear_128_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized, x = input_759_cast_fp16)[name = string("linear_128_cast_fp16")];
+            fp16 var_3438_to_fp16 = const()[name = string("op_3438_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3439_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3438_to_fp16)[name = string("op_3439_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3439_cast_fp16)[name = string("input_765_cast_fp16")];
+            tensor<int32, [1]> key_29_axes_0 = const()[name = string("key_29_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291291648)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291293760)))];
+            tensor<fp16, [1, 14, 1024]> key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = string("key_29_cast_fp16")];
+            bool input_767_interleave_0 = const()[name = string("input_767_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_767_cast_fp16 = concat(axis = var_69, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = string("input_767_cast_fp16")];
+            tensor<int32, [3]> var_3461_begin_0 = const()[name = string("op_3461_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_3461_end_0 = const()[name = string("op_3461_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3461_end_mask_0 = const()[name = string("op_3461_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_3461_cast_fp16 = slice_by_index(begin = var_3461_begin_0, end = var_3461_end_0, end_mask = var_3461_end_mask_0, x = cache_57_cast_fp16)[name = string("op_3461_cast_fp16")];
+            bool var_3467_interleave_0 = const()[name = string("op_3467_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3467_cast_fp16 = concat(axis = var_69, interleave = var_3467_interleave_0, values = (var_3461_cast_fp16, key_29_cast_fp16))[name = string("op_3467_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291295872))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292082368))))[name = string("encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292082560)))];
+            tensor<fp16, [1, 14, 1024]> linear_129_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized, x = key_29_cast_fp16)[name = string("linear_129_cast_fp16")];
+            tensor<int32, [4]> var_3472 = const()[name = string("op_3472"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_85_cast_fp16 = reshape(shape = var_3472, x = linear_129_cast_fp16)[name = string("q_85_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292084672))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292871168))))[name = string("encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292871360)))];
+            tensor<fp16, [1, 56, 1024]> linear_130_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_130_cast_fp16")];
+            tensor<int32, [4]> var_3477 = const()[name = string("op_3477"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_57_cast_fp16 = reshape(shape = var_3477, x = linear_130_cast_fp16)[name = string("k_57_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292873472))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293659968))))[name = string("encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293660160)))];
+            tensor<fp16, [1, 56, 1024]> linear_131_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_131_cast_fp16")];
+            tensor<int32, [4]> var_3482 = const()[name = string("op_3482"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_29_cast_fp16 = reshape(shape = var_3482, x = linear_131_cast_fp16)[name = string("v_29_cast_fp16")];
+            tensor<int32, [4]> value_37_perm_0 = const()[name = string("value_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293662272)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3495_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = string("op_3495_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293664384)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3497_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = string("op_3497_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_29_perm_0 = const()[name = string("q_with_bias_v_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_371_transpose_x_0 = const()[name = string("x_371_transpose_x_0"), val = bool(false)];
+            bool x_371_transpose_y_0 = const()[name = string("x_371_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_3499_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293666496))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293780224))))[name = string("op_3499_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3497_cast_fp16)[name = string("transpose_236")];
+            tensor<fp16, [1, 8, 14, 111]> x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = op_3499_to_fp16_quantized)[name = string("x_371_cast_fp16")];
+            tensor<int32, [8]> x_373_pad_0 = const()[name = string("x_373_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_373_mode_0 = const()[name = string("x_373_mode_0"), val = string("constant")];
+            fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_373_cast_fp16 = pad(constant_val = const_261_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = string("x_373_cast_fp16")];
+            tensor<int32, [4]> var_3507 = const()[name = string("op_3507"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_375_cast_fp16 = reshape(shape = var_3507, x = x_373_cast_fp16)[name = string("x_375_cast_fp16")];
+            tensor<int32, [4]> var_3511_begin_0 = const()[name = string("op_3511_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3511_end_0 = const()[name = string("op_3511_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_3511_end_mask_0 = const()[name = string("op_3511_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_3511_cast_fp16 = slice_by_index(begin = var_3511_begin_0, end = var_3511_end_0, end_mask = var_3511_end_mask_0, x = x_375_cast_fp16)[name = string("op_3511_cast_fp16")];
+            tensor<int32, [4]> var_3512 = const()[name = string("op_3512"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_57_cast_fp16 = reshape(shape = var_3512, x = var_3511_cast_fp16)[name = string("matrix_bd_57_cast_fp16")];
+            bool matrix_ac_29_transpose_x_0 = const()[name = string("matrix_ac_29_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_29_transpose_y_0 = const()[name = string("matrix_ac_29_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_124_perm_0 = const()[name = string("transpose_124_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_125_perm_0 = const()[name = string("transpose_125_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57_cast_fp16)[name = string("transpose_234")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_124 = transpose(perm = transpose_124_perm_0, x = var_3495_cast_fp16)[name = string("transpose_235")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_124, y = transpose_125)[name = string("matrix_ac_29_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_59_begin_0 = const()[name = string("matrix_bd_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_59_end_0 = const()[name = string("matrix_bd_59_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_59_end_mask_0 = const()[name = string("matrix_bd_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = string("matrix_bd_59_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3521_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = string("op_3521_cast_fp16")];
+            fp16 _inversed_scores_57_y_0_to_fp16 = const()[name = string("_inversed_scores_57_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_57_cast_fp16 = mul(x = var_3521_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = string("_inversed_scores_57_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_59_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_11)[name = string("scores_59_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3527_cast_fp16 = softmax(axis = var_60, x = scores_59_cast_fp16)[name = string("op_3527_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_769_cast_fp16 = select(a = var_45_to_fp16, b = var_3527_cast_fp16, cond = mask_11)[name = string("input_769_cast_fp16")];
+            bool x_377_transpose_x_0 = const()[name = string("x_377_transpose_x_0"), val = bool(false)];
+            bool x_377_transpose_y_0 = const()[name = string("x_377_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = v_29_cast_fp16)[name = string("transpose_233")];
+            tensor<fp16, [1, 8, 14, 128]> x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_37_cast_fp16)[name = string("x_377_cast_fp16")];
+            tensor<int32, [4]> var_3531_perm_0 = const()[name = string("op_3531_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3532 = const()[name = string("op_3532"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_3531_cast_fp16 = transpose(perm = var_3531_perm_0, x = x_377_cast_fp16)[name = string("transpose_232")];
+            tensor<fp16, [1, 14, 1024]> input_771_cast_fp16 = reshape(shape = var_3532, x = var_3531_cast_fp16)[name = string("input_771_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293780544))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294567040))))[name = string("encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294567232)))];
+            tensor<fp16, [1, 14, 1024]> linear_133_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized, x = input_771_cast_fp16)[name = string("linear_133_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = string("input_775_cast_fp16")];
+            tensor<int32, [1]> x_381_axes_0 = const()[name = string("x_381_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294569344)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294571456)))];
+            tensor<fp16, [1, 14, 1024]> x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = string("x_381_cast_fp16")];
+            tensor<int32, [3]> input_777_perm_0 = const()[name = string("input_777_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_779_pad_type_0 = const()[name = string("input_779_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_779_strides_0 = const()[name = string("input_779_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_779_pad_0 = const()[name = string("input_779_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_779_dilations_0 = const()[name = string("input_779_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_779_groups_0 = const()[name = string("input_779_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294573568))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296670784))))[name = string("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = string("transpose_231")];
+            tensor<fp16, [1, 2048, 14]> input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_777_cast_fp16)[name = string("input_779_cast_fp16")];
+            int32 x_383_split_num_splits_0 = const()[name = string("x_383_split_num_splits_0"), val = int32(2)];
+            int32 x_383_split_axis_0 = const()[name = string("x_383_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_383_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = string("x_383_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = string("x_383_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = string("x_383_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_781_cast_fp16 = select(a = var_45_to_fp16, b = x_383_cast_fp16, cond = var_576)[name = string("input_781_cast_fp16")];
+            bool new_x_59_interleave_0 = const()[name = string("new_x_59_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_59_cast_fp16 = concat(axis = var_60, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = string("new_x_59_cast_fp16")];
+            tensor<int32, [3]> var_3571_begin_0 = const()[name = string("op_3571_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_3571_end_0 = const()[name = string("op_3571_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_3571_end_mask_0 = const()[name = string("op_3571_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3571_cast_fp16 = slice_by_index(begin = var_3571_begin_0, end = var_3571_end_0, end_mask = var_3571_end_mask_0, x = new_x_59_cast_fp16)[name = string("op_3571_cast_fp16")];
+            string x_385_pad_type_0 = const()[name = string("x_385_pad_type_0"), val = string("valid")];
+            int32 x_385_groups_0 = const()[name = string("x_385_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_385_strides_0 = const()[name = string("x_385_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_385_pad_0 = const()[name = string("x_385_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_385_dilations_0 = const()[name = string("x_385_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296674944))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296684224))))[name = string("encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_59_cast_fp16)[name = string("x_385_cast_fp16")];
+            tensor<int32, [3]> input_783_perm_0 = const()[name = string("input_783_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_387_axes_0 = const()[name = string("x_387_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296686336)))];
+            tensor<fp16, [1024]> encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296688448)))];
+            tensor<fp16, [1, 14, 1024]> input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = string("transpose_230")];
+            tensor<fp16, [1, 14, 1024]> x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = string("x_387_cast_fp16")];
+            tensor<int32, [3]> input_785_perm_0 = const()[name = string("input_785_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = string("transpose_229")];
+            tensor<fp16, [1, 1024, 14]> input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = string("input_787_cast_fp16")];
+            string x_389_pad_type_0 = const()[name = string("x_389_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_389_strides_0 = const()[name = string("x_389_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_389_pad_0 = const()[name = string("x_389_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_389_dilations_0 = const()[name = string("x_389_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_389_groups_0 = const()[name = string("x_389_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296690560))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297739200))))[name = string("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_787_cast_fp16)[name = string("x_389_cast_fp16")];
+            tensor<int32, [3]> input_789_perm_0 = const()[name = string("input_789_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = string("transpose_228")];
+            tensor<fp16, [1, 14, 1024]> input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = string("input_791_cast_fp16")];
+            tensor<int32, [1]> input_793_axes_0 = const()[name = string("input_793_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297741312)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297743424)))];
+            tensor<fp16, [1, 14, 1024]> input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = string("input_793_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297745536))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300891328))))[name = string("encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_14_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300891520)))];
+            tensor<fp16, [1, 14, 4096]> linear_134_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized, x = input_793_cast_fp16)[name = string("linear_134_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = string("input_797_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300899776))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304045568))))[name = string("encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304045760)))];
+            tensor<fp16, [1, 14, 1024]> linear_135_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized, x = input_797_cast_fp16)[name = string("linear_135_cast_fp16")];
+            fp16 var_3614_to_fp16 = const()[name = string("op_3614_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3615_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3614_to_fp16)[name = string("op_3615_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3615_cast_fp16)[name = string("input_803_cast_fp16")];
+            tensor<int32, [1]> input_805_axes_0 = const()[name = string("input_805_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304047872)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304049984)))];
+            tensor<fp16, [1, 14, 1024]> input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = string("input_805_cast_fp16")];
+            tensor<int32, [4]> cache_61_begin_0 = const()[name = string("cache_61_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_61_end_0 = const()[name = string("cache_61_end_0"), val = tensor<int32, [4]>([16, 1, 42, 1024])];
+            tensor<bool, [4]> cache_61_end_mask_0 = const()[name = string("cache_61_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_61_squeeze_mask_0 = const()[name = string("cache_61_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_61_cast_fp16")];
+            tensor<int32, [4]> cache_63_begin_0 = const()[name = string("cache_63_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_63_end_0 = const()[name = string("cache_63_end_0"), val = tensor<int32, [4]>([16, 1, 1024, 8])];
+            tensor<bool, [4]> cache_63_end_mask_0 = const()[name = string("cache_63_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_63_squeeze_mask_0 = const()[name = string("cache_63_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_63_cast_fp16")];
+            tensor<int32, [1]> input_807_axes_0 = const()[name = string("input_807_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304052096)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304054208)))];
+            tensor<fp16, [1, 14, 1024]> input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = string("input_807_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304056320))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307202112))))[name = string("encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_15_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307202304)))];
+            tensor<fp16, [1, 14, 4096]> linear_136_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized, x = input_807_cast_fp16)[name = string("linear_136_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = string("input_811_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307210560))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310356352))))[name = string("encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310356544)))];
+            tensor<fp16, [1, 14, 1024]> linear_137_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("linear_137_cast_fp16")];
+            fp16 var_3651_to_fp16 = const()[name = string("op_3651_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3652_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3651_to_fp16)[name = string("op_3652_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3652_cast_fp16)[name = string("input_817_cast_fp16")];
+            tensor<int32, [1]> key_31_axes_0 = const()[name = string("key_31_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310358656)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310360768)))];
+            tensor<fp16, [1, 14, 1024]> key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = string("key_31_cast_fp16")];
+            bool input_819_interleave_0 = const()[name = string("input_819_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_819_cast_fp16 = concat(axis = var_69, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = string("input_819_cast_fp16")];
+            tensor<int32, [3]> var_3674_begin_0 = const()[name = string("op_3674_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_3674_end_0 = const()[name = string("op_3674_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3674_end_mask_0 = const()[name = string("op_3674_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_3674_cast_fp16 = slice_by_index(begin = var_3674_begin_0, end = var_3674_end_0, end_mask = var_3674_end_mask_0, x = cache_61_cast_fp16)[name = string("op_3674_cast_fp16")];
+            bool var_3680_interleave_0 = const()[name = string("op_3680_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3680_cast_fp16 = concat(axis = var_69, interleave = var_3680_interleave_0, values = (var_3674_cast_fp16, key_31_cast_fp16))[name = string("op_3680_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310362880))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311149376))))[name = string("encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311149568)))];
+            tensor<fp16, [1, 14, 1024]> linear_138_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized, x = key_31_cast_fp16)[name = string("linear_138_cast_fp16")];
+            tensor<int32, [4]> var_3685 = const()[name = string("op_3685"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_91_cast_fp16 = reshape(shape = var_3685, x = linear_138_cast_fp16)[name = string("q_91_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311151680))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311938176))))[name = string("encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311938368)))];
+            tensor<fp16, [1, 56, 1024]> linear_139_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_139_cast_fp16")];
+            tensor<int32, [4]> var_3690 = const()[name = string("op_3690"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_61_cast_fp16 = reshape(shape = var_3690, x = linear_139_cast_fp16)[name = string("k_61_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311940480))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312726976))))[name = string("encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312727168)))];
+            tensor<fp16, [1, 56, 1024]> linear_140_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_140_cast_fp16")];
+            tensor<int32, [4]> var_3695 = const()[name = string("op_3695"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_31_cast_fp16 = reshape(shape = var_3695, x = linear_140_cast_fp16)[name = string("v_31_cast_fp16")];
+            tensor<int32, [4]> value_39_perm_0 = const()[name = string("value_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312729280)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3708_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = string("op_3708_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312731392)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3710_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = string("op_3710_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_31_perm_0 = const()[name = string("q_with_bias_v_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_397_transpose_x_0 = const()[name = string("x_397_transpose_x_0"), val = bool(false)];
+            bool x_397_transpose_y_0 = const()[name = string("x_397_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_3712_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312733504))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312847232))))[name = string("op_3712_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3710_cast_fp16)[name = string("transpose_227")];
+            tensor<fp16, [1, 8, 14, 111]> x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = op_3712_to_fp16_quantized)[name = string("x_397_cast_fp16")];
+            tensor<int32, [8]> x_399_pad_0 = const()[name = string("x_399_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_399_mode_0 = const()[name = string("x_399_mode_0"), val = string("constant")];
+            fp16 const_274_to_fp16 = const()[name = string("const_274_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_399_cast_fp16 = pad(constant_val = const_274_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = string("x_399_cast_fp16")];
+            tensor<int32, [4]> var_3720 = const()[name = string("op_3720"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_401_cast_fp16 = reshape(shape = var_3720, x = x_399_cast_fp16)[name = string("x_401_cast_fp16")];
+            tensor<int32, [4]> var_3724_begin_0 = const()[name = string("op_3724_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3724_end_0 = const()[name = string("op_3724_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_3724_end_mask_0 = const()[name = string("op_3724_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_3724_cast_fp16 = slice_by_index(begin = var_3724_begin_0, end = var_3724_end_0, end_mask = var_3724_end_mask_0, x = x_401_cast_fp16)[name = string("op_3724_cast_fp16")];
+            tensor<int32, [4]> var_3725 = const()[name = string("op_3725"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_61_cast_fp16 = reshape(shape = var_3725, x = var_3724_cast_fp16)[name = string("matrix_bd_61_cast_fp16")];
+            bool matrix_ac_31_transpose_x_0 = const()[name = string("matrix_ac_31_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_31_transpose_y_0 = const()[name = string("matrix_ac_31_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_126_perm_0 = const()[name = string("transpose_126_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_127_perm_0 = const()[name = string("transpose_127_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61_cast_fp16)[name = string("transpose_225")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_126 = transpose(perm = transpose_126_perm_0, x = var_3708_cast_fp16)[name = string("transpose_226")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_126, y = transpose_127)[name = string("matrix_ac_31_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_63_begin_0 = const()[name = string("matrix_bd_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_63_end_0 = const()[name = string("matrix_bd_63_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_63_end_mask_0 = const()[name = string("matrix_bd_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = string("matrix_bd_63_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3734_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = string("op_3734_cast_fp16")];
+            fp16 _inversed_scores_61_y_0_to_fp16 = const()[name = string("_inversed_scores_61_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_61_cast_fp16 = mul(x = var_3734_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = string("_inversed_scores_61_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_63_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_11)[name = string("scores_63_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3740_cast_fp16 = softmax(axis = var_60, x = scores_63_cast_fp16)[name = string("op_3740_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_821_cast_fp16 = select(a = var_45_to_fp16, b = var_3740_cast_fp16, cond = mask_11)[name = string("input_821_cast_fp16")];
+            bool x_403_transpose_x_0 = const()[name = string("x_403_transpose_x_0"), val = bool(false)];
+            bool x_403_transpose_y_0 = const()[name = string("x_403_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = v_31_cast_fp16)[name = string("transpose_224")];
+            tensor<fp16, [1, 8, 14, 128]> x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_39_cast_fp16)[name = string("x_403_cast_fp16")];
+            tensor<int32, [4]> var_3744_perm_0 = const()[name = string("op_3744_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3745 = const()[name = string("op_3745"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_3744_cast_fp16 = transpose(perm = var_3744_perm_0, x = x_403_cast_fp16)[name = string("transpose_223")];
+            tensor<fp16, [1, 14, 1024]> input_823_cast_fp16 = reshape(shape = var_3745, x = var_3744_cast_fp16)[name = string("input_823_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312847552))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313634048))))[name = string("encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313634240)))];
+            tensor<fp16, [1, 14, 1024]> linear_142_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized, x = input_823_cast_fp16)[name = string("linear_142_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = string("input_827_cast_fp16")];
+            tensor<int32, [1]> x_407_axes_0 = const()[name = string("x_407_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313636352)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313638464)))];
+            tensor<fp16, [1, 14, 1024]> x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = string("x_407_cast_fp16")];
+            tensor<int32, [3]> input_829_perm_0 = const()[name = string("input_829_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_831_pad_type_0 = const()[name = string("input_831_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_831_strides_0 = const()[name = string("input_831_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_831_pad_0 = const()[name = string("input_831_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_831_dilations_0 = const()[name = string("input_831_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_831_groups_0 = const()[name = string("input_831_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313640576))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315737792))))[name = string("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = string("transpose_222")];
+            tensor<fp16, [1, 2048, 14]> input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_829_cast_fp16)[name = string("input_831_cast_fp16")];
+            int32 x_409_split_num_splits_0 = const()[name = string("x_409_split_num_splits_0"), val = int32(2)];
+            int32 x_409_split_axis_0 = const()[name = string("x_409_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_409_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = string("x_409_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = string("x_409_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = string("x_409_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_833_cast_fp16 = select(a = var_45_to_fp16, b = x_409_cast_fp16, cond = var_576)[name = string("input_833_cast_fp16")];
+            bool new_x_63_interleave_0 = const()[name = string("new_x_63_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_63_cast_fp16 = concat(axis = var_60, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = string("new_x_63_cast_fp16")];
+            tensor<int32, [3]> var_3784_begin_0 = const()[name = string("op_3784_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_3784_end_0 = const()[name = string("op_3784_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_3784_end_mask_0 = const()[name = string("op_3784_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3784_cast_fp16 = slice_by_index(begin = var_3784_begin_0, end = var_3784_end_0, end_mask = var_3784_end_mask_0, x = new_x_63_cast_fp16)[name = string("op_3784_cast_fp16")];
+            string x_411_pad_type_0 = const()[name = string("x_411_pad_type_0"), val = string("valid")];
+            int32 x_411_groups_0 = const()[name = string("x_411_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_411_strides_0 = const()[name = string("x_411_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_411_pad_0 = const()[name = string("x_411_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_411_dilations_0 = const()[name = string("x_411_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315741952))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315751232))))[name = string("encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_63_cast_fp16)[name = string("x_411_cast_fp16")];
+            tensor<int32, [3]> input_835_perm_0 = const()[name = string("input_835_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_413_axes_0 = const()[name = string("x_413_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315753344)))];
+            tensor<fp16, [1024]> encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315755456)))];
+            tensor<fp16, [1, 14, 1024]> input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = string("transpose_221")];
+            tensor<fp16, [1, 14, 1024]> x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = string("x_413_cast_fp16")];
+            tensor<int32, [3]> input_837_perm_0 = const()[name = string("input_837_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = string("transpose_220")];
+            tensor<fp16, [1, 1024, 14]> input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = string("input_839_cast_fp16")];
+            string x_415_pad_type_0 = const()[name = string("x_415_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_415_strides_0 = const()[name = string("x_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_415_pad_0 = const()[name = string("x_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_415_dilations_0 = const()[name = string("x_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_415_groups_0 = const()[name = string("x_415_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315757568))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316806208))))[name = string("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_839_cast_fp16)[name = string("x_415_cast_fp16")];
+            tensor<int32, [3]> input_841_perm_0 = const()[name = string("input_841_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = string("transpose_219")];
+            tensor<fp16, [1, 14, 1024]> input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = string("input_843_cast_fp16")];
+            tensor<int32, [1]> input_845_axes_0 = const()[name = string("input_845_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316808320)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316810432)))];
+            tensor<fp16, [1, 14, 1024]> input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = string("input_845_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316812544))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319958336))))[name = string("encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_15_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319958528)))];
+            tensor<fp16, [1, 14, 4096]> linear_143_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized, x = input_845_cast_fp16)[name = string("linear_143_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = string("input_849_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319966784))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323112576))))[name = string("encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323112768)))];
+            tensor<fp16, [1, 14, 1024]> linear_144_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized, x = input_849_cast_fp16)[name = string("linear_144_cast_fp16")];
+            fp16 var_3827_to_fp16 = const()[name = string("op_3827_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3828_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3827_to_fp16)[name = string("op_3828_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3828_cast_fp16)[name = string("input_855_cast_fp16")];
+            tensor<int32, [1]> input_857_axes_0 = const()[name = string("input_857_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323114880)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323116992)))];
+            tensor<fp16, [1, 14, 1024]> input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = string("input_857_cast_fp16")];
+            tensor<int32, [4]> cache_65_begin_0 = const()[name = string("cache_65_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_65_end_0 = const()[name = string("cache_65_end_0"), val = tensor<int32, [4]>([17, 1, 42, 1024])];
+            tensor<bool, [4]> cache_65_end_mask_0 = const()[name = string("cache_65_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_65_squeeze_mask_0 = const()[name = string("cache_65_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_65_cast_fp16")];
+            tensor<int32, [4]> cache_67_begin_0 = const()[name = string("cache_67_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_67_end_0 = const()[name = string("cache_67_end_0"), val = tensor<int32, [4]>([17, 1, 1024, 8])];
+            tensor<bool, [4]> cache_67_end_mask_0 = const()[name = string("cache_67_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_67_squeeze_mask_0 = const()[name = string("cache_67_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_67_cast_fp16 = slice_by_index(begin = cache_67_begin_0, end = cache_67_end_0, end_mask = cache_67_end_mask_0, squeeze_mask = cache_67_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_67_cast_fp16")];
+            tensor<int32, [1]> input_859_axes_0 = const()[name = string("input_859_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323119104)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323121216)))];
+            tensor<fp16, [1, 14, 1024]> input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = string("input_859_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323123328))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326269120))))[name = string("encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_16_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326269312)))];
+            tensor<fp16, [1, 14, 4096]> linear_145_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized, x = input_859_cast_fp16)[name = string("linear_145_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = string("input_863_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326277568))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329423360))))[name = string("encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329423552)))];
+            tensor<fp16, [1, 14, 1024]> linear_146_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized, x = input_863_cast_fp16)[name = string("linear_146_cast_fp16")];
+            fp16 var_3864_to_fp16 = const()[name = string("op_3864_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3865_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3864_to_fp16)[name = string("op_3865_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3865_cast_fp16)[name = string("input_869_cast_fp16")];
+            tensor<int32, [1]> key_33_axes_0 = const()[name = string("key_33_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329425664)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329427776)))];
+            tensor<fp16, [1, 14, 1024]> key_33_cast_fp16 = layer_norm(axes = key_33_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = string("key_33_cast_fp16")];
+            bool input_871_interleave_0 = const()[name = string("input_871_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_871_cast_fp16 = concat(axis = var_69, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_33_cast_fp16))[name = string("input_871_cast_fp16")];
+            tensor<int32, [3]> var_3887_begin_0 = const()[name = string("op_3887_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_3887_end_0 = const()[name = string("op_3887_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3887_end_mask_0 = const()[name = string("op_3887_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_3887_cast_fp16 = slice_by_index(begin = var_3887_begin_0, end = var_3887_end_0, end_mask = var_3887_end_mask_0, x = cache_65_cast_fp16)[name = string("op_3887_cast_fp16")];
+            bool var_3893_interleave_0 = const()[name = string("op_3893_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3893_cast_fp16 = concat(axis = var_69, interleave = var_3893_interleave_0, values = (var_3887_cast_fp16, key_33_cast_fp16))[name = string("op_3893_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329429888))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330216384))))[name = string("encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330216576)))];
+            tensor<fp16, [1, 14, 1024]> linear_147_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized, x = key_33_cast_fp16)[name = string("linear_147_cast_fp16")];
+            tensor<int32, [4]> var_3898 = const()[name = string("op_3898"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_97_cast_fp16 = reshape(shape = var_3898, x = linear_147_cast_fp16)[name = string("q_97_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330218688))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331005184))))[name = string("encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331005376)))];
+            tensor<fp16, [1, 56, 1024]> linear_148_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_148_cast_fp16")];
+            tensor<int32, [4]> var_3903 = const()[name = string("op_3903"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_65_cast_fp16 = reshape(shape = var_3903, x = linear_148_cast_fp16)[name = string("k_65_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331007488))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331793984))))[name = string("encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331794176)))];
+            tensor<fp16, [1, 56, 1024]> linear_149_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_149_cast_fp16")];
+            tensor<int32, [4]> var_3908 = const()[name = string("op_3908"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_33_cast_fp16 = reshape(shape = var_3908, x = linear_149_cast_fp16)[name = string("v_33_cast_fp16")];
+            tensor<int32, [4]> value_41_perm_0 = const()[name = string("value_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331796288)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3921_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = string("op_3921_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331798400)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3923_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = string("op_3923_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_33_perm_0 = const()[name = string("q_with_bias_v_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_423_transpose_x_0 = const()[name = string("x_423_transpose_x_0"), val = bool(false)];
+            bool x_423_transpose_y_0 = const()[name = string("x_423_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_3925_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331800512))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331914240))))[name = string("op_3925_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_33_cast_fp16 = transpose(perm = q_with_bias_v_33_perm_0, x = var_3923_cast_fp16)[name = string("transpose_218")];
+            tensor<fp16, [1, 8, 14, 111]> x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_33_cast_fp16, y = op_3925_to_fp16_quantized)[name = string("x_423_cast_fp16")];
+            tensor<int32, [8]> x_425_pad_0 = const()[name = string("x_425_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_425_mode_0 = const()[name = string("x_425_mode_0"), val = string("constant")];
+            fp16 const_287_to_fp16 = const()[name = string("const_287_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_425_cast_fp16 = pad(constant_val = const_287_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = string("x_425_cast_fp16")];
+            tensor<int32, [4]> var_3933 = const()[name = string("op_3933"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_427_cast_fp16 = reshape(shape = var_3933, x = x_425_cast_fp16)[name = string("x_427_cast_fp16")];
+            tensor<int32, [4]> var_3937_begin_0 = const()[name = string("op_3937_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3937_end_0 = const()[name = string("op_3937_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_3937_end_mask_0 = const()[name = string("op_3937_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_3937_cast_fp16 = slice_by_index(begin = var_3937_begin_0, end = var_3937_end_0, end_mask = var_3937_end_mask_0, x = x_427_cast_fp16)[name = string("op_3937_cast_fp16")];
+            tensor<int32, [4]> var_3938 = const()[name = string("op_3938"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_65_cast_fp16 = reshape(shape = var_3938, x = var_3937_cast_fp16)[name = string("matrix_bd_65_cast_fp16")];
+            bool matrix_ac_33_transpose_x_0 = const()[name = string("matrix_ac_33_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_33_transpose_y_0 = const()[name = string("matrix_ac_33_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_128_perm_0 = const()[name = string("transpose_128_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_129_perm_0 = const()[name = string("transpose_129_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65_cast_fp16)[name = string("transpose_216")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_128 = transpose(perm = transpose_128_perm_0, x = var_3921_cast_fp16)[name = string("transpose_217")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_33_cast_fp16 = matmul(transpose_x = matrix_ac_33_transpose_x_0, transpose_y = matrix_ac_33_transpose_y_0, x = transpose_128, y = transpose_129)[name = string("matrix_ac_33_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_67_begin_0 = const()[name = string("matrix_bd_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_67_end_0 = const()[name = string("matrix_bd_67_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_67_end_mask_0 = const()[name = string("matrix_bd_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_67_cast_fp16 = slice_by_index(begin = matrix_bd_67_begin_0, end = matrix_bd_67_end_0, end_mask = matrix_bd_67_end_mask_0, x = matrix_bd_65_cast_fp16)[name = string("matrix_bd_67_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3947_cast_fp16 = add(x = matrix_ac_33_cast_fp16, y = matrix_bd_67_cast_fp16)[name = string("op_3947_cast_fp16")];
+            fp16 _inversed_scores_65_y_0_to_fp16 = const()[name = string("_inversed_scores_65_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_65_cast_fp16 = mul(x = var_3947_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = string("_inversed_scores_65_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_67_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_11)[name = string("scores_67_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3953_cast_fp16 = softmax(axis = var_60, x = scores_67_cast_fp16)[name = string("op_3953_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_873_cast_fp16 = select(a = var_45_to_fp16, b = var_3953_cast_fp16, cond = mask_11)[name = string("input_873_cast_fp16")];
+            bool x_429_transpose_x_0 = const()[name = string("x_429_transpose_x_0"), val = bool(false)];
+            bool x_429_transpose_y_0 = const()[name = string("x_429_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = v_33_cast_fp16)[name = string("transpose_215")];
+            tensor<fp16, [1, 8, 14, 128]> x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_41_cast_fp16)[name = string("x_429_cast_fp16")];
+            tensor<int32, [4]> var_3957_perm_0 = const()[name = string("op_3957_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3958 = const()[name = string("op_3958"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_3957_cast_fp16 = transpose(perm = var_3957_perm_0, x = x_429_cast_fp16)[name = string("transpose_214")];
+            tensor<fp16, [1, 14, 1024]> input_875_cast_fp16 = reshape(shape = var_3958, x = var_3957_cast_fp16)[name = string("input_875_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331914560))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332701056))))[name = string("encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332701248)))];
+            tensor<fp16, [1, 14, 1024]> linear_151_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized, x = input_875_cast_fp16)[name = string("linear_151_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = string("input_879_cast_fp16")];
+            tensor<int32, [1]> x_433_axes_0 = const()[name = string("x_433_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332703360)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332705472)))];
+            tensor<fp16, [1, 14, 1024]> x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = string("x_433_cast_fp16")];
+            tensor<int32, [3]> input_881_perm_0 = const()[name = string("input_881_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_883_pad_type_0 = const()[name = string("input_883_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_883_strides_0 = const()[name = string("input_883_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_883_pad_0 = const()[name = string("input_883_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_883_dilations_0 = const()[name = string("input_883_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_883_groups_0 = const()[name = string("input_883_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332707584))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334804800))))[name = string("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = string("transpose_213")];
+            tensor<fp16, [1, 2048, 14]> input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_881_cast_fp16)[name = string("input_883_cast_fp16")];
+            int32 x_435_split_num_splits_0 = const()[name = string("x_435_split_num_splits_0"), val = int32(2)];
+            int32 x_435_split_axis_0 = const()[name = string("x_435_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_435_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = string("x_435_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = string("x_435_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = string("x_435_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_885_cast_fp16 = select(a = var_45_to_fp16, b = x_435_cast_fp16, cond = var_576)[name = string("input_885_cast_fp16")];
+            bool new_x_67_interleave_0 = const()[name = string("new_x_67_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_67_cast_fp16 = concat(axis = var_60, interleave = new_x_67_interleave_0, values = (cache_67_cast_fp16, input_885_cast_fp16))[name = string("new_x_67_cast_fp16")];
+            tensor<int32, [3]> var_3997_begin_0 = const()[name = string("op_3997_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_3997_end_0 = const()[name = string("op_3997_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_3997_end_mask_0 = const()[name = string("op_3997_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3997_cast_fp16 = slice_by_index(begin = var_3997_begin_0, end = var_3997_end_0, end_mask = var_3997_end_mask_0, x = new_x_67_cast_fp16)[name = string("op_3997_cast_fp16")];
+            string x_437_pad_type_0 = const()[name = string("x_437_pad_type_0"), val = string("valid")];
+            int32 x_437_groups_0 = const()[name = string("x_437_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_437_strides_0 = const()[name = string("x_437_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_437_pad_0 = const()[name = string("x_437_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_437_dilations_0 = const()[name = string("x_437_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334808960))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334818240))))[name = string("encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_67_cast_fp16)[name = string("x_437_cast_fp16")];
+            tensor<int32, [3]> input_887_perm_0 = const()[name = string("input_887_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_439_axes_0 = const()[name = string("x_439_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334820352)))];
+            tensor<fp16, [1024]> encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334822464)))];
+            tensor<fp16, [1, 14, 1024]> input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = string("transpose_212")];
+            tensor<fp16, [1, 14, 1024]> x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = string("x_439_cast_fp16")];
+            tensor<int32, [3]> input_889_perm_0 = const()[name = string("input_889_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = string("transpose_211")];
+            tensor<fp16, [1, 1024, 14]> input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = string("input_891_cast_fp16")];
+            string x_441_pad_type_0 = const()[name = string("x_441_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_441_strides_0 = const()[name = string("x_441_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_441_pad_0 = const()[name = string("x_441_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_441_dilations_0 = const()[name = string("x_441_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_441_groups_0 = const()[name = string("x_441_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334824576))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335873216))))[name = string("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_891_cast_fp16)[name = string("x_441_cast_fp16")];
+            tensor<int32, [3]> input_893_perm_0 = const()[name = string("input_893_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = string("transpose_210")];
+            tensor<fp16, [1, 14, 1024]> input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = string("input_895_cast_fp16")];
+            tensor<int32, [1]> input_897_axes_0 = const()[name = string("input_897_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335875328)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335877440)))];
+            tensor<fp16, [1, 14, 1024]> input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = string("input_897_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335879552))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339025344))))[name = string("encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_16_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339025536)))];
+            tensor<fp16, [1, 14, 4096]> linear_152_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized, x = input_897_cast_fp16)[name = string("linear_152_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = string("input_901_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339033792))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342179584))))[name = string("encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342179776)))];
+            tensor<fp16, [1, 14, 1024]> linear_153_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized, x = input_901_cast_fp16)[name = string("linear_153_cast_fp16")];
+            fp16 var_4040_to_fp16 = const()[name = string("op_4040_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4041_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_4040_to_fp16)[name = string("op_4041_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_907_cast_fp16 = add(x = input_895_cast_fp16, y = var_4041_cast_fp16)[name = string("input_907_cast_fp16")];
+            tensor<int32, [1]> input_909_axes_0 = const()[name = string("input_909_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342181888)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342184000)))];
+            tensor<fp16, [1, 14, 1024]> input_909_cast_fp16 = layer_norm(axes = input_909_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_907_cast_fp16)[name = string("input_909_cast_fp16")];
+            tensor<int32, [4]> cache_69_begin_0 = const()[name = string("cache_69_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> cache_69_end_0 = const()[name = string("cache_69_end_0"), val = tensor<int32, [4]>([18, 1, 42, 1024])];
+            tensor<bool, [4]> cache_69_end_mask_0 = const()[name = string("cache_69_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_69_squeeze_mask_0 = const()[name = string("cache_69_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_69_cast_fp16 = slice_by_index(begin = cache_69_begin_0, end = cache_69_end_0, end_mask = cache_69_end_mask_0, squeeze_mask = cache_69_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_69_cast_fp16")];
+            tensor<int32, [4]> cache_71_begin_0 = const()[name = string("cache_71_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> cache_71_end_0 = const()[name = string("cache_71_end_0"), val = tensor<int32, [4]>([18, 1, 1024, 8])];
+            tensor<bool, [4]> cache_71_end_mask_0 = const()[name = string("cache_71_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_71_squeeze_mask_0 = const()[name = string("cache_71_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_71_cast_fp16 = slice_by_index(begin = cache_71_begin_0, end = cache_71_end_0, end_mask = cache_71_end_mask_0, squeeze_mask = cache_71_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_71_cast_fp16")];
+            tensor<int32, [1]> input_911_axes_0 = const()[name = string("input_911_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342186112)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342188224)))];
+            tensor<fp16, [1, 14, 1024]> input_911_cast_fp16 = layer_norm(axes = input_911_axes_0, beta = encoder_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_feed_forward1_weight_to_fp16, x = input_909_cast_fp16)[name = string("input_911_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342190336))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345336128))))[name = string("encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_17_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345336320)))];
+            tensor<fp16, [1, 14, 4096]> linear_154_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized, x = input_911_cast_fp16)[name = string("linear_154_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_915_cast_fp16 = silu(x = linear_154_cast_fp16)[name = string("input_915_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345344576))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348490368))))[name = string("encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348490560)))];
+            tensor<fp16, [1, 14, 1024]> linear_155_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized, x = input_915_cast_fp16)[name = string("linear_155_cast_fp16")];
+            fp16 var_4077_to_fp16 = const()[name = string("op_4077_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4078_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_4077_to_fp16)[name = string("op_4078_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_921_cast_fp16 = add(x = input_909_cast_fp16, y = var_4078_cast_fp16)[name = string("input_921_cast_fp16")];
+            tensor<int32, [1]> key_35_axes_0 = const()[name = string("key_35_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348492672)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348494784)))];
+            tensor<fp16, [1, 14, 1024]> key_35_cast_fp16 = layer_norm(axes = key_35_axes_0, beta = encoder_layers_17_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_self_att_weight_to_fp16, x = input_921_cast_fp16)[name = string("key_35_cast_fp16")];
+            bool input_923_interleave_0 = const()[name = string("input_923_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_923_cast_fp16 = concat(axis = var_69, interleave = input_923_interleave_0, values = (cache_69_cast_fp16, key_35_cast_fp16))[name = string("input_923_cast_fp16")];
+            tensor<int32, [3]> var_4100_begin_0 = const()[name = string("op_4100_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_4100_end_0 = const()[name = string("op_4100_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4100_end_mask_0 = const()[name = string("op_4100_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_4100_cast_fp16 = slice_by_index(begin = var_4100_begin_0, end = var_4100_end_0, end_mask = var_4100_end_mask_0, x = cache_69_cast_fp16)[name = string("op_4100_cast_fp16")];
+            bool var_4106_interleave_0 = const()[name = string("op_4106_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4106_cast_fp16 = concat(axis = var_69, interleave = var_4106_interleave_0, values = (var_4100_cast_fp16, key_35_cast_fp16))[name = string("op_4106_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348496896))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349283392))))[name = string("encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349283584)))];
+            tensor<fp16, [1, 14, 1024]> linear_156_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized, x = key_35_cast_fp16)[name = string("linear_156_cast_fp16")];
+            tensor<int32, [4]> var_4111 = const()[name = string("op_4111"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_103_cast_fp16 = reshape(shape = var_4111, x = linear_156_cast_fp16)[name = string("q_103_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349285696))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350072192))))[name = string("encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350072384)))];
+            tensor<fp16, [1, 56, 1024]> linear_157_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_157_cast_fp16")];
+            tensor<int32, [4]> var_4116 = const()[name = string("op_4116"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_69_cast_fp16 = reshape(shape = var_4116, x = linear_157_cast_fp16)[name = string("k_69_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350074496))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350860992))))[name = string("encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350861184)))];
+            tensor<fp16, [1, 56, 1024]> linear_158_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_158_cast_fp16")];
+            tensor<int32, [4]> var_4121 = const()[name = string("op_4121"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_35_cast_fp16 = reshape(shape = var_4121, x = linear_158_cast_fp16)[name = string("v_35_cast_fp16")];
+            tensor<int32, [4]> value_43_perm_0 = const()[name = string("value_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_17_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350863296)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4134_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_u_to_fp16)[name = string("op_4134_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_17_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350865408)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4136_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_v_to_fp16)[name = string("op_4136_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_35_perm_0 = const()[name = string("q_with_bias_v_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_449_transpose_x_0 = const()[name = string("x_449_transpose_x_0"), val = bool(false)];
+            bool x_449_transpose_y_0 = const()[name = string("x_449_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_4138_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350867520))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350981248))))[name = string("op_4138_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_35_cast_fp16 = transpose(perm = q_with_bias_v_35_perm_0, x = var_4136_cast_fp16)[name = string("transpose_209")];
+            tensor<fp16, [1, 8, 14, 111]> x_449_cast_fp16 = matmul(transpose_x = x_449_transpose_x_0, transpose_y = x_449_transpose_y_0, x = q_with_bias_v_35_cast_fp16, y = op_4138_to_fp16_quantized)[name = string("x_449_cast_fp16")];
+            tensor<int32, [8]> x_451_pad_0 = const()[name = string("x_451_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_451_mode_0 = const()[name = string("x_451_mode_0"), val = string("constant")];
+            fp16 const_300_to_fp16 = const()[name = string("const_300_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_451_cast_fp16 = pad(constant_val = const_300_to_fp16, mode = x_451_mode_0, pad = x_451_pad_0, x = x_449_cast_fp16)[name = string("x_451_cast_fp16")];
+            tensor<int32, [4]> var_4146 = const()[name = string("op_4146"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_453_cast_fp16 = reshape(shape = var_4146, x = x_451_cast_fp16)[name = string("x_453_cast_fp16")];
+            tensor<int32, [4]> var_4150_begin_0 = const()[name = string("op_4150_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4150_end_0 = const()[name = string("op_4150_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_4150_end_mask_0 = const()[name = string("op_4150_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_4150_cast_fp16 = slice_by_index(begin = var_4150_begin_0, end = var_4150_end_0, end_mask = var_4150_end_mask_0, x = x_453_cast_fp16)[name = string("op_4150_cast_fp16")];
+            tensor<int32, [4]> var_4151 = const()[name = string("op_4151"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_69_cast_fp16 = reshape(shape = var_4151, x = var_4150_cast_fp16)[name = string("matrix_bd_69_cast_fp16")];
+            bool matrix_ac_35_transpose_x_0 = const()[name = string("matrix_ac_35_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_35_transpose_y_0 = const()[name = string("matrix_ac_35_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_130_perm_0 = const()[name = string("transpose_130_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_131_perm_0 = const()[name = string("transpose_131_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69_cast_fp16)[name = string("transpose_207")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_130 = transpose(perm = transpose_130_perm_0, x = var_4134_cast_fp16)[name = string("transpose_208")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_35_cast_fp16 = matmul(transpose_x = matrix_ac_35_transpose_x_0, transpose_y = matrix_ac_35_transpose_y_0, x = transpose_130, y = transpose_131)[name = string("matrix_ac_35_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_71_begin_0 = const()[name = string("matrix_bd_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_71_end_0 = const()[name = string("matrix_bd_71_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_71_end_mask_0 = const()[name = string("matrix_bd_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_71_cast_fp16 = slice_by_index(begin = matrix_bd_71_begin_0, end = matrix_bd_71_end_0, end_mask = matrix_bd_71_end_mask_0, x = matrix_bd_69_cast_fp16)[name = string("matrix_bd_71_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_4160_cast_fp16 = add(x = matrix_ac_35_cast_fp16, y = matrix_bd_71_cast_fp16)[name = string("op_4160_cast_fp16")];
+            fp16 _inversed_scores_69_y_0_to_fp16 = const()[name = string("_inversed_scores_69_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_69_cast_fp16 = mul(x = var_4160_cast_fp16, y = _inversed_scores_69_y_0_to_fp16)[name = string("_inversed_scores_69_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_71_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_69_cast_fp16, cond = mask_11)[name = string("scores_71_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_4166_cast_fp16 = softmax(axis = var_60, x = scores_71_cast_fp16)[name = string("op_4166_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_925_cast_fp16 = select(a = var_45_to_fp16, b = var_4166_cast_fp16, cond = mask_11)[name = string("input_925_cast_fp16")];
+            bool x_455_transpose_x_0 = const()[name = string("x_455_transpose_x_0"), val = bool(false)];
+            bool x_455_transpose_y_0 = const()[name = string("x_455_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = v_35_cast_fp16)[name = string("transpose_206")];
+            tensor<fp16, [1, 8, 14, 128]> x_455_cast_fp16 = matmul(transpose_x = x_455_transpose_x_0, transpose_y = x_455_transpose_y_0, x = input_925_cast_fp16, y = value_43_cast_fp16)[name = string("x_455_cast_fp16")];
+            tensor<int32, [4]> var_4170_perm_0 = const()[name = string("op_4170_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4171 = const()[name = string("op_4171"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_4170_cast_fp16 = transpose(perm = var_4170_perm_0, x = x_455_cast_fp16)[name = string("transpose_205")];
+            tensor<fp16, [1, 14, 1024]> input_927_cast_fp16 = reshape(shape = var_4171, x = var_4170_cast_fp16)[name = string("input_927_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350981568))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351768064))))[name = string("encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351768256)))];
+            tensor<fp16, [1, 14, 1024]> linear_160_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized, x = input_927_cast_fp16)[name = string("linear_160_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_931_cast_fp16 = add(x = input_921_cast_fp16, y = linear_160_cast_fp16)[name = string("input_931_cast_fp16")];
+            tensor<int32, [1]> x_459_axes_0 = const()[name = string("x_459_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351770368)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351772480)))];
+            tensor<fp16, [1, 14, 1024]> x_459_cast_fp16 = layer_norm(axes = x_459_axes_0, beta = encoder_layers_17_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_conv_weight_to_fp16, x = input_931_cast_fp16)[name = string("x_459_cast_fp16")];
+            tensor<int32, [3]> input_933_perm_0 = const()[name = string("input_933_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_935_pad_type_0 = const()[name = string("input_935_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_935_strides_0 = const()[name = string("input_935_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_935_pad_0 = const()[name = string("input_935_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_935_dilations_0 = const()[name = string("input_935_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_935_groups_0 = const()[name = string("input_935_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351774592))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353871808))))[name = string("encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_933_cast_fp16 = transpose(perm = input_933_perm_0, x = x_459_cast_fp16)[name = string("transpose_204")];
+            tensor<fp16, [1, 2048, 14]> input_935_cast_fp16 = conv(dilations = input_935_dilations_0, groups = input_935_groups_0, pad = input_935_pad_0, pad_type = input_935_pad_type_0, strides = input_935_strides_0, weight = encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_933_cast_fp16)[name = string("input_935_cast_fp16")];
+            int32 x_461_split_num_splits_0 = const()[name = string("x_461_split_num_splits_0"), val = int32(2)];
+            int32 x_461_split_axis_0 = const()[name = string("x_461_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_461_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_461_split_cast_fp16_1 = split(axis = x_461_split_axis_0, num_splits = x_461_split_num_splits_0, x = input_935_cast_fp16)[name = string("x_461_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_461_split_1_sigmoid_cast_fp16 = sigmoid(x = x_461_split_cast_fp16_1)[name = string("x_461_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_461_cast_fp16 = mul(x = x_461_split_cast_fp16_0, y = x_461_split_1_sigmoid_cast_fp16)[name = string("x_461_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_937_cast_fp16 = select(a = var_45_to_fp16, b = x_461_cast_fp16, cond = var_576)[name = string("input_937_cast_fp16")];
+            bool new_x_71_interleave_0 = const()[name = string("new_x_71_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_71_cast_fp16 = concat(axis = var_60, interleave = new_x_71_interleave_0, values = (cache_71_cast_fp16, input_937_cast_fp16))[name = string("new_x_71_cast_fp16")];
+            tensor<int32, [3]> var_4210_begin_0 = const()[name = string("op_4210_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_4210_end_0 = const()[name = string("op_4210_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_4210_end_mask_0 = const()[name = string("op_4210_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4210_cast_fp16 = slice_by_index(begin = var_4210_begin_0, end = var_4210_end_0, end_mask = var_4210_end_mask_0, x = new_x_71_cast_fp16)[name = string("op_4210_cast_fp16")];
+            string x_463_pad_type_0 = const()[name = string("x_463_pad_type_0"), val = string("valid")];
+            int32 x_463_groups_0 = const()[name = string("x_463_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_463_strides_0 = const()[name = string("x_463_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_463_pad_0 = const()[name = string("x_463_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_463_dilations_0 = const()[name = string("x_463_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353875968))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353885248))))[name = string("encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_463_cast_fp16 = conv(dilations = x_463_dilations_0, groups = x_463_groups_0, pad = x_463_pad_0, pad_type = x_463_pad_type_0, strides = x_463_strides_0, weight = encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_71_cast_fp16)[name = string("x_463_cast_fp16")];
+            tensor<int32, [3]> input_939_perm_0 = const()[name = string("input_939_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_465_axes_0 = const()[name = string("x_465_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353887360)))];
+            tensor<fp16, [1024]> encoder_layers_17_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353889472)))];
+            tensor<fp16, [1, 14, 1024]> input_939_cast_fp16 = transpose(perm = input_939_perm_0, x = x_463_cast_fp16)[name = string("transpose_203")];
+            tensor<fp16, [1, 14, 1024]> x_465_cast_fp16 = layer_norm(axes = x_465_axes_0, beta = encoder_layers_17_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_conv_batch_norm_weight_to_fp16, x = input_939_cast_fp16)[name = string("x_465_cast_fp16")];
+            tensor<int32, [3]> input_941_perm_0 = const()[name = string("input_941_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_941_cast_fp16 = transpose(perm = input_941_perm_0, x = x_465_cast_fp16)[name = string("transpose_202")];
+            tensor<fp16, [1, 1024, 14]> input_943_cast_fp16 = silu(x = input_941_cast_fp16)[name = string("input_943_cast_fp16")];
+            string x_467_pad_type_0 = const()[name = string("x_467_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_467_strides_0 = const()[name = string("x_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_467_pad_0 = const()[name = string("x_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_467_dilations_0 = const()[name = string("x_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_467_groups_0 = const()[name = string("x_467_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353891584))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354940224))))[name = string("encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_467_cast_fp16 = conv(dilations = x_467_dilations_0, groups = x_467_groups_0, pad = x_467_pad_0, pad_type = x_467_pad_type_0, strides = x_467_strides_0, weight = encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_943_cast_fp16)[name = string("x_467_cast_fp16")];
+            tensor<int32, [3]> input_945_perm_0 = const()[name = string("input_945_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_945_cast_fp16 = transpose(perm = input_945_perm_0, x = x_467_cast_fp16)[name = string("transpose_201")];
+            tensor<fp16, [1, 14, 1024]> input_947_cast_fp16 = add(x = input_931_cast_fp16, y = input_945_cast_fp16)[name = string("input_947_cast_fp16")];
+            tensor<int32, [1]> input_949_axes_0 = const()[name = string("input_949_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354942336)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354944448)))];
+            tensor<fp16, [1, 14, 1024]> input_949_cast_fp16 = layer_norm(axes = input_949_axes_0, beta = encoder_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_feed_forward2_weight_to_fp16, x = input_947_cast_fp16)[name = string("input_949_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354946560))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358092352))))[name = string("encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_17_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358092544)))];
+            tensor<fp16, [1, 14, 4096]> linear_161_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("linear_161_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_953_cast_fp16 = silu(x = linear_161_cast_fp16)[name = string("input_953_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358100800))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361246592))))[name = string("encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361246784)))];
+            tensor<fp16, [1, 14, 1024]> linear_162_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized, x = input_953_cast_fp16)[name = string("linear_162_cast_fp16")];
+            fp16 var_4253_to_fp16 = const()[name = string("op_4253_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4254_cast_fp16 = mul(x = linear_162_cast_fp16, y = var_4253_to_fp16)[name = string("op_4254_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_959_cast_fp16 = add(x = input_947_cast_fp16, y = var_4254_cast_fp16)[name = string("input_959_cast_fp16")];
+            tensor<int32, [1]> input_961_axes_0 = const()[name = string("input_961_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361248896)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361251008)))];
+            tensor<fp16, [1, 14, 1024]> input_961_cast_fp16 = layer_norm(axes = input_961_axes_0, beta = encoder_layers_17_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_out_weight_to_fp16, x = input_959_cast_fp16)[name = string("input_961_cast_fp16")];
+            tensor<int32, [4]> cache_73_begin_0 = const()[name = string("cache_73_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> cache_73_end_0 = const()[name = string("cache_73_end_0"), val = tensor<int32, [4]>([19, 1, 42, 1024])];
+            tensor<bool, [4]> cache_73_end_mask_0 = const()[name = string("cache_73_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_73_squeeze_mask_0 = const()[name = string("cache_73_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_73_cast_fp16 = slice_by_index(begin = cache_73_begin_0, end = cache_73_end_0, end_mask = cache_73_end_mask_0, squeeze_mask = cache_73_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_73_cast_fp16")];
+            tensor<int32, [4]> cache_75_begin_0 = const()[name = string("cache_75_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> cache_75_end_0 = const()[name = string("cache_75_end_0"), val = tensor<int32, [4]>([19, 1, 1024, 8])];
+            tensor<bool, [4]> cache_75_end_mask_0 = const()[name = string("cache_75_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_75_squeeze_mask_0 = const()[name = string("cache_75_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_75_cast_fp16 = slice_by_index(begin = cache_75_begin_0, end = cache_75_end_0, end_mask = cache_75_end_mask_0, squeeze_mask = cache_75_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_75_cast_fp16")];
+            tensor<int32, [1]> input_963_axes_0 = const()[name = string("input_963_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361253120)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361255232)))];
+            tensor<fp16, [1, 14, 1024]> input_963_cast_fp16 = layer_norm(axes = input_963_axes_0, beta = encoder_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_feed_forward1_weight_to_fp16, x = input_961_cast_fp16)[name = string("input_963_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361257344))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364403136))))[name = string("encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_18_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364403328)))];
+            tensor<fp16, [1, 14, 4096]> linear_163_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized, x = input_963_cast_fp16)[name = string("linear_163_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_967_cast_fp16 = silu(x = linear_163_cast_fp16)[name = string("input_967_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364411584))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367557376))))[name = string("encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367557568)))];
+            tensor<fp16, [1, 14, 1024]> linear_164_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized, x = input_967_cast_fp16)[name = string("linear_164_cast_fp16")];
+            fp16 var_4290_to_fp16 = const()[name = string("op_4290_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4291_cast_fp16 = mul(x = linear_164_cast_fp16, y = var_4290_to_fp16)[name = string("op_4291_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_973_cast_fp16 = add(x = input_961_cast_fp16, y = var_4291_cast_fp16)[name = string("input_973_cast_fp16")];
+            tensor<int32, [1]> key_37_axes_0 = const()[name = string("key_37_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367559680)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367561792)))];
+            tensor<fp16, [1, 14, 1024]> key_37_cast_fp16 = layer_norm(axes = key_37_axes_0, beta = encoder_layers_18_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_self_att_weight_to_fp16, x = input_973_cast_fp16)[name = string("key_37_cast_fp16")];
+            bool input_975_interleave_0 = const()[name = string("input_975_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_975_cast_fp16 = concat(axis = var_69, interleave = input_975_interleave_0, values = (cache_73_cast_fp16, key_37_cast_fp16))[name = string("input_975_cast_fp16")];
+            tensor<int32, [3]> var_4313_begin_0 = const()[name = string("op_4313_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_4313_end_0 = const()[name = string("op_4313_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4313_end_mask_0 = const()[name = string("op_4313_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_4313_cast_fp16 = slice_by_index(begin = var_4313_begin_0, end = var_4313_end_0, end_mask = var_4313_end_mask_0, x = cache_73_cast_fp16)[name = string("op_4313_cast_fp16")];
+            bool var_4319_interleave_0 = const()[name = string("op_4319_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4319_cast_fp16 = concat(axis = var_69, interleave = var_4319_interleave_0, values = (var_4313_cast_fp16, key_37_cast_fp16))[name = string("op_4319_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367563904))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368350400))))[name = string("encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368350592)))];
+            tensor<fp16, [1, 14, 1024]> linear_165_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized, x = key_37_cast_fp16)[name = string("linear_165_cast_fp16")];
+            tensor<int32, [4]> var_4324 = const()[name = string("op_4324"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_109_cast_fp16 = reshape(shape = var_4324, x = linear_165_cast_fp16)[name = string("q_109_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368352704))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369139200))))[name = string("encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369139392)))];
+            tensor<fp16, [1, 56, 1024]> linear_166_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_166_cast_fp16")];
+            tensor<int32, [4]> var_4329 = const()[name = string("op_4329"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_73_cast_fp16 = reshape(shape = var_4329, x = linear_166_cast_fp16)[name = string("k_73_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369141504))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369928000))))[name = string("encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369928192)))];
+            tensor<fp16, [1, 56, 1024]> linear_167_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_167_cast_fp16")];
+            tensor<int32, [4]> var_4334 = const()[name = string("op_4334"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_37_cast_fp16 = reshape(shape = var_4334, x = linear_167_cast_fp16)[name = string("v_37_cast_fp16")];
+            tensor<int32, [4]> value_45_perm_0 = const()[name = string("value_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_18_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369930304)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4347_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_u_to_fp16)[name = string("op_4347_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_18_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369932416)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4349_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_v_to_fp16)[name = string("op_4349_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_37_perm_0 = const()[name = string("q_with_bias_v_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_475_transpose_x_0 = const()[name = string("x_475_transpose_x_0"), val = bool(false)];
+            bool x_475_transpose_y_0 = const()[name = string("x_475_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_4351_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369934528))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370048256))))[name = string("op_4351_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_37_cast_fp16 = transpose(perm = q_with_bias_v_37_perm_0, x = var_4349_cast_fp16)[name = string("transpose_200")];
+            tensor<fp16, [1, 8, 14, 111]> x_475_cast_fp16 = matmul(transpose_x = x_475_transpose_x_0, transpose_y = x_475_transpose_y_0, x = q_with_bias_v_37_cast_fp16, y = op_4351_to_fp16_quantized)[name = string("x_475_cast_fp16")];
+            tensor<int32, [8]> x_477_pad_0 = const()[name = string("x_477_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_477_mode_0 = const()[name = string("x_477_mode_0"), val = string("constant")];
+            fp16 const_313_to_fp16 = const()[name = string("const_313_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_477_cast_fp16 = pad(constant_val = const_313_to_fp16, mode = x_477_mode_0, pad = x_477_pad_0, x = x_475_cast_fp16)[name = string("x_477_cast_fp16")];
+            tensor<int32, [4]> var_4359 = const()[name = string("op_4359"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_479_cast_fp16 = reshape(shape = var_4359, x = x_477_cast_fp16)[name = string("x_479_cast_fp16")];
+            tensor<int32, [4]> var_4363_begin_0 = const()[name = string("op_4363_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4363_end_0 = const()[name = string("op_4363_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_4363_end_mask_0 = const()[name = string("op_4363_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_4363_cast_fp16 = slice_by_index(begin = var_4363_begin_0, end = var_4363_end_0, end_mask = var_4363_end_mask_0, x = x_479_cast_fp16)[name = string("op_4363_cast_fp16")];
+            tensor<int32, [4]> var_4364 = const()[name = string("op_4364"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_73_cast_fp16 = reshape(shape = var_4364, x = var_4363_cast_fp16)[name = string("matrix_bd_73_cast_fp16")];
+            bool matrix_ac_37_transpose_x_0 = const()[name = string("matrix_ac_37_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_37_transpose_y_0 = const()[name = string("matrix_ac_37_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_132_perm_0 = const()[name = string("transpose_132_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_133_perm_0 = const()[name = string("transpose_133_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73_cast_fp16)[name = string("transpose_198")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_132 = transpose(perm = transpose_132_perm_0, x = var_4347_cast_fp16)[name = string("transpose_199")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_37_cast_fp16 = matmul(transpose_x = matrix_ac_37_transpose_x_0, transpose_y = matrix_ac_37_transpose_y_0, x = transpose_132, y = transpose_133)[name = string("matrix_ac_37_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_75_begin_0 = const()[name = string("matrix_bd_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_75_end_0 = const()[name = string("matrix_bd_75_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_75_end_mask_0 = const()[name = string("matrix_bd_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_75_cast_fp16 = slice_by_index(begin = matrix_bd_75_begin_0, end = matrix_bd_75_end_0, end_mask = matrix_bd_75_end_mask_0, x = matrix_bd_73_cast_fp16)[name = string("matrix_bd_75_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_4373_cast_fp16 = add(x = matrix_ac_37_cast_fp16, y = matrix_bd_75_cast_fp16)[name = string("op_4373_cast_fp16")];
+            fp16 _inversed_scores_73_y_0_to_fp16 = const()[name = string("_inversed_scores_73_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_73_cast_fp16 = mul(x = var_4373_cast_fp16, y = _inversed_scores_73_y_0_to_fp16)[name = string("_inversed_scores_73_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_75_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_73_cast_fp16, cond = mask_11)[name = string("scores_75_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_4379_cast_fp16 = softmax(axis = var_60, x = scores_75_cast_fp16)[name = string("op_4379_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_977_cast_fp16 = select(a = var_45_to_fp16, b = var_4379_cast_fp16, cond = mask_11)[name = string("input_977_cast_fp16")];
+            bool x_481_transpose_x_0 = const()[name = string("x_481_transpose_x_0"), val = bool(false)];
+            bool x_481_transpose_y_0 = const()[name = string("x_481_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = v_37_cast_fp16)[name = string("transpose_197")];
+            tensor<fp16, [1, 8, 14, 128]> x_481_cast_fp16 = matmul(transpose_x = x_481_transpose_x_0, transpose_y = x_481_transpose_y_0, x = input_977_cast_fp16, y = value_45_cast_fp16)[name = string("x_481_cast_fp16")];
+            tensor<int32, [4]> var_4383_perm_0 = const()[name = string("op_4383_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4384 = const()[name = string("op_4384"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_4383_cast_fp16 = transpose(perm = var_4383_perm_0, x = x_481_cast_fp16)[name = string("transpose_196")];
+            tensor<fp16, [1, 14, 1024]> input_979_cast_fp16 = reshape(shape = var_4384, x = var_4383_cast_fp16)[name = string("input_979_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370048576))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371097216))))[name = string("encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371099328)))];
+            tensor<fp16, [1, 14, 1024]> linear_169_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized, x = input_979_cast_fp16)[name = string("linear_169_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_983_cast_fp16 = add(x = input_973_cast_fp16, y = linear_169_cast_fp16)[name = string("input_983_cast_fp16")];
+            tensor<int32, [1]> x_485_axes_0 = const()[name = string("x_485_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371101440)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371103552)))];
+            tensor<fp16, [1, 14, 1024]> x_485_cast_fp16 = layer_norm(axes = x_485_axes_0, beta = encoder_layers_18_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_conv_weight_to_fp16, x = input_983_cast_fp16)[name = string("x_485_cast_fp16")];
+            tensor<int32, [3]> input_985_perm_0 = const()[name = string("input_985_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_987_pad_type_0 = const()[name = string("input_987_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_987_strides_0 = const()[name = string("input_987_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_987_pad_0 = const()[name = string("input_987_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_987_dilations_0 = const()[name = string("input_987_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_987_groups_0 = const()[name = string("input_987_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371105664))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373202880))))[name = string("encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_985_cast_fp16 = transpose(perm = input_985_perm_0, x = x_485_cast_fp16)[name = string("transpose_195")];
+            tensor<fp16, [1, 2048, 14]> input_987_cast_fp16 = conv(dilations = input_987_dilations_0, groups = input_987_groups_0, pad = input_987_pad_0, pad_type = input_987_pad_type_0, strides = input_987_strides_0, weight = encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_985_cast_fp16)[name = string("input_987_cast_fp16")];
+            int32 x_487_split_num_splits_0 = const()[name = string("x_487_split_num_splits_0"), val = int32(2)];
+            int32 x_487_split_axis_0 = const()[name = string("x_487_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_487_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_487_split_cast_fp16_1 = split(axis = x_487_split_axis_0, num_splits = x_487_split_num_splits_0, x = input_987_cast_fp16)[name = string("x_487_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_487_split_1_sigmoid_cast_fp16 = sigmoid(x = x_487_split_cast_fp16_1)[name = string("x_487_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_487_cast_fp16 = mul(x = x_487_split_cast_fp16_0, y = x_487_split_1_sigmoid_cast_fp16)[name = string("x_487_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_989_cast_fp16 = select(a = var_45_to_fp16, b = x_487_cast_fp16, cond = var_576)[name = string("input_989_cast_fp16")];
+            bool new_x_75_interleave_0 = const()[name = string("new_x_75_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_75_cast_fp16 = concat(axis = var_60, interleave = new_x_75_interleave_0, values = (cache_75_cast_fp16, input_989_cast_fp16))[name = string("new_x_75_cast_fp16")];
+            tensor<int32, [3]> var_4423_begin_0 = const()[name = string("op_4423_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_4423_end_0 = const()[name = string("op_4423_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_4423_end_mask_0 = const()[name = string("op_4423_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4423_cast_fp16 = slice_by_index(begin = var_4423_begin_0, end = var_4423_end_0, end_mask = var_4423_end_mask_0, x = new_x_75_cast_fp16)[name = string("op_4423_cast_fp16")];
+            string x_489_pad_type_0 = const()[name = string("x_489_pad_type_0"), val = string("valid")];
+            int32 x_489_groups_0 = const()[name = string("x_489_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_489_strides_0 = const()[name = string("x_489_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_489_pad_0 = const()[name = string("x_489_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_489_dilations_0 = const()[name = string("x_489_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373207040))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373216320))))[name = string("encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_489_cast_fp16 = conv(dilations = x_489_dilations_0, groups = x_489_groups_0, pad = x_489_pad_0, pad_type = x_489_pad_type_0, strides = x_489_strides_0, weight = encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_75_cast_fp16)[name = string("x_489_cast_fp16")];
+            tensor<int32, [3]> input_991_perm_0 = const()[name = string("input_991_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_491_axes_0 = const()[name = string("x_491_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373218432)))];
+            tensor<fp16, [1024]> encoder_layers_18_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373220544)))];
+            tensor<fp16, [1, 14, 1024]> input_991_cast_fp16 = transpose(perm = input_991_perm_0, x = x_489_cast_fp16)[name = string("transpose_194")];
+            tensor<fp16, [1, 14, 1024]> x_491_cast_fp16 = layer_norm(axes = x_491_axes_0, beta = encoder_layers_18_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_conv_batch_norm_weight_to_fp16, x = input_991_cast_fp16)[name = string("x_491_cast_fp16")];
+            tensor<int32, [3]> input_993_perm_0 = const()[name = string("input_993_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_993_cast_fp16 = transpose(perm = input_993_perm_0, x = x_491_cast_fp16)[name = string("transpose_193")];
+            tensor<fp16, [1, 1024, 14]> input_995_cast_fp16 = silu(x = input_993_cast_fp16)[name = string("input_995_cast_fp16")];
+            string x_493_pad_type_0 = const()[name = string("x_493_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_493_strides_0 = const()[name = string("x_493_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_493_pad_0 = const()[name = string("x_493_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_493_dilations_0 = const()[name = string("x_493_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_493_groups_0 = const()[name = string("x_493_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373222656))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374271296))))[name = string("encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_493_cast_fp16 = conv(dilations = x_493_dilations_0, groups = x_493_groups_0, pad = x_493_pad_0, pad_type = x_493_pad_type_0, strides = x_493_strides_0, weight = encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_995_cast_fp16)[name = string("x_493_cast_fp16")];
+            tensor<int32, [3]> input_997_perm_0 = const()[name = string("input_997_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_997_cast_fp16 = transpose(perm = input_997_perm_0, x = x_493_cast_fp16)[name = string("transpose_192")];
+            tensor<fp16, [1, 14, 1024]> input_999_cast_fp16 = add(x = input_983_cast_fp16, y = input_997_cast_fp16)[name = string("input_999_cast_fp16")];
+            tensor<int32, [1]> input_1001_axes_0 = const()[name = string("input_1001_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374273408)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374275520)))];
+            tensor<fp16, [1, 14, 1024]> input_1001_cast_fp16 = layer_norm(axes = input_1001_axes_0, beta = encoder_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_feed_forward2_weight_to_fp16, x = input_999_cast_fp16)[name = string("input_1001_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374277632))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378472000))))[name = string("encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_18_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378480256)))];
+            tensor<fp16, [1, 14, 4096]> linear_170_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1001_cast_fp16)[name = string("linear_170_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1005_cast_fp16 = silu(x = linear_170_cast_fp16)[name = string("input_1005_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378488512))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382682880))))[name = string("encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_18_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382684992)))];
+            tensor<fp16, [1, 14, 1024]> linear_171_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1005_cast_fp16)[name = string("linear_171_cast_fp16")];
+            fp16 var_4466_to_fp16 = const()[name = string("op_4466_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4467_cast_fp16 = mul(x = linear_171_cast_fp16, y = var_4466_to_fp16)[name = string("op_4467_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1011_cast_fp16 = add(x = input_999_cast_fp16, y = var_4467_cast_fp16)[name = string("input_1011_cast_fp16")];
+            tensor<int32, [1]> input_1013_axes_0 = const()[name = string("input_1013_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382687104)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382689216)))];
+            tensor<fp16, [1, 14, 1024]> input_1013_cast_fp16 = layer_norm(axes = input_1013_axes_0, beta = encoder_layers_18_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_out_weight_to_fp16, x = input_1011_cast_fp16)[name = string("input_1013_cast_fp16")];
+            tensor<int32, [4]> cache_77_begin_0 = const()[name = string("cache_77_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> cache_77_end_0 = const()[name = string("cache_77_end_0"), val = tensor<int32, [4]>([20, 1, 42, 1024])];
+            tensor<bool, [4]> cache_77_end_mask_0 = const()[name = string("cache_77_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_77_squeeze_mask_0 = const()[name = string("cache_77_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_77_cast_fp16 = slice_by_index(begin = cache_77_begin_0, end = cache_77_end_0, end_mask = cache_77_end_mask_0, squeeze_mask = cache_77_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_77_cast_fp16")];
+            tensor<int32, [4]> cache_79_begin_0 = const()[name = string("cache_79_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> cache_79_end_0 = const()[name = string("cache_79_end_0"), val = tensor<int32, [4]>([20, 1, 1024, 8])];
+            tensor<bool, [4]> cache_79_end_mask_0 = const()[name = string("cache_79_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_79_squeeze_mask_0 = const()[name = string("cache_79_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_79_cast_fp16 = slice_by_index(begin = cache_79_begin_0, end = cache_79_end_0, end_mask = cache_79_end_mask_0, squeeze_mask = cache_79_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_79_cast_fp16")];
+            tensor<int32, [1]> input_1015_axes_0 = const()[name = string("input_1015_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382691328)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382693440)))];
+            tensor<fp16, [1, 14, 1024]> input_1015_cast_fp16 = layer_norm(axes = input_1015_axes_0, beta = encoder_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_feed_forward1_weight_to_fp16, x = input_1013_cast_fp16)[name = string("input_1015_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382695552))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386889920))))[name = string("encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_19_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386898176)))];
+            tensor<fp16, [1, 14, 4096]> linear_172_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1015_cast_fp16)[name = string("linear_172_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1019_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("input_1019_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386906432))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391100800))))[name = string("encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391102912)))];
+            tensor<fp16, [1, 14, 1024]> linear_173_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1019_cast_fp16)[name = string("linear_173_cast_fp16")];
+            fp16 var_4503_to_fp16 = const()[name = string("op_4503_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4504_cast_fp16 = mul(x = linear_173_cast_fp16, y = var_4503_to_fp16)[name = string("op_4504_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1025_cast_fp16 = add(x = input_1013_cast_fp16, y = var_4504_cast_fp16)[name = string("input_1025_cast_fp16")];
+            tensor<int32, [1]> key_39_axes_0 = const()[name = string("key_39_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391105024)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391107136)))];
+            tensor<fp16, [1, 14, 1024]> key_39_cast_fp16 = layer_norm(axes = key_39_axes_0, beta = encoder_layers_19_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_self_att_weight_to_fp16, x = input_1025_cast_fp16)[name = string("key_39_cast_fp16")];
+            bool input_1027_interleave_0 = const()[name = string("input_1027_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_1027_cast_fp16 = concat(axis = var_69, interleave = input_1027_interleave_0, values = (cache_77_cast_fp16, key_39_cast_fp16))[name = string("input_1027_cast_fp16")];
+            tensor<int32, [3]> var_4526_begin_0 = const()[name = string("op_4526_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_4526_end_0 = const()[name = string("op_4526_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4526_end_mask_0 = const()[name = string("op_4526_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_4526_cast_fp16 = slice_by_index(begin = var_4526_begin_0, end = var_4526_end_0, end_mask = var_4526_end_mask_0, x = cache_77_cast_fp16)[name = string("op_4526_cast_fp16")];
+            bool var_4532_interleave_0 = const()[name = string("op_4532_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4532_cast_fp16 = concat(axis = var_69, interleave = var_4532_interleave_0, values = (var_4526_cast_fp16, key_39_cast_fp16))[name = string("op_4532_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391109248))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392157888))))[name = string("encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392160000)))];
+            tensor<fp16, [1, 14, 1024]> linear_174_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized, x = key_39_cast_fp16)[name = string("linear_174_cast_fp16")];
+            tensor<int32, [4]> var_4537 = const()[name = string("op_4537"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_115_cast_fp16 = reshape(shape = var_4537, x = linear_174_cast_fp16)[name = string("q_115_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392162112))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393210752))))[name = string("encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393212864)))];
+            tensor<fp16, [1, 56, 1024]> linear_175_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_175_cast_fp16")];
+            tensor<int32, [4]> var_4542 = const()[name = string("op_4542"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_77_cast_fp16 = reshape(shape = var_4542, x = linear_175_cast_fp16)[name = string("k_77_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393214976))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394263616))))[name = string("encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394265728)))];
+            tensor<fp16, [1, 56, 1024]> linear_176_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_176_cast_fp16")];
+            tensor<int32, [4]> var_4547 = const()[name = string("op_4547"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_39_cast_fp16 = reshape(shape = var_4547, x = linear_176_cast_fp16)[name = string("v_39_cast_fp16")];
+            tensor<int32, [4]> value_47_perm_0 = const()[name = string("value_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_19_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394267840)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4560_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_u_to_fp16)[name = string("op_4560_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_19_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394269952)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4562_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_v_to_fp16)[name = string("op_4562_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_39_perm_0 = const()[name = string("q_with_bias_v_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_501_transpose_x_0 = const()[name = string("x_501_transpose_x_0"), val = bool(false)];
+            bool x_501_transpose_y_0 = const()[name = string("x_501_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_4564_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394272064))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394385792))))[name = string("op_4564_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_39_cast_fp16 = transpose(perm = q_with_bias_v_39_perm_0, x = var_4562_cast_fp16)[name = string("transpose_191")];
+            tensor<fp16, [1, 8, 14, 111]> x_501_cast_fp16 = matmul(transpose_x = x_501_transpose_x_0, transpose_y = x_501_transpose_y_0, x = q_with_bias_v_39_cast_fp16, y = op_4564_to_fp16_quantized)[name = string("x_501_cast_fp16")];
+            tensor<int32, [8]> x_503_pad_0 = const()[name = string("x_503_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_503_mode_0 = const()[name = string("x_503_mode_0"), val = string("constant")];
+            fp16 const_326_to_fp16 = const()[name = string("const_326_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_503_cast_fp16 = pad(constant_val = const_326_to_fp16, mode = x_503_mode_0, pad = x_503_pad_0, x = x_501_cast_fp16)[name = string("x_503_cast_fp16")];
+            tensor<int32, [4]> var_4572 = const()[name = string("op_4572"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_505_cast_fp16 = reshape(shape = var_4572, x = x_503_cast_fp16)[name = string("x_505_cast_fp16")];
+            tensor<int32, [4]> var_4576_begin_0 = const()[name = string("op_4576_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4576_end_0 = const()[name = string("op_4576_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_4576_end_mask_0 = const()[name = string("op_4576_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_4576_cast_fp16 = slice_by_index(begin = var_4576_begin_0, end = var_4576_end_0, end_mask = var_4576_end_mask_0, x = x_505_cast_fp16)[name = string("op_4576_cast_fp16")];
+            tensor<int32, [4]> var_4577 = const()[name = string("op_4577"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_77_cast_fp16 = reshape(shape = var_4577, x = var_4576_cast_fp16)[name = string("matrix_bd_77_cast_fp16")];
+            bool matrix_ac_39_transpose_x_0 = const()[name = string("matrix_ac_39_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_39_transpose_y_0 = const()[name = string("matrix_ac_39_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_134_perm_0 = const()[name = string("transpose_134_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_135_perm_0 = const()[name = string("transpose_135_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77_cast_fp16)[name = string("transpose_189")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_134 = transpose(perm = transpose_134_perm_0, x = var_4560_cast_fp16)[name = string("transpose_190")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_39_cast_fp16 = matmul(transpose_x = matrix_ac_39_transpose_x_0, transpose_y = matrix_ac_39_transpose_y_0, x = transpose_134, y = transpose_135)[name = string("matrix_ac_39_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_79_begin_0 = const()[name = string("matrix_bd_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_79_end_0 = const()[name = string("matrix_bd_79_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_79_end_mask_0 = const()[name = string("matrix_bd_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_79_cast_fp16 = slice_by_index(begin = matrix_bd_79_begin_0, end = matrix_bd_79_end_0, end_mask = matrix_bd_79_end_mask_0, x = matrix_bd_77_cast_fp16)[name = string("matrix_bd_79_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_4586_cast_fp16 = add(x = matrix_ac_39_cast_fp16, y = matrix_bd_79_cast_fp16)[name = string("op_4586_cast_fp16")];
+            fp16 _inversed_scores_77_y_0_to_fp16 = const()[name = string("_inversed_scores_77_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_77_cast_fp16 = mul(x = var_4586_cast_fp16, y = _inversed_scores_77_y_0_to_fp16)[name = string("_inversed_scores_77_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_79_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_77_cast_fp16, cond = mask_11)[name = string("scores_79_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_4592_cast_fp16 = softmax(axis = var_60, x = scores_79_cast_fp16)[name = string("op_4592_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_1029_cast_fp16 = select(a = var_45_to_fp16, b = var_4592_cast_fp16, cond = mask_11)[name = string("input_1029_cast_fp16")];
+            bool x_507_transpose_x_0 = const()[name = string("x_507_transpose_x_0"), val = bool(false)];
+            bool x_507_transpose_y_0 = const()[name = string("x_507_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = v_39_cast_fp16)[name = string("transpose_188")];
+            tensor<fp16, [1, 8, 14, 128]> x_507_cast_fp16 = matmul(transpose_x = x_507_transpose_x_0, transpose_y = x_507_transpose_y_0, x = input_1029_cast_fp16, y = value_47_cast_fp16)[name = string("x_507_cast_fp16")];
+            tensor<int32, [4]> var_4596_perm_0 = const()[name = string("op_4596_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4597 = const()[name = string("op_4597"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_4596_cast_fp16 = transpose(perm = var_4596_perm_0, x = x_507_cast_fp16)[name = string("transpose_187")];
+            tensor<fp16, [1, 14, 1024]> input_1031_cast_fp16 = reshape(shape = var_4597, x = var_4596_cast_fp16)[name = string("input_1031_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394386112))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395434752))))[name = string("encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395436864)))];
+            tensor<fp16, [1, 14, 1024]> linear_178_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized, x = input_1031_cast_fp16)[name = string("linear_178_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1035_cast_fp16 = add(x = input_1025_cast_fp16, y = linear_178_cast_fp16)[name = string("input_1035_cast_fp16")];
+            tensor<int32, [1]> x_511_axes_0 = const()[name = string("x_511_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395438976)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395441088)))];
+            tensor<fp16, [1, 14, 1024]> x_511_cast_fp16 = layer_norm(axes = x_511_axes_0, beta = encoder_layers_19_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_conv_weight_to_fp16, x = input_1035_cast_fp16)[name = string("x_511_cast_fp16")];
+            tensor<int32, [3]> input_1037_perm_0 = const()[name = string("input_1037_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1039_pad_type_0 = const()[name = string("input_1039_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1039_strides_0 = const()[name = string("input_1039_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1039_pad_0 = const()[name = string("input_1039_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1039_dilations_0 = const()[name = string("input_1039_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1039_groups_0 = const()[name = string("input_1039_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395443200))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397540416))))[name = string("encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_1037_cast_fp16 = transpose(perm = input_1037_perm_0, x = x_511_cast_fp16)[name = string("transpose_186")];
+            tensor<fp16, [1, 2048, 14]> input_1039_cast_fp16 = conv(dilations = input_1039_dilations_0, groups = input_1039_groups_0, pad = input_1039_pad_0, pad_type = input_1039_pad_type_0, strides = input_1039_strides_0, weight = encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1037_cast_fp16)[name = string("input_1039_cast_fp16")];
+            int32 x_513_split_num_splits_0 = const()[name = string("x_513_split_num_splits_0"), val = int32(2)];
+            int32 x_513_split_axis_0 = const()[name = string("x_513_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_513_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_513_split_cast_fp16_1 = split(axis = x_513_split_axis_0, num_splits = x_513_split_num_splits_0, x = input_1039_cast_fp16)[name = string("x_513_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_513_split_1_sigmoid_cast_fp16 = sigmoid(x = x_513_split_cast_fp16_1)[name = string("x_513_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_513_cast_fp16 = mul(x = x_513_split_cast_fp16_0, y = x_513_split_1_sigmoid_cast_fp16)[name = string("x_513_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_1041_cast_fp16 = select(a = var_45_to_fp16, b = x_513_cast_fp16, cond = var_576)[name = string("input_1041_cast_fp16")];
+            bool new_x_79_interleave_0 = const()[name = string("new_x_79_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_79_cast_fp16 = concat(axis = var_60, interleave = new_x_79_interleave_0, values = (cache_79_cast_fp16, input_1041_cast_fp16))[name = string("new_x_79_cast_fp16")];
+            tensor<int32, [3]> var_4636_begin_0 = const()[name = string("op_4636_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_4636_end_0 = const()[name = string("op_4636_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_4636_end_mask_0 = const()[name = string("op_4636_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4636_cast_fp16 = slice_by_index(begin = var_4636_begin_0, end = var_4636_end_0, end_mask = var_4636_end_mask_0, x = new_x_79_cast_fp16)[name = string("op_4636_cast_fp16")];
+            string x_515_pad_type_0 = const()[name = string("x_515_pad_type_0"), val = string("valid")];
+            int32 x_515_groups_0 = const()[name = string("x_515_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_515_strides_0 = const()[name = string("x_515_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_515_pad_0 = const()[name = string("x_515_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_515_dilations_0 = const()[name = string("x_515_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397544576))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397553856))))[name = string("encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_515_cast_fp16 = conv(dilations = x_515_dilations_0, groups = x_515_groups_0, pad = x_515_pad_0, pad_type = x_515_pad_type_0, strides = x_515_strides_0, weight = encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_79_cast_fp16)[name = string("x_515_cast_fp16")];
+            tensor<int32, [3]> input_1043_perm_0 = const()[name = string("input_1043_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_517_axes_0 = const()[name = string("x_517_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397555968)))];
+            tensor<fp16, [1024]> encoder_layers_19_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397558080)))];
+            tensor<fp16, [1, 14, 1024]> input_1043_cast_fp16 = transpose(perm = input_1043_perm_0, x = x_515_cast_fp16)[name = string("transpose_185")];
+            tensor<fp16, [1, 14, 1024]> x_517_cast_fp16 = layer_norm(axes = x_517_axes_0, beta = encoder_layers_19_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_conv_batch_norm_weight_to_fp16, x = input_1043_cast_fp16)[name = string("x_517_cast_fp16")];
+            tensor<int32, [3]> input_1045_perm_0 = const()[name = string("input_1045_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_1045_cast_fp16 = transpose(perm = input_1045_perm_0, x = x_517_cast_fp16)[name = string("transpose_184")];
+            tensor<fp16, [1, 1024, 14]> input_1047_cast_fp16 = silu(x = input_1045_cast_fp16)[name = string("input_1047_cast_fp16")];
+            string x_519_pad_type_0 = const()[name = string("x_519_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_519_strides_0 = const()[name = string("x_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_519_pad_0 = const()[name = string("x_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_519_dilations_0 = const()[name = string("x_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_519_groups_0 = const()[name = string("x_519_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397560192))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398608832))))[name = string("encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_519_cast_fp16 = conv(dilations = x_519_dilations_0, groups = x_519_groups_0, pad = x_519_pad_0, pad_type = x_519_pad_type_0, strides = x_519_strides_0, weight = encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1047_cast_fp16)[name = string("x_519_cast_fp16")];
+            tensor<int32, [3]> input_1049_perm_0 = const()[name = string("input_1049_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_1049_cast_fp16 = transpose(perm = input_1049_perm_0, x = x_519_cast_fp16)[name = string("transpose_183")];
+            tensor<fp16, [1, 14, 1024]> input_1051_cast_fp16 = add(x = input_1035_cast_fp16, y = input_1049_cast_fp16)[name = string("input_1051_cast_fp16")];
+            tensor<int32, [1]> input_1053_axes_0 = const()[name = string("input_1053_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398610944)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398613056)))];
+            tensor<fp16, [1, 14, 1024]> input_1053_cast_fp16 = layer_norm(axes = input_1053_axes_0, beta = encoder_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_feed_forward2_weight_to_fp16, x = input_1051_cast_fp16)[name = string("input_1053_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398615168))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402809536))))[name = string("encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_19_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402817792)))];
+            tensor<fp16, [1, 14, 4096]> linear_179_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1053_cast_fp16)[name = string("linear_179_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1057_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("input_1057_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402826048))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407020416))))[name = string("encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407022528)))];
+            tensor<fp16, [1, 14, 1024]> linear_180_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1057_cast_fp16)[name = string("linear_180_cast_fp16")];
+            fp16 var_4679_to_fp16 = const()[name = string("op_4679_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4680_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_4679_to_fp16)[name = string("op_4680_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1063_cast_fp16 = add(x = input_1051_cast_fp16, y = var_4680_cast_fp16)[name = string("input_1063_cast_fp16")];
+            tensor<int32, [1]> input_1065_axes_0 = const()[name = string("input_1065_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407024640)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407026752)))];
+            tensor<fp16, [1, 14, 1024]> input_1065_cast_fp16 = layer_norm(axes = input_1065_axes_0, beta = encoder_layers_19_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_out_weight_to_fp16, x = input_1063_cast_fp16)[name = string("input_1065_cast_fp16")];
+            tensor<int32, [4]> cache_81_begin_0 = const()[name = string("cache_81_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> cache_81_end_0 = const()[name = string("cache_81_end_0"), val = tensor<int32, [4]>([21, 1, 42, 1024])];
+            tensor<bool, [4]> cache_81_end_mask_0 = const()[name = string("cache_81_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_81_squeeze_mask_0 = const()[name = string("cache_81_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_81_cast_fp16 = slice_by_index(begin = cache_81_begin_0, end = cache_81_end_0, end_mask = cache_81_end_mask_0, squeeze_mask = cache_81_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_81_cast_fp16")];
+            tensor<int32, [4]> cache_83_begin_0 = const()[name = string("cache_83_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> cache_83_end_0 = const()[name = string("cache_83_end_0"), val = tensor<int32, [4]>([21, 1, 1024, 8])];
+            tensor<bool, [4]> cache_83_end_mask_0 = const()[name = string("cache_83_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_83_squeeze_mask_0 = const()[name = string("cache_83_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_83_cast_fp16 = slice_by_index(begin = cache_83_begin_0, end = cache_83_end_0, end_mask = cache_83_end_mask_0, squeeze_mask = cache_83_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_83_cast_fp16")];
+            tensor<int32, [1]> input_1067_axes_0 = const()[name = string("input_1067_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407028864)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407030976)))];
+            tensor<fp16, [1, 14, 1024]> input_1067_cast_fp16 = layer_norm(axes = input_1067_axes_0, beta = encoder_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_feed_forward1_weight_to_fp16, x = input_1065_cast_fp16)[name = string("input_1067_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407033088))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411227456))))[name = string("encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_20_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411235712)))];
+            tensor<fp16, [1, 14, 4096]> linear_181_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1067_cast_fp16)[name = string("linear_181_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1071_cast_fp16 = silu(x = linear_181_cast_fp16)[name = string("input_1071_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411243968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415438336))))[name = string("encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415440448)))];
+            tensor<fp16, [1, 14, 1024]> linear_182_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1071_cast_fp16)[name = string("linear_182_cast_fp16")];
+            fp16 var_4716_to_fp16 = const()[name = string("op_4716_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4717_cast_fp16 = mul(x = linear_182_cast_fp16, y = var_4716_to_fp16)[name = string("op_4717_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1077_cast_fp16 = add(x = input_1065_cast_fp16, y = var_4717_cast_fp16)[name = string("input_1077_cast_fp16")];
+            tensor<int32, [1]> key_41_axes_0 = const()[name = string("key_41_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415442560)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415444672)))];
+            tensor<fp16, [1, 14, 1024]> key_41_cast_fp16 = layer_norm(axes = key_41_axes_0, beta = encoder_layers_20_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_self_att_weight_to_fp16, x = input_1077_cast_fp16)[name = string("key_41_cast_fp16")];
+            bool input_1079_interleave_0 = const()[name = string("input_1079_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_1079_cast_fp16 = concat(axis = var_69, interleave = input_1079_interleave_0, values = (cache_81_cast_fp16, key_41_cast_fp16))[name = string("input_1079_cast_fp16")];
+            tensor<int32, [3]> var_4739_begin_0 = const()[name = string("op_4739_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_4739_end_0 = const()[name = string("op_4739_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4739_end_mask_0 = const()[name = string("op_4739_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_4739_cast_fp16 = slice_by_index(begin = var_4739_begin_0, end = var_4739_end_0, end_mask = var_4739_end_mask_0, x = cache_81_cast_fp16)[name = string("op_4739_cast_fp16")];
+            bool var_4745_interleave_0 = const()[name = string("op_4745_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4745_cast_fp16 = concat(axis = var_69, interleave = var_4745_interleave_0, values = (var_4739_cast_fp16, key_41_cast_fp16))[name = string("op_4745_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415446784))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416495424))))[name = string("encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416497536)))];
+            tensor<fp16, [1, 14, 1024]> linear_183_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized, x = key_41_cast_fp16)[name = string("linear_183_cast_fp16")];
+            tensor<int32, [4]> var_4750 = const()[name = string("op_4750"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_121_cast_fp16 = reshape(shape = var_4750, x = linear_183_cast_fp16)[name = string("q_121_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416499648))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417548288))))[name = string("encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417550400)))];
+            tensor<fp16, [1, 56, 1024]> linear_184_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_184_cast_fp16")];
+            tensor<int32, [4]> var_4755 = const()[name = string("op_4755"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_81_cast_fp16 = reshape(shape = var_4755, x = linear_184_cast_fp16)[name = string("k_81_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417552512))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418601152))))[name = string("encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418603264)))];
+            tensor<fp16, [1, 56, 1024]> linear_185_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_185_cast_fp16")];
+            tensor<int32, [4]> var_4760 = const()[name = string("op_4760"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_41_cast_fp16 = reshape(shape = var_4760, x = linear_185_cast_fp16)[name = string("v_41_cast_fp16")];
+            tensor<int32, [4]> value_49_perm_0 = const()[name = string("value_49_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_20_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418605376)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4773_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_u_to_fp16)[name = string("op_4773_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_20_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418607488)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4775_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_v_to_fp16)[name = string("op_4775_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_41_perm_0 = const()[name = string("q_with_bias_v_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_527_transpose_x_0 = const()[name = string("x_527_transpose_x_0"), val = bool(false)];
+            bool x_527_transpose_y_0 = const()[name = string("x_527_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_4777_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418609600))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418723328))))[name = string("op_4777_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_41_cast_fp16 = transpose(perm = q_with_bias_v_41_perm_0, x = var_4775_cast_fp16)[name = string("transpose_182")];
+            tensor<fp16, [1, 8, 14, 111]> x_527_cast_fp16 = matmul(transpose_x = x_527_transpose_x_0, transpose_y = x_527_transpose_y_0, x = q_with_bias_v_41_cast_fp16, y = op_4777_to_fp16_quantized)[name = string("x_527_cast_fp16")];
+            tensor<int32, [8]> x_529_pad_0 = const()[name = string("x_529_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_529_mode_0 = const()[name = string("x_529_mode_0"), val = string("constant")];
+            fp16 const_339_to_fp16 = const()[name = string("const_339_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_529_cast_fp16 = pad(constant_val = const_339_to_fp16, mode = x_529_mode_0, pad = x_529_pad_0, x = x_527_cast_fp16)[name = string("x_529_cast_fp16")];
+            tensor<int32, [4]> var_4785 = const()[name = string("op_4785"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_531_cast_fp16 = reshape(shape = var_4785, x = x_529_cast_fp16)[name = string("x_531_cast_fp16")];
+            tensor<int32, [4]> var_4789_begin_0 = const()[name = string("op_4789_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4789_end_0 = const()[name = string("op_4789_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_4789_end_mask_0 = const()[name = string("op_4789_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_4789_cast_fp16 = slice_by_index(begin = var_4789_begin_0, end = var_4789_end_0, end_mask = var_4789_end_mask_0, x = x_531_cast_fp16)[name = string("op_4789_cast_fp16")];
+            tensor<int32, [4]> var_4790 = const()[name = string("op_4790"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_81_cast_fp16 = reshape(shape = var_4790, x = var_4789_cast_fp16)[name = string("matrix_bd_81_cast_fp16")];
+            bool matrix_ac_41_transpose_x_0 = const()[name = string("matrix_ac_41_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_41_transpose_y_0 = const()[name = string("matrix_ac_41_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_136_perm_0 = const()[name = string("transpose_136_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_137_perm_0 = const()[name = string("transpose_137_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81_cast_fp16)[name = string("transpose_180")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_136 = transpose(perm = transpose_136_perm_0, x = var_4773_cast_fp16)[name = string("transpose_181")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_41_cast_fp16 = matmul(transpose_x = matrix_ac_41_transpose_x_0, transpose_y = matrix_ac_41_transpose_y_0, x = transpose_136, y = transpose_137)[name = string("matrix_ac_41_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_83_begin_0 = const()[name = string("matrix_bd_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_83_end_0 = const()[name = string("matrix_bd_83_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_83_end_mask_0 = const()[name = string("matrix_bd_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_83_cast_fp16 = slice_by_index(begin = matrix_bd_83_begin_0, end = matrix_bd_83_end_0, end_mask = matrix_bd_83_end_mask_0, x = matrix_bd_81_cast_fp16)[name = string("matrix_bd_83_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_4799_cast_fp16 = add(x = matrix_ac_41_cast_fp16, y = matrix_bd_83_cast_fp16)[name = string("op_4799_cast_fp16")];
+            fp16 _inversed_scores_81_y_0_to_fp16 = const()[name = string("_inversed_scores_81_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_81_cast_fp16 = mul(x = var_4799_cast_fp16, y = _inversed_scores_81_y_0_to_fp16)[name = string("_inversed_scores_81_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_83_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_81_cast_fp16, cond = mask_11)[name = string("scores_83_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_4805_cast_fp16 = softmax(axis = var_60, x = scores_83_cast_fp16)[name = string("op_4805_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_1081_cast_fp16 = select(a = var_45_to_fp16, b = var_4805_cast_fp16, cond = mask_11)[name = string("input_1081_cast_fp16")];
+            bool x_533_transpose_x_0 = const()[name = string("x_533_transpose_x_0"), val = bool(false)];
+            bool x_533_transpose_y_0 = const()[name = string("x_533_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = v_41_cast_fp16)[name = string("transpose_179")];
+            tensor<fp16, [1, 8, 14, 128]> x_533_cast_fp16 = matmul(transpose_x = x_533_transpose_x_0, transpose_y = x_533_transpose_y_0, x = input_1081_cast_fp16, y = value_49_cast_fp16)[name = string("x_533_cast_fp16")];
+            tensor<int32, [4]> var_4809_perm_0 = const()[name = string("op_4809_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4810 = const()[name = string("op_4810"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_4809_cast_fp16 = transpose(perm = var_4809_perm_0, x = x_533_cast_fp16)[name = string("transpose_178")];
+            tensor<fp16, [1, 14, 1024]> input_1083_cast_fp16 = reshape(shape = var_4810, x = var_4809_cast_fp16)[name = string("input_1083_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418723648))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419772288))))[name = string("encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419774400)))];
+            tensor<fp16, [1, 14, 1024]> linear_187_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized, x = input_1083_cast_fp16)[name = string("linear_187_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1087_cast_fp16 = add(x = input_1077_cast_fp16, y = linear_187_cast_fp16)[name = string("input_1087_cast_fp16")];
+            tensor<int32, [1]> x_537_axes_0 = const()[name = string("x_537_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419776512)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419778624)))];
+            tensor<fp16, [1, 14, 1024]> x_537_cast_fp16 = layer_norm(axes = x_537_axes_0, beta = encoder_layers_20_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_conv_weight_to_fp16, x = input_1087_cast_fp16)[name = string("x_537_cast_fp16")];
+            tensor<int32, [3]> input_1089_perm_0 = const()[name = string("input_1089_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1091_pad_type_0 = const()[name = string("input_1091_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1091_strides_0 = const()[name = string("input_1091_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1091_pad_0 = const()[name = string("input_1091_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1091_dilations_0 = const()[name = string("input_1091_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1091_groups_0 = const()[name = string("input_1091_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419780736))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421877952))))[name = string("encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_1089_cast_fp16 = transpose(perm = input_1089_perm_0, x = x_537_cast_fp16)[name = string("transpose_177")];
+            tensor<fp16, [1, 2048, 14]> input_1091_cast_fp16 = conv(dilations = input_1091_dilations_0, groups = input_1091_groups_0, pad = input_1091_pad_0, pad_type = input_1091_pad_type_0, strides = input_1091_strides_0, weight = encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1089_cast_fp16)[name = string("input_1091_cast_fp16")];
+            int32 x_539_split_num_splits_0 = const()[name = string("x_539_split_num_splits_0"), val = int32(2)];
+            int32 x_539_split_axis_0 = const()[name = string("x_539_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_539_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_539_split_cast_fp16_1 = split(axis = x_539_split_axis_0, num_splits = x_539_split_num_splits_0, x = input_1091_cast_fp16)[name = string("x_539_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_539_split_1_sigmoid_cast_fp16 = sigmoid(x = x_539_split_cast_fp16_1)[name = string("x_539_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_539_cast_fp16 = mul(x = x_539_split_cast_fp16_0, y = x_539_split_1_sigmoid_cast_fp16)[name = string("x_539_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_1093_cast_fp16 = select(a = var_45_to_fp16, b = x_539_cast_fp16, cond = var_576)[name = string("input_1093_cast_fp16")];
+            bool new_x_83_interleave_0 = const()[name = string("new_x_83_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_83_cast_fp16 = concat(axis = var_60, interleave = new_x_83_interleave_0, values = (cache_83_cast_fp16, input_1093_cast_fp16))[name = string("new_x_83_cast_fp16")];
+            tensor<int32, [3]> var_4849_begin_0 = const()[name = string("op_4849_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_4849_end_0 = const()[name = string("op_4849_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_4849_end_mask_0 = const()[name = string("op_4849_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4849_cast_fp16 = slice_by_index(begin = var_4849_begin_0, end = var_4849_end_0, end_mask = var_4849_end_mask_0, x = new_x_83_cast_fp16)[name = string("op_4849_cast_fp16")];
+            string x_541_pad_type_0 = const()[name = string("x_541_pad_type_0"), val = string("valid")];
+            int32 x_541_groups_0 = const()[name = string("x_541_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_541_strides_0 = const()[name = string("x_541_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_541_pad_0 = const()[name = string("x_541_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_541_dilations_0 = const()[name = string("x_541_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421882112))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421891392))))[name = string("encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_541_cast_fp16 = conv(dilations = x_541_dilations_0, groups = x_541_groups_0, pad = x_541_pad_0, pad_type = x_541_pad_type_0, strides = x_541_strides_0, weight = encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_83_cast_fp16)[name = string("x_541_cast_fp16")];
+            tensor<int32, [3]> input_1095_perm_0 = const()[name = string("input_1095_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_543_axes_0 = const()[name = string("x_543_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421893504)))];
+            tensor<fp16, [1024]> encoder_layers_20_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421895616)))];
+            tensor<fp16, [1, 14, 1024]> input_1095_cast_fp16 = transpose(perm = input_1095_perm_0, x = x_541_cast_fp16)[name = string("transpose_176")];
+            tensor<fp16, [1, 14, 1024]> x_543_cast_fp16 = layer_norm(axes = x_543_axes_0, beta = encoder_layers_20_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_conv_batch_norm_weight_to_fp16, x = input_1095_cast_fp16)[name = string("x_543_cast_fp16")];
+            tensor<int32, [3]> input_1097_perm_0 = const()[name = string("input_1097_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_1097_cast_fp16 = transpose(perm = input_1097_perm_0, x = x_543_cast_fp16)[name = string("transpose_175")];
+            tensor<fp16, [1, 1024, 14]> input_1099_cast_fp16 = silu(x = input_1097_cast_fp16)[name = string("input_1099_cast_fp16")];
+            string x_545_pad_type_0 = const()[name = string("x_545_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_545_strides_0 = const()[name = string("x_545_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_545_pad_0 = const()[name = string("x_545_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_545_dilations_0 = const()[name = string("x_545_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_545_groups_0 = const()[name = string("x_545_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421897728))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422946368))))[name = string("encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_545_cast_fp16 = conv(dilations = x_545_dilations_0, groups = x_545_groups_0, pad = x_545_pad_0, pad_type = x_545_pad_type_0, strides = x_545_strides_0, weight = encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1099_cast_fp16)[name = string("x_545_cast_fp16")];
+            tensor<int32, [3]> input_1101_perm_0 = const()[name = string("input_1101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_1101_cast_fp16 = transpose(perm = input_1101_perm_0, x = x_545_cast_fp16)[name = string("transpose_174")];
+            tensor<fp16, [1, 14, 1024]> input_1103_cast_fp16 = add(x = input_1087_cast_fp16, y = input_1101_cast_fp16)[name = string("input_1103_cast_fp16")];
+            tensor<int32, [1]> input_1105_axes_0 = const()[name = string("input_1105_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422948480)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422950592)))];
+            tensor<fp16, [1, 14, 1024]> input_1105_cast_fp16 = layer_norm(axes = input_1105_axes_0, beta = encoder_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_feed_forward2_weight_to_fp16, x = input_1103_cast_fp16)[name = string("input_1105_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422952704))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427147072))))[name = string("encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_20_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427155328)))];
+            tensor<fp16, [1, 14, 4096]> linear_188_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1105_cast_fp16)[name = string("linear_188_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1109_cast_fp16 = silu(x = linear_188_cast_fp16)[name = string("input_1109_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427163584))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431357952))))[name = string("encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431360064)))];
+            tensor<fp16, [1, 14, 1024]> linear_189_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1109_cast_fp16)[name = string("linear_189_cast_fp16")];
+            fp16 var_4892_to_fp16 = const()[name = string("op_4892_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4893_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_4892_to_fp16)[name = string("op_4893_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1115_cast_fp16 = add(x = input_1103_cast_fp16, y = var_4893_cast_fp16)[name = string("input_1115_cast_fp16")];
+            tensor<int32, [1]> input_1117_axes_0 = const()[name = string("input_1117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431362176)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431364288)))];
+            tensor<fp16, [1, 14, 1024]> input_1117_cast_fp16 = layer_norm(axes = input_1117_axes_0, beta = encoder_layers_20_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_out_weight_to_fp16, x = input_1115_cast_fp16)[name = string("input_1117_cast_fp16")];
+            tensor<int32, [4]> cache_85_begin_0 = const()[name = string("cache_85_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> cache_85_end_0 = const()[name = string("cache_85_end_0"), val = tensor<int32, [4]>([22, 1, 42, 1024])];
+            tensor<bool, [4]> cache_85_end_mask_0 = const()[name = string("cache_85_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_85_squeeze_mask_0 = const()[name = string("cache_85_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_85_cast_fp16 = slice_by_index(begin = cache_85_begin_0, end = cache_85_end_0, end_mask = cache_85_end_mask_0, squeeze_mask = cache_85_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_85_cast_fp16")];
+            tensor<int32, [4]> cache_87_begin_0 = const()[name = string("cache_87_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> cache_87_end_0 = const()[name = string("cache_87_end_0"), val = tensor<int32, [4]>([22, 1, 1024, 8])];
+            tensor<bool, [4]> cache_87_end_mask_0 = const()[name = string("cache_87_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_87_squeeze_mask_0 = const()[name = string("cache_87_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_87_cast_fp16 = slice_by_index(begin = cache_87_begin_0, end = cache_87_end_0, end_mask = cache_87_end_mask_0, squeeze_mask = cache_87_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_87_cast_fp16")];
+            tensor<int32, [1]> input_1119_axes_0 = const()[name = string("input_1119_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431366400)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431368512)))];
+            tensor<fp16, [1, 14, 1024]> input_1119_cast_fp16 = layer_norm(axes = input_1119_axes_0, beta = encoder_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_feed_forward1_weight_to_fp16, x = input_1117_cast_fp16)[name = string("input_1119_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431370624))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435564992))))[name = string("encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_21_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435573248)))];
+            tensor<fp16, [1, 14, 4096]> linear_190_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1119_cast_fp16)[name = string("linear_190_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1123_cast_fp16 = silu(x = linear_190_cast_fp16)[name = string("input_1123_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435581504))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439775872))))[name = string("encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439777984)))];
+            tensor<fp16, [1, 14, 1024]> linear_191_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1123_cast_fp16)[name = string("linear_191_cast_fp16")];
+            fp16 var_4929_to_fp16 = const()[name = string("op_4929_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4930_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_4929_to_fp16)[name = string("op_4930_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1129_cast_fp16 = add(x = input_1117_cast_fp16, y = var_4930_cast_fp16)[name = string("input_1129_cast_fp16")];
+            tensor<int32, [1]> key_43_axes_0 = const()[name = string("key_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439780096)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439782208)))];
+            tensor<fp16, [1, 14, 1024]> key_43_cast_fp16 = layer_norm(axes = key_43_axes_0, beta = encoder_layers_21_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_self_att_weight_to_fp16, x = input_1129_cast_fp16)[name = string("key_43_cast_fp16")];
+            bool input_1131_interleave_0 = const()[name = string("input_1131_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_1131_cast_fp16 = concat(axis = var_69, interleave = input_1131_interleave_0, values = (cache_85_cast_fp16, key_43_cast_fp16))[name = string("input_1131_cast_fp16")];
+            tensor<int32, [3]> var_4952_begin_0 = const()[name = string("op_4952_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_4952_end_0 = const()[name = string("op_4952_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4952_end_mask_0 = const()[name = string("op_4952_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_4952_cast_fp16 = slice_by_index(begin = var_4952_begin_0, end = var_4952_end_0, end_mask = var_4952_end_mask_0, x = cache_85_cast_fp16)[name = string("op_4952_cast_fp16")];
+            bool var_4958_interleave_0 = const()[name = string("op_4958_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4958_cast_fp16 = concat(axis = var_69, interleave = var_4958_interleave_0, values = (var_4952_cast_fp16, key_43_cast_fp16))[name = string("op_4958_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439784320))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440832960))))[name = string("encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440835072)))];
+            tensor<fp16, [1, 14, 1024]> linear_192_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized, x = key_43_cast_fp16)[name = string("linear_192_cast_fp16")];
+            tensor<int32, [4]> var_4963 = const()[name = string("op_4963"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_127_cast_fp16 = reshape(shape = var_4963, x = linear_192_cast_fp16)[name = string("q_127_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440837184))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441885824))))[name = string("encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441887936)))];
+            tensor<fp16, [1, 56, 1024]> linear_193_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_193_cast_fp16")];
+            tensor<int32, [4]> var_4968 = const()[name = string("op_4968"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_85_cast_fp16 = reshape(shape = var_4968, x = linear_193_cast_fp16)[name = string("k_85_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441890048))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442938688))))[name = string("encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442940800)))];
+            tensor<fp16, [1, 56, 1024]> linear_194_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_194_cast_fp16")];
+            tensor<int32, [4]> var_4973 = const()[name = string("op_4973"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_43_cast_fp16 = reshape(shape = var_4973, x = linear_194_cast_fp16)[name = string("v_43_cast_fp16")];
+            tensor<int32, [4]> value_51_perm_0 = const()[name = string("value_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_21_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442942912)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4986_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_u_to_fp16)[name = string("op_4986_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_21_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442945024)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4988_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_v_to_fp16)[name = string("op_4988_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_43_perm_0 = const()[name = string("q_with_bias_v_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_553_transpose_x_0 = const()[name = string("x_553_transpose_x_0"), val = bool(false)];
+            bool x_553_transpose_y_0 = const()[name = string("x_553_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_4990_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442947136))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443060864))))[name = string("op_4990_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_43_cast_fp16 = transpose(perm = q_with_bias_v_43_perm_0, x = var_4988_cast_fp16)[name = string("transpose_173")];
+            tensor<fp16, [1, 8, 14, 111]> x_553_cast_fp16 = matmul(transpose_x = x_553_transpose_x_0, transpose_y = x_553_transpose_y_0, x = q_with_bias_v_43_cast_fp16, y = op_4990_to_fp16_quantized)[name = string("x_553_cast_fp16")];
+            tensor<int32, [8]> x_555_pad_0 = const()[name = string("x_555_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_555_mode_0 = const()[name = string("x_555_mode_0"), val = string("constant")];
+            fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_555_cast_fp16 = pad(constant_val = const_352_to_fp16, mode = x_555_mode_0, pad = x_555_pad_0, x = x_553_cast_fp16)[name = string("x_555_cast_fp16")];
+            tensor<int32, [4]> var_4998 = const()[name = string("op_4998"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_557_cast_fp16 = reshape(shape = var_4998, x = x_555_cast_fp16)[name = string("x_557_cast_fp16")];
+            tensor<int32, [4]> var_5002_begin_0 = const()[name = string("op_5002_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5002_end_0 = const()[name = string("op_5002_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_5002_end_mask_0 = const()[name = string("op_5002_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_5002_cast_fp16 = slice_by_index(begin = var_5002_begin_0, end = var_5002_end_0, end_mask = var_5002_end_mask_0, x = x_557_cast_fp16)[name = string("op_5002_cast_fp16")];
+            tensor<int32, [4]> var_5003 = const()[name = string("op_5003"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_85_cast_fp16 = reshape(shape = var_5003, x = var_5002_cast_fp16)[name = string("matrix_bd_85_cast_fp16")];
+            bool matrix_ac_43_transpose_x_0 = const()[name = string("matrix_ac_43_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_43_transpose_y_0 = const()[name = string("matrix_ac_43_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_138_perm_0 = const()[name = string("transpose_138_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_139_perm_0 = const()[name = string("transpose_139_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85_cast_fp16)[name = string("transpose_171")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_138 = transpose(perm = transpose_138_perm_0, x = var_4986_cast_fp16)[name = string("transpose_172")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_43_cast_fp16 = matmul(transpose_x = matrix_ac_43_transpose_x_0, transpose_y = matrix_ac_43_transpose_y_0, x = transpose_138, y = transpose_139)[name = string("matrix_ac_43_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_87_begin_0 = const()[name = string("matrix_bd_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_87_end_0 = const()[name = string("matrix_bd_87_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_87_end_mask_0 = const()[name = string("matrix_bd_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_87_cast_fp16 = slice_by_index(begin = matrix_bd_87_begin_0, end = matrix_bd_87_end_0, end_mask = matrix_bd_87_end_mask_0, x = matrix_bd_85_cast_fp16)[name = string("matrix_bd_87_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_5012_cast_fp16 = add(x = matrix_ac_43_cast_fp16, y = matrix_bd_87_cast_fp16)[name = string("op_5012_cast_fp16")];
+            fp16 _inversed_scores_85_y_0_to_fp16 = const()[name = string("_inversed_scores_85_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_85_cast_fp16 = mul(x = var_5012_cast_fp16, y = _inversed_scores_85_y_0_to_fp16)[name = string("_inversed_scores_85_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_87_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_85_cast_fp16, cond = mask_11)[name = string("scores_87_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_5018_cast_fp16 = softmax(axis = var_60, x = scores_87_cast_fp16)[name = string("op_5018_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_1133_cast_fp16 = select(a = var_45_to_fp16, b = var_5018_cast_fp16, cond = mask_11)[name = string("input_1133_cast_fp16")];
+            bool x_559_transpose_x_0 = const()[name = string("x_559_transpose_x_0"), val = bool(false)];
+            bool x_559_transpose_y_0 = const()[name = string("x_559_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = v_43_cast_fp16)[name = string("transpose_170")];
+            tensor<fp16, [1, 8, 14, 128]> x_559_cast_fp16 = matmul(transpose_x = x_559_transpose_x_0, transpose_y = x_559_transpose_y_0, x = input_1133_cast_fp16, y = value_51_cast_fp16)[name = string("x_559_cast_fp16")];
+            tensor<int32, [4]> var_5022_perm_0 = const()[name = string("op_5022_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5023 = const()[name = string("op_5023"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_5022_cast_fp16 = transpose(perm = var_5022_perm_0, x = x_559_cast_fp16)[name = string("transpose_169")];
+            tensor<fp16, [1, 14, 1024]> input_1135_cast_fp16 = reshape(shape = var_5023, x = var_5022_cast_fp16)[name = string("input_1135_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443061184)))];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445158400)))];
+            tensor<fp16, [1, 14, 1024]> linear_196_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_out_weight_to_fp16, x = input_1135_cast_fp16)[name = string("linear_196_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1139_cast_fp16 = add(x = input_1129_cast_fp16, y = linear_196_cast_fp16)[name = string("input_1139_cast_fp16")];
+            tensor<int32, [1]> x_563_axes_0 = const()[name = string("x_563_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445160512)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445162624)))];
+            tensor<fp16, [1, 14, 1024]> x_563_cast_fp16 = layer_norm(axes = x_563_axes_0, beta = encoder_layers_21_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_conv_weight_to_fp16, x = input_1139_cast_fp16)[name = string("x_563_cast_fp16")];
+            tensor<int32, [3]> input_1141_perm_0 = const()[name = string("input_1141_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1143_pad_type_0 = const()[name = string("input_1143_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1143_strides_0 = const()[name = string("input_1143_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1143_pad_0 = const()[name = string("input_1143_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1143_dilations_0 = const()[name = string("input_1143_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1143_groups_0 = const()[name = string("input_1143_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445164736))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447261952))))[name = string("encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_1141_cast_fp16 = transpose(perm = input_1141_perm_0, x = x_563_cast_fp16)[name = string("transpose_168")];
+            tensor<fp16, [1, 2048, 14]> input_1143_cast_fp16 = conv(dilations = input_1143_dilations_0, groups = input_1143_groups_0, pad = input_1143_pad_0, pad_type = input_1143_pad_type_0, strides = input_1143_strides_0, weight = encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1141_cast_fp16)[name = string("input_1143_cast_fp16")];
+            int32 x_565_split_num_splits_0 = const()[name = string("x_565_split_num_splits_0"), val = int32(2)];
+            int32 x_565_split_axis_0 = const()[name = string("x_565_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_565_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_565_split_cast_fp16_1 = split(axis = x_565_split_axis_0, num_splits = x_565_split_num_splits_0, x = input_1143_cast_fp16)[name = string("x_565_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_565_split_1_sigmoid_cast_fp16 = sigmoid(x = x_565_split_cast_fp16_1)[name = string("x_565_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_565_cast_fp16 = mul(x = x_565_split_cast_fp16_0, y = x_565_split_1_sigmoid_cast_fp16)[name = string("x_565_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_1145_cast_fp16 = select(a = var_45_to_fp16, b = x_565_cast_fp16, cond = var_576)[name = string("input_1145_cast_fp16")];
+            bool new_x_87_interleave_0 = const()[name = string("new_x_87_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_87_cast_fp16 = concat(axis = var_60, interleave = new_x_87_interleave_0, values = (cache_87_cast_fp16, input_1145_cast_fp16))[name = string("new_x_87_cast_fp16")];
+            tensor<int32, [3]> var_5062_begin_0 = const()[name = string("op_5062_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_5062_end_0 = const()[name = string("op_5062_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_5062_end_mask_0 = const()[name = string("op_5062_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_5062_cast_fp16 = slice_by_index(begin = var_5062_begin_0, end = var_5062_end_0, end_mask = var_5062_end_mask_0, x = new_x_87_cast_fp16)[name = string("op_5062_cast_fp16")];
+            string x_567_pad_type_0 = const()[name = string("x_567_pad_type_0"), val = string("valid")];
+            int32 x_567_groups_0 = const()[name = string("x_567_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_567_strides_0 = const()[name = string("x_567_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_567_pad_0 = const()[name = string("x_567_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_567_dilations_0 = const()[name = string("x_567_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447266112))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447275392))))[name = string("encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_567_cast_fp16 = conv(dilations = x_567_dilations_0, groups = x_567_groups_0, pad = x_567_pad_0, pad_type = x_567_pad_type_0, strides = x_567_strides_0, weight = encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_87_cast_fp16)[name = string("x_567_cast_fp16")];
+            tensor<int32, [3]> input_1147_perm_0 = const()[name = string("input_1147_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_569_axes_0 = const()[name = string("x_569_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447277504)))];
+            tensor<fp16, [1024]> encoder_layers_21_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447279616)))];
+            tensor<fp16, [1, 14, 1024]> input_1147_cast_fp16 = transpose(perm = input_1147_perm_0, x = x_567_cast_fp16)[name = string("transpose_167")];
+            tensor<fp16, [1, 14, 1024]> x_569_cast_fp16 = layer_norm(axes = x_569_axes_0, beta = encoder_layers_21_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_conv_batch_norm_weight_to_fp16, x = input_1147_cast_fp16)[name = string("x_569_cast_fp16")];
+            tensor<int32, [3]> input_1149_perm_0 = const()[name = string("input_1149_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_1149_cast_fp16 = transpose(perm = input_1149_perm_0, x = x_569_cast_fp16)[name = string("transpose_166")];
+            tensor<fp16, [1, 1024, 14]> input_1151_cast_fp16 = silu(x = input_1149_cast_fp16)[name = string("input_1151_cast_fp16")];
+            string x_571_pad_type_0 = const()[name = string("x_571_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_571_strides_0 = const()[name = string("x_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_571_pad_0 = const()[name = string("x_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_571_dilations_0 = const()[name = string("x_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_571_groups_0 = const()[name = string("x_571_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447281728))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448330368))))[name = string("encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_571_cast_fp16 = conv(dilations = x_571_dilations_0, groups = x_571_groups_0, pad = x_571_pad_0, pad_type = x_571_pad_type_0, strides = x_571_strides_0, weight = encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1151_cast_fp16)[name = string("x_571_cast_fp16")];
+            tensor<int32, [3]> input_1153_perm_0 = const()[name = string("input_1153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_1153_cast_fp16 = transpose(perm = input_1153_perm_0, x = x_571_cast_fp16)[name = string("transpose_165")];
+            tensor<fp16, [1, 14, 1024]> input_1155_cast_fp16 = add(x = input_1139_cast_fp16, y = input_1153_cast_fp16)[name = string("input_1155_cast_fp16")];
+            tensor<int32, [1]> input_1157_axes_0 = const()[name = string("input_1157_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448332480)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448334592)))];
+            tensor<fp16, [1, 14, 1024]> input_1157_cast_fp16 = layer_norm(axes = input_1157_axes_0, beta = encoder_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_feed_forward2_weight_to_fp16, x = input_1155_cast_fp16)[name = string("input_1157_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_21_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448336704)))];
+            tensor<fp16, [4096]> encoder_layers_21_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456725376)))];
+            tensor<fp16, [1, 14, 4096]> linear_197_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear1_weight_to_fp16, x = input_1157_cast_fp16)[name = string("linear_197_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1161_cast_fp16 = silu(x = linear_197_cast_fp16)[name = string("input_1161_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_21_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456733632)))];
+            tensor<fp16, [1024]> encoder_layers_21_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465122304)))];
+            tensor<fp16, [1, 14, 1024]> linear_198_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear2_weight_to_fp16, x = input_1161_cast_fp16)[name = string("linear_198_cast_fp16")];
+            fp16 var_5105_to_fp16 = const()[name = string("op_5105_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_5106_cast_fp16 = mul(x = linear_198_cast_fp16, y = var_5105_to_fp16)[name = string("op_5106_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1167_cast_fp16 = add(x = input_1155_cast_fp16, y = var_5106_cast_fp16)[name = string("input_1167_cast_fp16")];
+            tensor<int32, [1]> input_1169_axes_0 = const()[name = string("input_1169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465124416)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465126528)))];
+            tensor<fp16, [1, 14, 1024]> input_1169_cast_fp16 = layer_norm(axes = input_1169_axes_0, beta = encoder_layers_21_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_out_weight_to_fp16, x = input_1167_cast_fp16)[name = string("input_1169_cast_fp16")];
+            tensor<int32, [4]> cache_89_begin_0 = const()[name = string("cache_89_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> cache_89_end_0 = const()[name = string("cache_89_end_0"), val = tensor<int32, [4]>([23, 1, 42, 1024])];
+            tensor<bool, [4]> cache_89_end_mask_0 = const()[name = string("cache_89_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_89_squeeze_mask_0 = const()[name = string("cache_89_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_89_cast_fp16 = slice_by_index(begin = cache_89_begin_0, end = cache_89_end_0, end_mask = cache_89_end_mask_0, squeeze_mask = cache_89_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_89_cast_fp16")];
+            tensor<int32, [4]> cache_91_begin_0 = const()[name = string("cache_91_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> cache_91_end_0 = const()[name = string("cache_91_end_0"), val = tensor<int32, [4]>([23, 1, 1024, 8])];
+            tensor<bool, [4]> cache_91_end_mask_0 = const()[name = string("cache_91_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_91_squeeze_mask_0 = const()[name = string("cache_91_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_91_cast_fp16 = slice_by_index(begin = cache_91_begin_0, end = cache_91_end_0, end_mask = cache_91_end_mask_0, squeeze_mask = cache_91_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_91_cast_fp16")];
+            tensor<int32, [1]> input_1171_axes_0 = const()[name = string("input_1171_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465128640)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465130752)))];
+            tensor<fp16, [1, 14, 1024]> input_1171_cast_fp16 = layer_norm(axes = input_1171_axes_0, beta = encoder_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_feed_forward1_weight_to_fp16, x = input_1169_cast_fp16)[name = string("input_1171_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_22_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465132864)))];
+            tensor<fp16, [4096]> encoder_layers_22_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473521536)))];
+            tensor<fp16, [1, 14, 4096]> linear_199_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear1_weight_to_fp16, x = input_1171_cast_fp16)[name = string("linear_199_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1175_cast_fp16 = silu(x = linear_199_cast_fp16)[name = string("input_1175_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_22_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473529792)))];
+            tensor<fp16, [1024]> encoder_layers_22_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481918464)))];
+            tensor<fp16, [1, 14, 1024]> linear_200_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear2_weight_to_fp16, x = input_1175_cast_fp16)[name = string("linear_200_cast_fp16")];
+            fp16 var_5142_to_fp16 = const()[name = string("op_5142_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_5143_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_5142_to_fp16)[name = string("op_5143_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1181_cast_fp16 = add(x = input_1169_cast_fp16, y = var_5143_cast_fp16)[name = string("input_1181_cast_fp16")];
+            tensor<int32, [1]> key_45_axes_0 = const()[name = string("key_45_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481920576)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481922688)))];
+            tensor<fp16, [1, 14, 1024]> key_45_cast_fp16 = layer_norm(axes = key_45_axes_0, beta = encoder_layers_22_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_self_att_weight_to_fp16, x = input_1181_cast_fp16)[name = string("key_45_cast_fp16")];
+            bool input_1183_interleave_0 = const()[name = string("input_1183_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_1183_cast_fp16 = concat(axis = var_69, interleave = input_1183_interleave_0, values = (cache_89_cast_fp16, key_45_cast_fp16))[name = string("input_1183_cast_fp16")];
+            tensor<int32, [3]> var_5165_begin_0 = const()[name = string("op_5165_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_5165_end_0 = const()[name = string("op_5165_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_5165_end_mask_0 = const()[name = string("op_5165_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_5165_cast_fp16 = slice_by_index(begin = var_5165_begin_0, end = var_5165_end_0, end_mask = var_5165_end_mask_0, x = cache_89_cast_fp16)[name = string("op_5165_cast_fp16")];
+            bool var_5171_interleave_0 = const()[name = string("op_5171_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_5171_cast_fp16 = concat(axis = var_69, interleave = var_5171_interleave_0, values = (var_5165_cast_fp16, key_45_cast_fp16))[name = string("op_5171_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481924800)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484022016)))];
+            tensor<fp16, [1, 14, 1024]> linear_201_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_q_weight_to_fp16, x = key_45_cast_fp16)[name = string("linear_201_cast_fp16")];
+            tensor<int32, [4]> var_5176 = const()[name = string("op_5176"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_133_cast_fp16 = reshape(shape = var_5176, x = linear_201_cast_fp16)[name = string("q_133_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484024128)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486121344)))];
+            tensor<fp16, [1, 56, 1024]> linear_202_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_k_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_202_cast_fp16")];
+            tensor<int32, [4]> var_5181 = const()[name = string("op_5181"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_89_cast_fp16 = reshape(shape = var_5181, x = linear_202_cast_fp16)[name = string("k_89_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486123456)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488220672)))];
+            tensor<fp16, [1, 56, 1024]> linear_203_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_v_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_203_cast_fp16")];
+            tensor<int32, [4]> var_5186 = const()[name = string("op_5186"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_45_cast_fp16 = reshape(shape = var_5186, x = linear_203_cast_fp16)[name = string("v_45_cast_fp16")];
+            tensor<int32, [4]> value_53_perm_0 = const()[name = string("value_53_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_22_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488222784)))];
+            tensor<fp16, [1, 14, 8, 128]> var_5199_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_u_to_fp16)[name = string("op_5199_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_22_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488224896)))];
+            tensor<fp16, [1, 14, 8, 128]> var_5201_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_v_to_fp16)[name = string("op_5201_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_45_perm_0 = const()[name = string("q_with_bias_v_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_579_transpose_x_0 = const()[name = string("x_579_transpose_x_0"), val = bool(false)];
+            bool x_579_transpose_y_0 = const()[name = string("x_579_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_5203_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488227008))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488340736))))[name = string("op_5203_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_45_cast_fp16 = transpose(perm = q_with_bias_v_45_perm_0, x = var_5201_cast_fp16)[name = string("transpose_164")];
+            tensor<fp16, [1, 8, 14, 111]> x_579_cast_fp16 = matmul(transpose_x = x_579_transpose_x_0, transpose_y = x_579_transpose_y_0, x = q_with_bias_v_45_cast_fp16, y = op_5203_to_fp16_quantized)[name = string("x_579_cast_fp16")];
+            tensor<int32, [8]> x_581_pad_0 = const()[name = string("x_581_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_581_mode_0 = const()[name = string("x_581_mode_0"), val = string("constant")];
+            fp16 const_365_to_fp16 = const()[name = string("const_365_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_581_cast_fp16 = pad(constant_val = const_365_to_fp16, mode = x_581_mode_0, pad = x_581_pad_0, x = x_579_cast_fp16)[name = string("x_581_cast_fp16")];
+            tensor<int32, [4]> var_5211 = const()[name = string("op_5211"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_583_cast_fp16 = reshape(shape = var_5211, x = x_581_cast_fp16)[name = string("x_583_cast_fp16")];
+            tensor<int32, [4]> var_5215_begin_0 = const()[name = string("op_5215_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5215_end_0 = const()[name = string("op_5215_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_5215_end_mask_0 = const()[name = string("op_5215_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_5215_cast_fp16 = slice_by_index(begin = var_5215_begin_0, end = var_5215_end_0, end_mask = var_5215_end_mask_0, x = x_583_cast_fp16)[name = string("op_5215_cast_fp16")];
+            tensor<int32, [4]> var_5216 = const()[name = string("op_5216"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_89_cast_fp16 = reshape(shape = var_5216, x = var_5215_cast_fp16)[name = string("matrix_bd_89_cast_fp16")];
+            bool matrix_ac_45_transpose_x_0 = const()[name = string("matrix_ac_45_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_45_transpose_y_0 = const()[name = string("matrix_ac_45_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_140_perm_0 = const()[name = string("transpose_140_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_141_perm_0 = const()[name = string("transpose_141_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89_cast_fp16)[name = string("transpose_162")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_140 = transpose(perm = transpose_140_perm_0, x = var_5199_cast_fp16)[name = string("transpose_163")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_45_cast_fp16 = matmul(transpose_x = matrix_ac_45_transpose_x_0, transpose_y = matrix_ac_45_transpose_y_0, x = transpose_140, y = transpose_141)[name = string("matrix_ac_45_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_91_begin_0 = const()[name = string("matrix_bd_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_91_end_0 = const()[name = string("matrix_bd_91_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_91_end_mask_0 = const()[name = string("matrix_bd_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_91_cast_fp16 = slice_by_index(begin = matrix_bd_91_begin_0, end = matrix_bd_91_end_0, end_mask = matrix_bd_91_end_mask_0, x = matrix_bd_89_cast_fp16)[name = string("matrix_bd_91_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_5225_cast_fp16 = add(x = matrix_ac_45_cast_fp16, y = matrix_bd_91_cast_fp16)[name = string("op_5225_cast_fp16")];
+            fp16 _inversed_scores_89_y_0_to_fp16 = const()[name = string("_inversed_scores_89_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_89_cast_fp16 = mul(x = var_5225_cast_fp16, y = _inversed_scores_89_y_0_to_fp16)[name = string("_inversed_scores_89_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_91_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_89_cast_fp16, cond = mask_11)[name = string("scores_91_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_5231_cast_fp16 = softmax(axis = var_60, x = scores_91_cast_fp16)[name = string("op_5231_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_1185_cast_fp16 = select(a = var_45_to_fp16, b = var_5231_cast_fp16, cond = mask_11)[name = string("input_1185_cast_fp16")];
+            bool x_585_transpose_x_0 = const()[name = string("x_585_transpose_x_0"), val = bool(false)];
+            bool x_585_transpose_y_0 = const()[name = string("x_585_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = v_45_cast_fp16)[name = string("transpose_161")];
+            tensor<fp16, [1, 8, 14, 128]> x_585_cast_fp16 = matmul(transpose_x = x_585_transpose_x_0, transpose_y = x_585_transpose_y_0, x = input_1185_cast_fp16, y = value_53_cast_fp16)[name = string("x_585_cast_fp16")];
+            tensor<int32, [4]> var_5235_perm_0 = const()[name = string("op_5235_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5236 = const()[name = string("op_5236"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_5235_cast_fp16 = transpose(perm = var_5235_perm_0, x = x_585_cast_fp16)[name = string("transpose_160")];
+            tensor<fp16, [1, 14, 1024]> input_1187_cast_fp16 = reshape(shape = var_5236, x = var_5235_cast_fp16)[name = string("input_1187_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488341056)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490438272)))];
+            tensor<fp16, [1, 14, 1024]> linear_205_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_out_weight_to_fp16, x = input_1187_cast_fp16)[name = string("linear_205_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1191_cast_fp16 = add(x = input_1181_cast_fp16, y = linear_205_cast_fp16)[name = string("input_1191_cast_fp16")];
+            tensor<int32, [1]> x_589_axes_0 = const()[name = string("x_589_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490440384)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490442496)))];
+            tensor<fp16, [1, 14, 1024]> x_589_cast_fp16 = layer_norm(axes = x_589_axes_0, beta = encoder_layers_22_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_conv_weight_to_fp16, x = input_1191_cast_fp16)[name = string("x_589_cast_fp16")];
+            tensor<int32, [3]> input_1193_perm_0 = const()[name = string("input_1193_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1195_pad_type_0 = const()[name = string("input_1195_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1195_strides_0 = const()[name = string("input_1195_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1195_pad_0 = const()[name = string("input_1195_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1195_dilations_0 = const()[name = string("input_1195_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1195_groups_0 = const()[name = string("input_1195_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490444608))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492541824))))[name = string("encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_1193_cast_fp16 = transpose(perm = input_1193_perm_0, x = x_589_cast_fp16)[name = string("transpose_159")];
+            tensor<fp16, [1, 2048, 14]> input_1195_cast_fp16 = conv(dilations = input_1195_dilations_0, groups = input_1195_groups_0, pad = input_1195_pad_0, pad_type = input_1195_pad_type_0, strides = input_1195_strides_0, weight = encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1193_cast_fp16)[name = string("input_1195_cast_fp16")];
+            int32 x_591_split_num_splits_0 = const()[name = string("x_591_split_num_splits_0"), val = int32(2)];
+            int32 x_591_split_axis_0 = const()[name = string("x_591_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_591_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_591_split_cast_fp16_1 = split(axis = x_591_split_axis_0, num_splits = x_591_split_num_splits_0, x = input_1195_cast_fp16)[name = string("x_591_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_591_split_1_sigmoid_cast_fp16 = sigmoid(x = x_591_split_cast_fp16_1)[name = string("x_591_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_591_cast_fp16 = mul(x = x_591_split_cast_fp16_0, y = x_591_split_1_sigmoid_cast_fp16)[name = string("x_591_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_1197_cast_fp16 = select(a = var_45_to_fp16, b = x_591_cast_fp16, cond = var_576)[name = string("input_1197_cast_fp16")];
+            bool new_x_91_interleave_0 = const()[name = string("new_x_91_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_91_cast_fp16 = concat(axis = var_60, interleave = new_x_91_interleave_0, values = (cache_91_cast_fp16, input_1197_cast_fp16))[name = string("new_x_91_cast_fp16")];
+            tensor<int32, [3]> var_5275_begin_0 = const()[name = string("op_5275_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_5275_end_0 = const()[name = string("op_5275_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_5275_end_mask_0 = const()[name = string("op_5275_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_5275_cast_fp16 = slice_by_index(begin = var_5275_begin_0, end = var_5275_end_0, end_mask = var_5275_end_mask_0, x = new_x_91_cast_fp16)[name = string("op_5275_cast_fp16")];
+            string x_593_pad_type_0 = const()[name = string("x_593_pad_type_0"), val = string("valid")];
+            int32 x_593_groups_0 = const()[name = string("x_593_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_593_strides_0 = const()[name = string("x_593_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_593_pad_0 = const()[name = string("x_593_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_593_dilations_0 = const()[name = string("x_593_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492545984))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492555264))))[name = string("encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_593_cast_fp16 = conv(dilations = x_593_dilations_0, groups = x_593_groups_0, pad = x_593_pad_0, pad_type = x_593_pad_type_0, strides = x_593_strides_0, weight = encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_91_cast_fp16)[name = string("x_593_cast_fp16")];
+            tensor<int32, [3]> input_1199_perm_0 = const()[name = string("input_1199_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_595_axes_0 = const()[name = string("x_595_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492557376)))];
+            tensor<fp16, [1024]> encoder_layers_22_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492559488)))];
+            tensor<fp16, [1, 14, 1024]> input_1199_cast_fp16 = transpose(perm = input_1199_perm_0, x = x_593_cast_fp16)[name = string("transpose_158")];
+            tensor<fp16, [1, 14, 1024]> x_595_cast_fp16 = layer_norm(axes = x_595_axes_0, beta = encoder_layers_22_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_conv_batch_norm_weight_to_fp16, x = input_1199_cast_fp16)[name = string("x_595_cast_fp16")];
+            tensor<int32, [3]> input_1201_perm_0 = const()[name = string("input_1201_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_1201_cast_fp16 = transpose(perm = input_1201_perm_0, x = x_595_cast_fp16)[name = string("transpose_157")];
+            tensor<fp16, [1, 1024, 14]> input_1203_cast_fp16 = silu(x = input_1201_cast_fp16)[name = string("input_1203_cast_fp16")];
+            string x_597_pad_type_0 = const()[name = string("x_597_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_597_strides_0 = const()[name = string("x_597_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_597_pad_0 = const()[name = string("x_597_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_597_dilations_0 = const()[name = string("x_597_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_597_groups_0 = const()[name = string("x_597_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492561600))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493610240))))[name = string("encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_597_cast_fp16 = conv(dilations = x_597_dilations_0, groups = x_597_groups_0, pad = x_597_pad_0, pad_type = x_597_pad_type_0, strides = x_597_strides_0, weight = encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1203_cast_fp16)[name = string("x_597_cast_fp16")];
+            tensor<int32, [3]> input_1205_perm_0 = const()[name = string("input_1205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_1205_cast_fp16 = transpose(perm = input_1205_perm_0, x = x_597_cast_fp16)[name = string("transpose_156")];
+            tensor<fp16, [1, 14, 1024]> input_1207_cast_fp16 = add(x = input_1191_cast_fp16, y = input_1205_cast_fp16)[name = string("input_1207_cast_fp16")];
+            tensor<int32, [1]> input_1209_axes_0 = const()[name = string("input_1209_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493612352)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493614464)))];
+            tensor<fp16, [1, 14, 1024]> input_1209_cast_fp16 = layer_norm(axes = input_1209_axes_0, beta = encoder_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_feed_forward2_weight_to_fp16, x = input_1207_cast_fp16)[name = string("input_1209_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_22_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493616576)))];
+            tensor<fp16, [4096]> encoder_layers_22_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502005248)))];
+            tensor<fp16, [1, 14, 4096]> linear_206_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear1_weight_to_fp16, x = input_1209_cast_fp16)[name = string("linear_206_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1213_cast_fp16 = silu(x = linear_206_cast_fp16)[name = string("input_1213_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_22_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502013504)))];
+            tensor<fp16, [1024]> encoder_layers_22_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510402176)))];
+            tensor<fp16, [1, 14, 1024]> linear_207_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear2_weight_to_fp16, x = input_1213_cast_fp16)[name = string("linear_207_cast_fp16")];
+            fp16 var_5318_to_fp16 = const()[name = string("op_5318_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_5319_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_5318_to_fp16)[name = string("op_5319_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1219_cast_fp16 = add(x = input_1207_cast_fp16, y = var_5319_cast_fp16)[name = string("input_1219_cast_fp16")];
+            tensor<int32, [1]> input_1221_axes_0 = const()[name = string("input_1221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510404288)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510406400)))];
+            tensor<fp16, [1, 14, 1024]> input_1221_cast_fp16 = layer_norm(axes = input_1221_axes_0, beta = encoder_layers_22_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_out_weight_to_fp16, x = input_1219_cast_fp16)[name = string("input_1221_cast_fp16")];
+            tensor<int32, [4]> cache_93_begin_0 = const()[name = string("cache_93_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> cache_93_end_0 = const()[name = string("cache_93_end_0"), val = tensor<int32, [4]>([24, 1, 42, 1024])];
+            tensor<bool, [4]> cache_93_end_mask_0 = const()[name = string("cache_93_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_93_squeeze_mask_0 = const()[name = string("cache_93_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_93_cast_fp16 = slice_by_index(begin = cache_93_begin_0, end = cache_93_end_0, end_mask = cache_93_end_mask_0, squeeze_mask = cache_93_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_93_cast_fp16")];
+            tensor<int32, [4]> cache_begin_0 = const()[name = string("cache_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> cache_end_0 = const()[name = string("cache_end_0"), val = tensor<int32, [4]>([24, 1, 1024, 8])];
+            tensor<bool, [4]> cache_end_mask_0 = const()[name = string("cache_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_squeeze_mask_0 = const()[name = string("cache_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_cast_fp16")];
+            tensor<int32, [1]> input_1223_axes_0 = const()[name = string("input_1223_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510408512)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510410624)))];
+            tensor<fp16, [1, 14, 1024]> input_1223_cast_fp16 = layer_norm(axes = input_1223_axes_0, beta = encoder_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_feed_forward1_weight_to_fp16, x = input_1221_cast_fp16)[name = string("input_1223_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_23_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510412736)))];
+            tensor<fp16, [4096]> encoder_layers_23_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518801408)))];
+            tensor<fp16, [1, 14, 4096]> linear_208_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear1_weight_to_fp16, x = input_1223_cast_fp16)[name = string("linear_208_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1227_cast_fp16 = silu(x = linear_208_cast_fp16)[name = string("input_1227_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_23_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518809664)))];
+            tensor<fp16, [1024]> encoder_layers_23_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527198336)))];
+            tensor<fp16, [1, 14, 1024]> linear_209_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear2_weight_to_fp16, x = input_1227_cast_fp16)[name = string("linear_209_cast_fp16")];
+            fp16 var_5355_to_fp16 = const()[name = string("op_5355_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_5356_cast_fp16 = mul(x = linear_209_cast_fp16, y = var_5355_to_fp16)[name = string("op_5356_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1233_cast_fp16 = add(x = input_1221_cast_fp16, y = var_5356_cast_fp16)[name = string("input_1233_cast_fp16")];
+            tensor<int32, [1]> key_axes_0 = const()[name = string("key_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527200448)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527202560)))];
+            tensor<fp16, [1, 14, 1024]> key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_23_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_self_att_weight_to_fp16, x = input_1233_cast_fp16)[name = string("key_cast_fp16")];
+            bool input_1235_interleave_0 = const()[name = string("input_1235_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_1235_cast_fp16 = concat(axis = var_69, interleave = input_1235_interleave_0, values = (cache_93_cast_fp16, key_cast_fp16))[name = string("input_1235_cast_fp16")];
+            tensor<int32, [3]> var_5378_begin_0 = const()[name = string("op_5378_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_5378_end_0 = const()[name = string("op_5378_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_5378_end_mask_0 = const()[name = string("op_5378_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_5378_cast_fp16 = slice_by_index(begin = var_5378_begin_0, end = var_5378_end_0, end_mask = var_5378_end_mask_0, x = cache_93_cast_fp16)[name = string("op_5378_cast_fp16")];
+            bool cache_last_channel_cur_interleave_0 = const()[name = string("cache_last_channel_cur_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> cache_last_channel_cur_cast_fp16 = concat(axis = var_69, interleave = cache_last_channel_cur_interleave_0, values = (var_5378_cast_fp16, key_cast_fp16))[name = string("cache_last_channel_cur_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527204672)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529301888)))];
+            tensor<fp16, [1, 14, 1024]> linear_210_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = string("linear_210_cast_fp16")];
+            tensor<int32, [4]> var_5389 = const()[name = string("op_5389"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_139_cast_fp16 = reshape(shape = var_5389, x = linear_210_cast_fp16)[name = string("q_139_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529304000)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531401216)))];
+            tensor<fp16, [1, 56, 1024]> linear_211_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_k_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_211_cast_fp16")];
+            tensor<int32, [4]> var_5394 = const()[name = string("op_5394"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_93_cast_fp16 = reshape(shape = var_5394, x = linear_211_cast_fp16)[name = string("k_93_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531403328)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533500544)))];
+            tensor<fp16, [1, 56, 1024]> linear_212_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_v_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_212_cast_fp16")];
+            tensor<int32, [4]> var_5399 = const()[name = string("op_5399"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_cast_fp16 = reshape(shape = var_5399, x = linear_212_cast_fp16)[name = string("v_cast_fp16")];
+            tensor<int32, [4]> value_perm_0 = const()[name = string("value_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_23_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533502656)))];
+            tensor<fp16, [1, 14, 8, 128]> var_5412_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_u_to_fp16)[name = string("op_5412_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_23_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533504768)))];
+            tensor<fp16, [1, 14, 8, 128]> var_5414_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_v_to_fp16)[name = string("op_5414_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_perm_0 = const()[name = string("q_with_bias_v_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_605_transpose_x_0 = const()[name = string("x_605_transpose_x_0"), val = bool(false)];
+            bool x_605_transpose_y_0 = const()[name = string("x_605_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_5416_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533506880))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533620608))))[name = string("op_5416_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_5414_cast_fp16)[name = string("transpose_155")];
+            tensor<fp16, [1, 8, 14, 111]> x_605_cast_fp16 = matmul(transpose_x = x_605_transpose_x_0, transpose_y = x_605_transpose_y_0, x = q_with_bias_v_cast_fp16, y = op_5416_to_fp16_quantized)[name = string("x_605_cast_fp16")];
+            tensor<int32, [8]> x_607_pad_0 = const()[name = string("x_607_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_607_mode_0 = const()[name = string("x_607_mode_0"), val = string("constant")];
+            fp16 const_378_to_fp16 = const()[name = string("const_378_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_607_cast_fp16 = pad(constant_val = const_378_to_fp16, mode = x_607_mode_0, pad = x_607_pad_0, x = x_605_cast_fp16)[name = string("x_607_cast_fp16")];
+            tensor<int32, [4]> var_5424 = const()[name = string("op_5424"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_609_cast_fp16 = reshape(shape = var_5424, x = x_607_cast_fp16)[name = string("x_609_cast_fp16")];
+            tensor<int32, [4]> var_5428_begin_0 = const()[name = string("op_5428_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5428_end_0 = const()[name = string("op_5428_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_5428_end_mask_0 = const()[name = string("op_5428_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_5428_cast_fp16 = slice_by_index(begin = var_5428_begin_0, end = var_5428_end_0, end_mask = var_5428_end_mask_0, x = x_609_cast_fp16)[name = string("op_5428_cast_fp16")];
+            tensor<int32, [4]> var_5429 = const()[name = string("op_5429"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_93_cast_fp16 = reshape(shape = var_5429, x = var_5428_cast_fp16)[name = string("matrix_bd_93_cast_fp16")];
+            bool matrix_ac_transpose_x_0 = const()[name = string("matrix_ac_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_transpose_y_0 = const()[name = string("matrix_ac_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_142_perm_0 = const()[name = string("transpose_142_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_143_perm_0 = const()[name = string("transpose_143_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93_cast_fp16)[name = string("transpose_153")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_142 = transpose(perm = transpose_142_perm_0, x = var_5412_cast_fp16)[name = string("transpose_154")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_142, y = transpose_143)[name = string("matrix_ac_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_begin_0 = const()[name = string("matrix_bd_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_end_0 = const()[name = string("matrix_bd_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_end_mask_0 = const()[name = string("matrix_bd_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_93_cast_fp16)[name = string("matrix_bd_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_5438_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = string("op_5438_cast_fp16")];
+            fp16 _inversed_scores_93_y_0_to_fp16 = const()[name = string("_inversed_scores_93_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_93_cast_fp16 = mul(x = var_5438_cast_fp16, y = _inversed_scores_93_y_0_to_fp16)[name = string("_inversed_scores_93_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_93_cast_fp16, cond = mask_11)[name = string("scores_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_5444_cast_fp16 = softmax(axis = var_60, x = scores_cast_fp16)[name = string("op_5444_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_1237_cast_fp16 = select(a = var_45_to_fp16, b = var_5444_cast_fp16, cond = mask_11)[name = string("input_1237_cast_fp16")];
+            bool x_611_transpose_x_0 = const()[name = string("x_611_transpose_x_0"), val = bool(false)];
+            bool x_611_transpose_y_0 = const()[name = string("x_611_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = string("transpose_152")];
+            tensor<fp16, [1, 8, 14, 128]> x_611_cast_fp16 = matmul(transpose_x = x_611_transpose_x_0, transpose_y = x_611_transpose_y_0, x = input_1237_cast_fp16, y = value_cast_fp16)[name = string("x_611_cast_fp16")];
+            tensor<int32, [4]> var_5448_perm_0 = const()[name = string("op_5448_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5449 = const()[name = string("op_5449"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_5448_cast_fp16 = transpose(perm = var_5448_perm_0, x = x_611_cast_fp16)[name = string("transpose_151")];
+            tensor<fp16, [1, 14, 1024]> input_1239_cast_fp16 = reshape(shape = var_5449, x = var_5448_cast_fp16)[name = string("input_1239_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533620928)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535718144)))];
+            tensor<fp16, [1, 14, 1024]> linear_214_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_out_weight_to_fp16, x = input_1239_cast_fp16)[name = string("linear_214_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1243_cast_fp16 = add(x = input_1233_cast_fp16, y = linear_214_cast_fp16)[name = string("input_1243_cast_fp16")];
+            tensor<int32, [1]> x_615_axes_0 = const()[name = string("x_615_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535720256)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535722368)))];
+            tensor<fp16, [1, 14, 1024]> x_615_cast_fp16 = layer_norm(axes = x_615_axes_0, beta = encoder_layers_23_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_conv_weight_to_fp16, x = input_1243_cast_fp16)[name = string("x_615_cast_fp16")];
+            tensor<int32, [3]> input_1245_perm_0 = const()[name = string("input_1245_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1247_pad_type_0 = const()[name = string("input_1247_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1247_strides_0 = const()[name = string("input_1247_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1247_pad_0 = const()[name = string("input_1247_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1247_dilations_0 = const()[name = string("input_1247_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1247_groups_0 = const()[name = string("input_1247_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535724480))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537821696))))[name = string("encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_1245_cast_fp16 = transpose(perm = input_1245_perm_0, x = x_615_cast_fp16)[name = string("transpose_150")];
+            tensor<fp16, [1, 2048, 14]> input_1247_cast_fp16 = conv(dilations = input_1247_dilations_0, groups = input_1247_groups_0, pad = input_1247_pad_0, pad_type = input_1247_pad_type_0, strides = input_1247_strides_0, weight = encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1245_cast_fp16)[name = string("input_1247_cast_fp16")];
+            int32 x_617_split_num_splits_0 = const()[name = string("x_617_split_num_splits_0"), val = int32(2)];
+            int32 x_617_split_axis_0 = const()[name = string("x_617_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_617_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_617_split_cast_fp16_1 = split(axis = x_617_split_axis_0, num_splits = x_617_split_num_splits_0, x = input_1247_cast_fp16)[name = string("x_617_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_617_split_1_sigmoid_cast_fp16 = sigmoid(x = x_617_split_cast_fp16_1)[name = string("x_617_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_617_cast_fp16 = mul(x = x_617_split_cast_fp16_0, y = x_617_split_1_sigmoid_cast_fp16)[name = string("x_617_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_1249_cast_fp16 = select(a = var_45_to_fp16, b = x_617_cast_fp16, cond = var_576)[name = string("input_1249_cast_fp16")];
+            bool new_x_interleave_0 = const()[name = string("new_x_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_cast_fp16 = concat(axis = var_60, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_1249_cast_fp16))[name = string("new_x_cast_fp16")];
+            tensor<int32, [3]> cache_last_time_cur_begin_0 = const()[name = string("cache_last_time_cur_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> cache_last_time_cur_end_0 = const()[name = string("cache_last_time_cur_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> cache_last_time_cur_end_mask_0 = const()[name = string("cache_last_time_cur_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = new_x_cast_fp16)[name = string("cache_last_time_cur_cast_fp16")];
+            string x_619_pad_type_0 = const()[name = string("x_619_pad_type_0"), val = string("valid")];
+            int32 x_619_groups_0 = const()[name = string("x_619_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_619_strides_0 = const()[name = string("x_619_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_619_pad_0 = const()[name = string("x_619_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_619_dilations_0 = const()[name = string("x_619_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537825856))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537835136))))[name = string("encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_619_cast_fp16 = conv(dilations = x_619_dilations_0, groups = x_619_groups_0, pad = x_619_pad_0, pad_type = x_619_pad_type_0, strides = x_619_strides_0, weight = encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_cast_fp16)[name = string("x_619_cast_fp16")];
+            tensor<int32, [3]> input_1251_perm_0 = const()[name = string("input_1251_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_621_axes_0 = const()[name = string("x_621_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537837248)))];
+            tensor<fp16, [1024]> encoder_layers_23_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537839360)))];
+            tensor<fp16, [1, 14, 1024]> input_1251_cast_fp16 = transpose(perm = input_1251_perm_0, x = x_619_cast_fp16)[name = string("transpose_149")];
+            tensor<fp16, [1, 14, 1024]> x_621_cast_fp16 = layer_norm(axes = x_621_axes_0, beta = encoder_layers_23_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_conv_batch_norm_weight_to_fp16, x = input_1251_cast_fp16)[name = string("x_621_cast_fp16")];
+            tensor<int32, [3]> input_1253_perm_0 = const()[name = string("input_1253_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_1253_cast_fp16 = transpose(perm = input_1253_perm_0, x = x_621_cast_fp16)[name = string("transpose_148")];
+            tensor<fp16, [1, 1024, 14]> input_1255_cast_fp16 = silu(x = input_1253_cast_fp16)[name = string("input_1255_cast_fp16")];
+            string x_623_pad_type_0 = const()[name = string("x_623_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_623_strides_0 = const()[name = string("x_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_623_pad_0 = const()[name = string("x_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_623_dilations_0 = const()[name = string("x_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_623_groups_0 = const()[name = string("x_623_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537841472))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538890112))))[name = string("encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_623_cast_fp16 = conv(dilations = x_623_dilations_0, groups = x_623_groups_0, pad = x_623_pad_0, pad_type = x_623_pad_type_0, strides = x_623_strides_0, weight = encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1255_cast_fp16)[name = string("x_623_cast_fp16")];
+            tensor<int32, [3]> input_1257_perm_0 = const()[name = string("input_1257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_1257_cast_fp16 = transpose(perm = input_1257_perm_0, x = x_623_cast_fp16)[name = string("transpose_147")];
+            tensor<fp16, [1, 14, 1024]> input_1259_cast_fp16 = add(x = input_1243_cast_fp16, y = input_1257_cast_fp16)[name = string("input_1259_cast_fp16")];
+            tensor<int32, [1]> input_1261_axes_0 = const()[name = string("input_1261_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538892224)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538894336)))];
+            tensor<fp16, [1, 14, 1024]> input_1261_cast_fp16 = layer_norm(axes = input_1261_axes_0, beta = encoder_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_feed_forward2_weight_to_fp16, x = input_1259_cast_fp16)[name = string("input_1261_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_23_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538896448)))];
+            tensor<fp16, [4096]> encoder_layers_23_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547285120)))];
+            tensor<fp16, [1, 14, 4096]> linear_215_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear1_weight_to_fp16, x = input_1261_cast_fp16)[name = string("linear_215_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1265_cast_fp16 = silu(x = linear_215_cast_fp16)[name = string("input_1265_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_23_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547293376)))];
+            tensor<fp16, [1024]> encoder_layers_23_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555682048)))];
+            tensor<fp16, [1, 14, 1024]> linear_216_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear2_weight_to_fp16, x = input_1265_cast_fp16)[name = string("linear_216_cast_fp16")];
+            fp16 var_5531_to_fp16 = const()[name = string("op_5531_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_5532_cast_fp16 = mul(x = linear_216_cast_fp16, y = var_5531_to_fp16)[name = string("op_5532_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1271_cast_fp16 = add(x = input_1259_cast_fp16, y = var_5532_cast_fp16)[name = string("input_1271_cast_fp16")];
+            tensor<int32, [1]> audio_signal_axes_0 = const()[name = string("audio_signal_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555684160)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555686272)))];
+            tensor<fp16, [1, 14, 1024]> audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_23_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_out_weight_to_fp16, x = input_1271_cast_fp16)[name = string("audio_signal_cast_fp16")];
+            int32 obj_5_axis_0 = const()[name = string("obj_5_axis_0"), val = int32(0)];
+            tensor<fp16, [24, 1, 42, 1024]> obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_485_cast_fp16, var_698_cast_fp16, var_911_cast_fp16, var_1124_cast_fp16, var_1337_cast_fp16, var_1550_cast_fp16, var_1763_cast_fp16, var_1976_cast_fp16, var_2189_cast_fp16, var_2402_cast_fp16, var_2615_cast_fp16, var_2828_cast_fp16, var_3041_cast_fp16, var_3254_cast_fp16, var_3467_cast_fp16, var_3680_cast_fp16, var_3893_cast_fp16, var_4106_cast_fp16, var_4319_cast_fp16, var_4532_cast_fp16, var_4745_cast_fp16, var_4958_cast_fp16, var_5171_cast_fp16, cache_last_channel_cur_cast_fp16))[name = string("obj_5_cast_fp16")];
+            int32 obj_7_axis_0 = const()[name = string("obj_7_axis_0"), val = int32(0)];
+            tensor<fp16, [24, 1, 1024, 8]> obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_589_cast_fp16, var_802_cast_fp16, var_1015_cast_fp16, var_1228_cast_fp16, var_1441_cast_fp16, var_1654_cast_fp16, var_1867_cast_fp16, var_2080_cast_fp16, var_2293_cast_fp16, var_2506_cast_fp16, var_2719_cast_fp16, var_2932_cast_fp16, var_3145_cast_fp16, var_3358_cast_fp16, var_3571_cast_fp16, var_3784_cast_fp16, var_3997_cast_fp16, var_4210_cast_fp16, var_4423_cast_fp16, var_4636_cast_fp16, var_4849_cast_fp16, var_5062_cast_fp16, var_5275_cast_fp16, cache_last_time_cur_cast_fp16))[name = string("obj_7_cast_fp16")];
+            tensor<int32, [1]> var_5548 = add(x = cache_len, y = max_audio_length_1)[name = string("op_5548")];
+            string var_5548_promoted_to_fp16_dtype_0 = const()[name = string("op_5548_promoted_to_fp16_dtype_0"), val = string("fp16")];
+            fp16 const_384_to_fp16 = const()[name = string("const_384_to_fp16"), val = fp16(-inf)];
+            fp16 var_50_promoted_to_fp16 = const()[name = string("op_50_promoted_to_fp16"), val = fp16(0x1.5p+5)];
+            tensor<fp16, [1]> var_5548_to_fp16 = cast(dtype = var_5548_promoted_to_fp16_dtype_0, x = var_5548)[name = string("cast_10")];
+            tensor<fp16, [1]> clip_1_cast_fp16 = clip(alpha = const_384_to_fp16, beta = var_50_promoted_to_fp16, x = var_5548_to_fp16)[name = string("clip_1_cast_fp16")];
+            int32 one_hot_1_batch_dims_0 = const()[name = string("one_hot_1_batch_dims_0"), val = int32(0)];
+            bool one_hot_1_validate_indices_0 = const()[name = string("one_hot_1_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [128, 128]> to_onehot_identity_table_to_fp16 = const()[name = string("to_onehot_identity_table_to_fp16"), val = tensor<fp16, [128, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555688384)))];
+            string prompt_id_to_int16_dtype_0 = const()[name = string("prompt_id_to_int16_dtype_0"), val = string("int16")];
+            string cast_230_dtype_0 = const()[name = string("cast_230_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
+            tensor<int16, [1]> prompt_id_to_int16 = cast(dtype = prompt_id_to_int16_dtype_0, x = prompt_id)[name = string("cast_9")];
+            tensor<int32, [1]> cast_230 = cast(dtype = cast_230_dtype_0, x = prompt_id_to_int16)[name = string("cast_8")];
+            tensor<bool, [1]> greater_equal_0 = greater_equal(x = cast_230, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
+            int32 slice_by_index_2 = const()[name = string("slice_by_index_2"), val = int32(128)];
+            tensor<int32, [1]> add_0 = add(x = cast_230, y = slice_by_index_2)[name = string("add_0")];
+            tensor<int32, [1]> select_0 = select(a = cast_230, b = add_0, cond = greater_equal_0)[name = string("select_0")];
+            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
+            string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
+            tensor<int16, [1]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_7")];
+            tensor<int32, [1]> cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_6")];
+            tensor<bool, [1]> greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
+            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(128)];
+            tensor<int32, [1]> add_0_1 = add(x = cast_0, y = slice_by_index_0)[name = string("add_0_1")];
+            tensor<int32, [1]> select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")];
+            int32 greater_equal_0_y_0_2 = const()[name = string("greater_equal_0_y_0_2"), val = int32(0)];
+            tensor<bool, [1]> greater_equal_0_2 = greater_equal(x = select_0_1, y = greater_equal_0_y_0_2)[name = string("greater_equal_0_2")];
+            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(128)];
+            tensor<int32, [1]> add_0_2 = add(x = select_0_1, y = slice_by_index_0_1)[name = string("add_0_2")];
+            tensor<int32, [1]> select_0_2 = select(a = select_0_1, b = add_0_2, cond = greater_equal_0_2)[name = string("select_0_2")];
+            int32 one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(0)];
+            tensor<fp16, [1, 128]> one_hot_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = one_hot_1_batch_dims_0, indices = select_0_2, validate_indices = one_hot_1_validate_indices_0, x = to_onehot_identity_table_to_fp16)[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16")];
+            tensor<int32, [1]> var_5594_axes_0 = const()[name = string("op_5594_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 128]> var_5594_cast_fp16 = expand_dims(axes = var_5594_axes_0, x = one_hot_1_cast_fp16_cast_uint16_cast_uint16)[name = string("op_5594_cast_fp16")];
+            tensor<int32, [3]> one_hot_reps_0 = const()[name = string("one_hot_reps_0"), val = tensor<int32, [3]>([1, 14, 1])];
+            tensor<fp16, [1, 14, 128]> one_hot_cast_fp16 = tile(reps = one_hot_reps_0, x = var_5594_cast_fp16)[name = string("one_hot_cast_fp16")];
+            int32 var_5603 = const()[name = string("op_5603"), val = int32(-1)];
+            bool input_1273_interleave_0 = const()[name = string("input_1273_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 14, 1152]> input_1273_cast_fp16 = concat(axis = var_5603, interleave = input_1273_interleave_0, values = (audio_signal_cast_fp16, one_hot_cast_fp16))[name = string("input_1273_cast_fp16")];
+            tensor<fp16, [2048, 1152]> prompt_kernel_0_weight_to_fp16 = const()[name = string("prompt_kernel_0_weight_to_fp16"), val = tensor<fp16, [2048, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555721216)))];
+            tensor<fp16, [2048]> prompt_kernel_0_bias_to_fp16 = const()[name = string("prompt_kernel_0_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560439872)))];
+            tensor<fp16, [1, 14, 2048]> linear_217_cast_fp16 = linear(bias = prompt_kernel_0_bias_to_fp16, weight = prompt_kernel_0_weight_to_fp16, x = input_1273_cast_fp16)[name = string("linear_217_cast_fp16")];
+            tensor<fp16, [1, 14, 2048]> input_1277_cast_fp16 = relu(x = linear_217_cast_fp16)[name = string("input_1277_cast_fp16")];
+            tensor<fp16, [1024, 2048]> prompt_kernel_2_weight_to_fp16 = const()[name = string("prompt_kernel_2_weight_to_fp16"), val = tensor<fp16, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560444032)))];
+            tensor<fp16, [1024]> prompt_kernel_2_bias_to_fp16 = const()[name = string("prompt_kernel_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564638400)))];
+            tensor<fp16, [1, 14, 1024]> linear_218_cast_fp16 = linear(bias = prompt_kernel_2_bias_to_fp16, weight = prompt_kernel_2_weight_to_fp16, x = input_1277_cast_fp16)[name = string("linear_218_cast_fp16")];
+            string conditioned_cast_fp16_to_fp32_dtype_0 = const()[name = string("conditioned_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [3]> transpose_72_perm_0_1 = const()[name = string("transpose_72_perm_0_1"), val = tensor<int32, [3]>([0, 2, 1])];
+            string var_5621_dtype_0 = const()[name = string("op_5621_dtype_0"), val = string("int32")];
+            tensor<int32, [4]> var_5624_perm_0 = const()[name = string("op_5624_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string var_5624_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5624_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [4]> var_5627_perm_0 = const()[name = string("op_5627_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string var_5627_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5627_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            string var_5632_dtype_0 = const()[name = string("op_5632_dtype_0"), val = string("int32")];
+            tensor<fp16, [640, 1024]> joint_enc_weight_to_fp16 = const()[name = string("joint_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564640512)))];
+            tensor<fp16, [640]> joint_enc_bias_to_fp16 = const()[name = string("joint_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565951296)))];
+            tensor<fp16, [1, 14, 640]> linear_219_cast_fp16 = linear(bias = joint_enc_bias_to_fp16, weight = joint_enc_weight_to_fp16, x = linear_218_cast_fp16)[name = string("linear_219_cast_fp16")];
+            string linear_219_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_219_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 14, 640]> encoder_proj = cast(dtype = linear_219_cast_fp16_to_fp32_dtype_0, x = linear_219_cast_fp16)[name = string("cast_0")];
+            tensor<int32, [1]> cache_len_out = cast(dtype = var_5632_dtype_0, x = clip_1_cast_fp16)[name = string("cast_1")];
+            tensor<fp16, [1, 24, 1024, 8]> var_5627_cast_fp16 = transpose(perm = var_5627_perm_0, x = obj_7_cast_fp16)[name = string("transpose_144")];
+            tensor<fp32, [1, 24, 1024, 8]> cache_time_out = cast(dtype = var_5627_cast_fp16_to_fp32_dtype_0, x = var_5627_cast_fp16)[name = string("cast_2")];
+            tensor<fp16, [1, 24, 42, 1024]> var_5624_cast_fp16 = transpose(perm = var_5624_perm_0, x = obj_5_cast_fp16)[name = string("transpose_145")];
+            tensor<fp32, [1, 24, 42, 1024]> cache_channel_out = cast(dtype = var_5624_cast_fp16_to_fp32_dtype_0, x = var_5624_cast_fp16)[name = string("cast_3")];
+            tensor<int32, [1]> encoded_length = cast(dtype = var_5621_dtype_0, x = clip_0_cast_fp16)[name = string("cast_4")];
+            tensor<fp16, [1, 1024, 14]> transpose_72_1 = transpose(perm = transpose_72_perm_0_1, x = linear_218_cast_fp16)[name = string("transpose_146")];
+            tensor<fp32, [1, 1024, 14]> encoded = cast(dtype = conditioned_cast_fp16_to_fp32_dtype_0, x = transpose_72_1)[name = string("cast_5")];
+        } -> (encoded, encoded_length, cache_channel_out, cache_time_out, cache_len_out, encoder_proj);
+}
\ No newline at end of file