diff --git a/latin/1120ms/decoder.mlmodelc/weights/weight.bin b/latin/1120ms/decoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..b25cf03aaabac63a0f946036e3e3ce9c5f8abffb
--- /dev/null
+++ b/latin/1120ms/decoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:923df635dad8f9688d46afdf1689c0913cc977f9944db30e35a78e5b41799903
+size 16739072
diff --git a/latin/1120ms/decoder_joint.mlmodelc/weights/weight.bin b/latin/1120ms/decoder_joint.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..68d2b0afd28b8baf6308f64f419042ce35a5579d
--- /dev/null
+++ b/latin/1120ms/decoder_joint.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5792bda2b88daae2ccb66689069da286e801f7bd168a95aa430590efa93e8e04
+size 22498714
diff --git a/latin/1120ms/encoder.mlmodelc/weights/weight.bin b/latin/1120ms/encoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..22d4a9d6056019a89615f7756e310887fb6faff0
--- /dev/null
+++ b/latin/1120ms/encoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3dbd37659c7ef51d2dc691afac42fbdd1f04998fa62816ff0f2eebec5cc3d01
+size 564646848
diff --git a/latin/2240ms/decoder.mlmodelc/analytics/coremldata.bin b/latin/2240ms/decoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..700cbb18eb40be2dca6a0570bd2e7ceef5712de9
--- /dev/null
+++ b/latin/2240ms/decoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2e55e9add404d86292a04d8e25a6aee3833095d6191f73bbfa2167445384e0c
+size 243
diff --git a/latin/2240ms/decoder.mlmodelc/weights/weight.bin b/latin/2240ms/decoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..b25cf03aaabac63a0f946036e3e3ce9c5f8abffb
--- /dev/null
+++ b/latin/2240ms/decoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:923df635dad8f9688d46afdf1689c0913cc977f9944db30e35a78e5b41799903
+size 16739072
diff --git a/latin/2240ms/decoder_joint.mlmodelc/analytics/coremldata.bin b/latin/2240ms/decoder_joint.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..01a4d8e659ce38221f7d62422cc52c6dd728461e
--- /dev/null
+++ b/latin/2240ms/decoder_joint.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0908aa04510db0fde1188519ae0a067789a7826613e8f96c47832ded7d0de131
+size 243
diff --git a/latin/2240ms/decoder_joint.mlmodelc/weights/weight.bin b/latin/2240ms/decoder_joint.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..68d2b0afd28b8baf6308f64f419042ce35a5579d
--- /dev/null
+++ b/latin/2240ms/decoder_joint.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5792bda2b88daae2ccb66689069da286e801f7bd168a95aa430590efa93e8e04
+size 22498714
diff --git a/latin/2240ms/encoder.mlmodelc/analytics/coremldata.bin b/latin/2240ms/encoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..0795a6e1d2a0f3a23218f5abe9f246a6a404d080
--- /dev/null
+++ b/latin/2240ms/encoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af569fd95237bdf8b91d38094691bfc990fcf3534316fa615bac5964c02809e4
+size 243
diff --git a/latin/2240ms/encoder.mlmodelc/weights/weight.bin b/latin/2240ms/encoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..15a9977eabe4d212f6775bc2c6653c508556d54f
--- /dev/null
+++ b/latin/2240ms/encoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e00be98049a22e095452c020f183d2b23728e145cc814ba031436931b4f2e8f
+size 565336640
diff --git a/latin/2240ms/joint.mlmodelc/analytics/coremldata.bin b/latin/2240ms/joint.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..1448cf9be5817f18c92b98cd90017254d6521f55
--- /dev/null
+++ b/latin/2240ms/joint.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6dbd3bf1b03e529be08a4673cb1d5889cbc5ac40ff6c7df0c47e3676c1ef6dc7
+size 243
diff --git a/latin/2240ms/joint.mlmodelc/weights/weight.bin b/latin/2240ms/joint.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..baa3ed0a821090eb9e42550135aba83fb8e89bbe
--- /dev/null
+++ b/latin/2240ms/joint.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66b9d7405fad48fdd8413a7275c04d44933cb9beb80d318c158f80e52ff43fba
+size 5759706
diff --git a/latin/2240ms/preprocessor.mlmodelc/analytics/coremldata.bin b/latin/2240ms/preprocessor.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7d40c48169474732ff5d227c86c093a242482865
--- /dev/null
+++ b/latin/2240ms/preprocessor.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e918fd75105ef01a971d29b5ec28f531467b42dd60978c29afb1914c4af838af
+size 243
diff --git a/latin/2240ms/preprocessor.mlmodelc/weights/weight.bin b/latin/2240ms/preprocessor.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411
--- /dev/null
+++ b/latin/2240ms/preprocessor.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3
+size 592384
diff --git a/latin/4480ms/decoder.mlmodelc/analytics/coremldata.bin b/latin/4480ms/decoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7306458767336f10cbfb307063718cf305cc2b0f
--- /dev/null
+++ b/latin/4480ms/decoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:895e6c97a82782b2fd429e6fcfc7201bbb2e6fe668d8d5f5636e05654f334b8e
+size 243
diff --git a/latin/4480ms/decoder.mlmodelc/coremldata.bin b/latin/4480ms/decoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..81fe951046888805a7a4d02bbc0633cfe2f122ea
--- /dev/null
+++ b/latin/4480ms/decoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:76a3ae4c7f2dcd6cd4bf3a1ccbe40ad0c0aecd693e1cdebd8cb85fa30f44a8b9
+size 433
diff --git a/latin/4480ms/decoder.mlmodelc/model.mil b/latin/4480ms/decoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..7394ffc00c3770474de6dc940583a31934b72c03
--- /dev/null
+++ b/latin/4480ms/decoder.mlmodelc/model.mil
@@ -0,0 +1,64 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            tensor<int32, []> y_axis_0 = const()[name = tensor<string, []>("y_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> y_batch_dims_0 = const()[name = tensor<string, []>("y_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> y_validate_indices_0 = const()[name = tensor<string, []>("y_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [2829, 640]> module_prediction_embed_weight_to_fp16 = const()[name = tensor<string, []>("module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [2829, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<string, []> token_to_int16_dtype_0 = const()[name = tensor<string, []>("token_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = tensor<string, []>("cast_8")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = tensor<string, []>("y_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, []> split_0_num_splits_0 = const()[name = tensor<string, []>("split_0_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_0_axis_0 = const()[name = tensor<string, []>("split_0_axis_0"), val = tensor<int32, []>(0)];
+            tensor<string, []> h_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("h_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = tensor<string, []>("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = tensor<string, []>("split_0_cast_fp16")];
+            tensor<int32, []> split_1_num_splits_0 = const()[name = tensor<string, []>("split_1_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_1_axis_0 = const()[name = tensor<string, []>("split_1_axis_0"), val = tensor<int32, []>(0)];
+            tensor<string, []> c_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("c_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = tensor<string, []>("cast_6")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = tensor<string, []>("split_1_cast_fp16")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = tensor<string, []>("input_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = tensor<string, []>("input_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> input_lstm_layer_0_direction_0 = const()[name = tensor<string, []>("input_lstm_layer_0_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_lstm_layer_0_output_sequence_0 = const()[name = tensor<string, []>("input_lstm_layer_0_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_lstm_layer_0_recurrent_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_lstm_layer_0_cell_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_lstm_layer_0_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = tensor<string, []>("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3621248)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = tensor<string, []>("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6898112)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = tensor<string, []>("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10174976)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_2")];
+            tensor<fp16, [1, 1, 640]> input_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_lstm_layer_0_cast_fp16_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = tensor<string, []>("input_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = tensor<string, []>("input_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> input_direction_0 = const()[name = tensor<string, []>("input_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_output_sequence_0 = const()[name = tensor<string, []>("input_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_recurrent_activation_0 = const()[name = tensor<string, []>("input_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_cell_activation_0 = const()[name = tensor<string, []>("input_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_activation_0 = const()[name = tensor<string, []>("input_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = tensor<string, []>("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10180160)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = tensor<string, []>("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13457024)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = tensor<string, []>("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16733888)))];
+            tensor<fp16, [1, 1, 640]> input_cast_fp16_0, tensor<fp16, [1, 640]> input_cast_fp16_1, tensor<fp16, [1, 640]> input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_3_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_lstm_layer_0_cast_fp16_0)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, []> obj_3_axis_0 = const()[name = tensor<string, []>("obj_3_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_cast_fp16_1, input_cast_fp16_1))[name = tensor<string, []>("obj_3_cast_fp16")];
+            tensor<string, []> obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_3_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, []> obj_axis_0 = const()[name = tensor<string, []>("obj_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_lstm_layer_0_cast_fp16_2, input_cast_fp16_2))[name = tensor<string, []>("obj_cast_fp16")];
+            tensor<string, []> obj_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = tensor<string, []>("transpose_0_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
+            tensor<string, []> transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("transpose_0_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp16, [1, 640, 1]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = tensor<string, []>("transpose_1")];
+            tensor<fp32, [1, 640, 1]> decoder_out = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = tensor<string, []>("cast_3")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = tensor<string, []>("cast_4")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = tensor<string, []>("cast_5")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = tensor<string, []>("token_length_tmp")];
+        } -> (decoder_out, h_out, c_out);
+}
\ No newline at end of file
diff --git a/latin/4480ms/decoder.mlmodelc/weights/weight.bin b/latin/4480ms/decoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..b25cf03aaabac63a0f946036e3e3ce9c5f8abffb
--- /dev/null
+++ b/latin/4480ms/decoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:923df635dad8f9688d46afdf1689c0913cc977f9944db30e35a78e5b41799903
+size 16739072
diff --git a/latin/4480ms/decoder_joint.mlmodelc/analytics/coremldata.bin b/latin/4480ms/decoder_joint.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..01a4d8e659ce38221f7d62422cc52c6dd728461e
--- /dev/null
+++ b/latin/4480ms/decoder_joint.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0908aa04510db0fde1188519ae0a067789a7826613e8f96c47832ded7d0de131
+size 243
diff --git a/latin/4480ms/decoder_joint.mlmodelc/coremldata.bin b/latin/4480ms/decoder_joint.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..40ab419974d2f10ed16576675826a4ea354ee61b
--- /dev/null
+++ b/latin/4480ms/decoder_joint.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a72f522aac0b6dc7490c1c71a270c607611a4eb357f1300bd3984852a98ff32
+size 454
diff --git a/latin/4480ms/decoder_joint.mlmodelc/model.mil b/latin/4480ms/decoder_joint.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..a74099fa94fca54b433d1c19fba076c3ebec7948
--- /dev/null
+++ b/latin/4480ms/decoder_joint.mlmodelc/model.mil
@@ -0,0 +1,83 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [1, 1024, 1]> encoder, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            tensor<int32, []> y_axis_0 = const()[name = tensor<string, []>("y_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> y_batch_dims_0 = const()[name = tensor<string, []>("y_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> y_validate_indices_0 = const()[name = tensor<string, []>("y_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [2829, 640]> decoder_module_prediction_embed_weight_to_fp16 = const()[name = tensor<string, []>("decoder_module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [2829, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<string, []> token_to_int16_dtype_0 = const()[name = tensor<string, []>("token_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = tensor<string, []>("cast_9")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = tensor<string, []>("y_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, []> split_0_num_splits_0 = const()[name = tensor<string, []>("split_0_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_0_axis_0 = const()[name = tensor<string, []>("split_0_axis_0"), val = tensor<int32, []>(0)];
+            tensor<string, []> h_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("h_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = tensor<string, []>("cast_8")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = tensor<string, []>("split_0_cast_fp16")];
+            tensor<int32, []> split_1_num_splits_0 = const()[name = tensor<string, []>("split_1_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_1_axis_0 = const()[name = tensor<string, []>("split_1_axis_0"), val = tensor<int32, []>(0)];
+            tensor<string, []> c_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("c_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = tensor<string, []>("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = tensor<string, []>("split_1_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = tensor<string, []>("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = tensor<string, []>("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> input_5_lstm_layer_0_direction_0 = const()[name = tensor<string, []>("input_5_lstm_layer_0_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_5_lstm_layer_0_output_sequence_0 = const()[name = tensor<string, []>("input_5_lstm_layer_0_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_5_lstm_layer_0_recurrent_activation_0 = const()[name = tensor<string, []>("input_5_lstm_layer_0_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_5_lstm_layer_0_cell_activation_0 = const()[name = tensor<string, []>("input_5_lstm_layer_0_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_5_lstm_layer_0_activation_0 = const()[name = tensor<string, []>("input_5_lstm_layer_0_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = tensor<string, []>("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3621248)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = tensor<string, []>("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6898112)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = tensor<string, []>("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10174976)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_4")];
+            tensor<fp16, [1, 1, 640]> input_5_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_5_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = tensor<string, []>("input_5_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_5_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = tensor<string, []>("input_5_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> input_5_direction_0 = const()[name = tensor<string, []>("input_5_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_5_output_sequence_0 = const()[name = tensor<string, []>("input_5_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_5_recurrent_activation_0 = const()[name = tensor<string, []>("input_5_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_5_cell_activation_0 = const()[name = tensor<string, []>("input_5_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_5_activation_0 = const()[name = tensor<string, []>("input_5_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = tensor<string, []>("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10180160)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = tensor<string, []>("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13457024)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = tensor<string, []>("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16733888)))];
+            tensor<fp16, [1, 1, 640]> input_5_cast_fp16_0, tensor<fp16, [1, 640]> input_5_cast_fp16_1, tensor<fp16, [1, 640]> input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<int32, []> obj_3_axis_0 = const()[name = tensor<string, []>("obj_3_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = tensor<string, []>("obj_3_cast_fp16")];
+            tensor<string, []> obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_3_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, []> obj_axis_0 = const()[name = tensor<string, []>("obj_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = tensor<string, []>("obj_cast_fp16")];
+            tensor<string, []> obj_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [3]> transpose_1_perm_0 = const()[name = tensor<string, []>("transpose_1_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [3]> input_7_perm_0 = const()[name = tensor<string, []>("input_7_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> encoder_to_fp16_dtype_0 = const()[name = tensor<string, []>("encoder_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [640, 1024]> joint_module_enc_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16739072)))];
+            tensor<fp16, [640]> joint_module_enc_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18049856)))];
+            tensor<fp16, [1, 1024, 1]> encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = tensor<string, []>("cast_4")];
+            tensor<fp16, [1, 1, 1024]> input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = encoder_to_fp16)[name = tensor<string, []>("transpose_2")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18051200)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18870464)))];
+            tensor<fp16, [1, 1, 640]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = tensor<string, []>("transpose_3")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_79_axes_0 = const()[name = tensor<string, []>("op_79_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_79_cast_fp16 = expand_dims(axes = var_79_axes_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_79_cast_fp16")];
+            tensor<int32, [1]> var_80_axes_0 = const()[name = tensor<string, []>("op_80_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_80_cast_fp16 = expand_dims(axes = var_80_axes_0, x = linear_1_cast_fp16)[name = tensor<string, []>("op_80_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_11_cast_fp16 = add(x = var_79_cast_fp16, y = var_80_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_13_cast_fp16 = relu(x = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
+            tensor<fp16, [2829, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [2829, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18871808)))];
+            tensor<fp16, [2829]> joint_module_joint_net_2_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [2829]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22492992)))];
+            tensor<fp16, [1, 1, 1, 2829]> linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<string, []> linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("linear_2_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 1, 1, 2829]> logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = tensor<string, []>("cast_3")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = tensor<string, []>("cast_5")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = tensor<string, []>("cast_6")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = tensor<string, []>("token_length_tmp")];
+        } -> (logits, h_out, c_out);
+}
\ No newline at end of file
diff --git a/latin/4480ms/decoder_joint.mlmodelc/weights/weight.bin b/latin/4480ms/decoder_joint.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..68d2b0afd28b8baf6308f64f419042ce35a5579d
--- /dev/null
+++ b/latin/4480ms/decoder_joint.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5792bda2b88daae2ccb66689069da286e801f7bd168a95aa430590efa93e8e04
+size 22498714
diff --git a/latin/4480ms/encoder.mlmodelc/analytics/coremldata.bin b/latin/4480ms/encoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..07f1338b410f7c45ec613f261758b3a988fda10f
--- /dev/null
+++ b/latin/4480ms/encoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a3d8e583dff43e6775dac37936793d4bf5006911c7d939a93d523dba70c81fc
+size 243
diff --git a/latin/4480ms/encoder.mlmodelc/coremldata.bin b/latin/4480ms/encoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..22fc2a1e318cdf3e64cb17fdf984bc5bcd6ef690
--- /dev/null
+++ b/latin/4480ms/encoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2559ad5418947a317a172c03693799b39e4966e3ebef05fe56470fb9166df98
+size 573
diff --git a/latin/4480ms/encoder.mlmodelc/model.mil b/latin/4480ms/encoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..1baf22502c12f7f01313667d6a61e58bfd1ad6fd
--- /dev/null
+++ b/latin/4480ms/encoder.mlmodelc/model.mil
@@ -0,0 +1,4321 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})]
+{
+    func main<ios17>(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, 457]> mel, tensor<int32, [1]> mel_length, tensor<int32, [1]> prompt_id) {
+            tensor<int32, [4]> value_3_perm_0 = const()[name = tensor<string, []>("value_3_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<string, []> cache_channel_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_channel_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<int32, [4]> value_5_perm_0 = const()[name = tensor<string, []>("value_5_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<string, []> cache_time_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_time_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<int32, []> var_59 = const()[name = tensor<string, []>("op_59"), val = tensor<int32, []>(-1)];
+            tensor<int32, []> var_68 = const()[name = tensor<string, []>("op_68"), val = tensor<int32, []>(1)];
+            tensor<int32, [3]> x_1_perm_0 = const()[name = tensor<string, []>("x_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> mel_to_fp16_dtype_0 = const()[name = tensor<string, []>("mel_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<int32, [1]> tensor_1_axes_0 = const()[name = tensor<string, []>("tensor_1_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 128, 457]> mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = tensor<string, []>("cast_18")];
+            tensor<fp16, [1, 457, 128]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = mel_to_fp16)[name = tensor<string, []>("transpose_367")];
+            tensor<fp16, [1, 1, 457, 128]> tensor_1_cast_fp16 = expand_dims(axes = tensor_1_axes_0, x = x_1_cast_fp16)[name = tensor<string, []>("tensor_1_cast_fp16")];
+            tensor<int32, [1, 457]> expand_dims_0 = const()[name = tensor<string, []>("expand_dims_0"), val = tensor<int32, [1, 457]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287, 288, 289, 290, 291, 292, 293, 294, 295, 296, 297, 298, 299, 300, 301, 302, 303, 304, 305, 306, 307, 308, 309, 310, 311, 312, 313, 314, 315, 316, 317, 318, 319, 320, 321, 322, 323, 324, 325, 326, 327, 328, 329, 330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 340, 341, 342, 343, 344, 345, 346, 347, 348, 349, 350, 351, 352, 353, 354, 355, 356, 357, 358, 359, 360, 361, 362, 363, 364, 365, 366, 367, 368, 369, 370, 371, 372, 373, 374, 375, 376, 377, 378, 379, 380, 381, 382, 383, 384, 385, 386, 387, 388, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404, 405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419, 420, 421, 422, 423, 424, 425, 426, 427, 428, 429, 430, 431, 432, 433, 434, 435, 436, 437, 438, 439, 440, 441, 442, 443, 444, 445, 446, 447, 448, 449, 450, 451, 452, 453, 454, 455, 456]])];
+            tensor<int32, [1]> var_137_axes_0 = const()[name = tensor<string, []>("op_137_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_137 = expand_dims(axes = var_137_axes_0, x = mel_length)[name = tensor<string, []>("op_137")];
+            tensor<bool, [1, 457]> time_mask_1 = less(x = expand_dims_0, y = var_137)[name = tensor<string, []>("time_mask_1")];
+            tensor<int32, [1]> var_139_axes_0 = const()[name = tensor<string, []>("op_139_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 457, 1]> var_139 = expand_dims(axes = var_139_axes_0, x = time_mask_1)[name = tensor<string, []>("op_139")];
+            tensor<int32, [3]> var_141_reps_0 = const()[name = tensor<string, []>("op_141_reps_0"), val = tensor<int32, [3]>([1, 1, 128])];
+            tensor<bool, [1, 457, 128]> var_141 = tile(reps = var_141_reps_0, x = var_139)[name = tensor<string, []>("op_141")];
+            tensor<int32, [1]> var_147_axes_0 = const()[name = tensor<string, []>("op_147_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<string, []> cast_4_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_4_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 457, 128]> var_141_to_fp16 = cast(dtype = cast_4_to_fp16_dtype_0, x = var_141)[name = tensor<string, []>("cast_17")];
+            tensor<fp16, [1, 1, 457, 128]> var_147_cast_fp16 = expand_dims(axes = var_147_axes_0, x = var_141_to_fp16)[name = tensor<string, []>("op_147_cast_fp16")];
+            tensor<fp16, [1, 1, 457, 128]> input_1_cast_fp16 = mul(x = tensor_1_cast_fp16, y = var_147_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<int32, [8]> input_3_pad_0 = const()[name = tensor<string, []>("input_3_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_3_mode_0 = const()[name = tensor<string, []>("input_3_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_9_to_fp16 = const()[name = tensor<string, []>("const_9_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 1, 460, 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 = tensor<string, []>("input_3_cast_fp16")];
+            tensor<string, []> tensor_3_pad_type_0 = const()[name = tensor<string, []>("tensor_3_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_3_strides_0 = const()[name = tensor<string, []>("tensor_3_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, [4]> tensor_3_pad_0 = const()[name = tensor<string, []>("tensor_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_3_dilations_0 = const()[name = tensor<string, []>("tensor_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> tensor_3_groups_0 = const()[name = tensor<string, []>("tensor_3_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_0_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_pre_encode_conv_0_weight_to_fp16_quantized"), quantized_data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64))), scale = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2752))), zero_point = tensor<int8, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2432)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3328)))];
+            tensor<fp16, [1, 256, 229, 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 = tensor<string, []>("tensor_3_cast_fp16")];
+            tensor<string, []> cast_2_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> var_160_promoted_to_fp16 = const()[name = tensor<string, []>("op_160_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> mel_length_to_fp16 = cast(dtype = cast_2_to_fp16_dtype_0, x = mel_length)[name = tensor<string, []>("cast_16")];
+            tensor<fp16, [1]> var_161_cast_fp16 = add(x = mel_length_to_fp16, y = var_160_promoted_to_fp16)[name = tensor<string, []>("op_161_cast_fp16")];
+            tensor<fp16, []> var_162_promoted_to_fp16 = const()[name = tensor<string, []>("op_162_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> var_163_cast_fp16 = add(x = var_161_cast_fp16, y = var_162_promoted_to_fp16)[name = tensor<string, []>("op_163_cast_fp16")];
+            tensor<fp16, []> var_164_promoted_to_fp16 = const()[name = tensor<string, []>("op_164_promoted_to_fp16"), val = tensor<fp16, []>(0x1.8p+1)];
+            tensor<fp16, [1]> var_165_cast_fp16 = sub(x = var_163_cast_fp16, y = var_164_promoted_to_fp16)[name = tensor<string, []>("op_165_cast_fp16")];
+            tensor<fp16, []> var_56_promoted_to_fp16 = const()[name = tensor<string, []>("op_56_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> floor_div_0_cast_fp16 = floor_div(x = var_165_cast_fp16, y = var_56_promoted_to_fp16)[name = tensor<string, []>("floor_div_0_cast_fp16")];
+            tensor<fp16, []> var_167_promoted_to_fp16 = const()[name = tensor<string, []>("op_167_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_3_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_167_promoted_to_fp16)[name = tensor<string, []>("current_lengths_3_cast_fp16")];
+            tensor<string, []> cast_5_dtype_0 = const()[name = tensor<string, []>("cast_5_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 229]> expand_dims_1 = const()[name = tensor<string, []>("expand_dims_1"), val = tensor<int32, [1, 229]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228]])];
+            tensor<int32, [1]> var_176_axes_0 = const()[name = tensor<string, []>("op_176_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_3_cast_fp16_to_int32 = cast(dtype = cast_5_dtype_0, x = current_lengths_3_cast_fp16)[name = tensor<string, []>("cast_15")];
+            tensor<int32, [1, 1]> var_176 = expand_dims(axes = var_176_axes_0, x = current_lengths_3_cast_fp16_to_int32)[name = tensor<string, []>("op_176")];
+            tensor<bool, [1, 229]> time_mask_3 = less(x = expand_dims_1, y = var_176)[name = tensor<string, []>("time_mask_3")];
+            tensor<int32, [1]> var_178_axes_0 = const()[name = tensor<string, []>("op_178_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 229, 1]> var_178 = expand_dims(axes = var_178_axes_0, x = time_mask_3)[name = tensor<string, []>("op_178")];
+            tensor<int32, [3]> var_180_reps_0 = const()[name = tensor<string, []>("op_180_reps_0"), val = tensor<int32, [3]>([1, 1, 65])];
+            tensor<bool, [1, 229, 65]> var_180 = tile(reps = var_180_reps_0, x = var_178)[name = tensor<string, []>("op_180")];
+            tensor<int32, [1]> var_186_axes_0 = const()[name = tensor<string, []>("op_186_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<string, []> cast_6_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_6_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 229, 65]> var_180_to_fp16 = cast(dtype = cast_6_to_fp16_dtype_0, x = var_180)[name = tensor<string, []>("cast_14")];
+            tensor<fp16, [1, 1, 229, 65]> var_186_cast_fp16 = expand_dims(axes = var_186_axes_0, x = var_180_to_fp16)[name = tensor<string, []>("op_186_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_3_reps_0 = const()[name = tensor<string, []>("expanded_mask_3_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 229, 65]> expanded_mask_3_cast_fp16 = tile(reps = expanded_mask_3_reps_0, x = var_186_cast_fp16)[name = tensor<string, []>("expanded_mask_3_cast_fp16")];
+            tensor<fp16, [1, 256, 229, 65]> input_5_cast_fp16 = mul(x = tensor_3_cast_fp16, y = expanded_mask_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [1, 256, 229, 65]> tensor_5_cast_fp16 = relu(x = input_5_cast_fp16)[name = tensor<string, []>("tensor_5_cast_fp16")];
+            tensor<fp16, [1, 256, 229, 65]> input_7_cast_fp16 = mul(x = tensor_5_cast_fp16, y = expanded_mask_3_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<int32, [8]> input_9_pad_0 = const()[name = tensor<string, []>("input_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_9_mode_0 = const()[name = tensor<string, []>("input_9_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_23_to_fp16 = const()[name = tensor<string, []>("const_23_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 256, 232, 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 = tensor<string, []>("input_9_cast_fp16")];
+            tensor<string, []> tensor_7_pad_type_0 = const()[name = tensor<string, []>("tensor_7_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_7_strides_0 = const()[name = tensor<string, []>("tensor_7_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, []> tensor_7_groups_0 = const()[name = tensor<string, []>("tensor_7_groups_0"), val = tensor<int32, []>(256)];
+            tensor<int32, [4]> tensor_7_pad_0 = const()[name = tensor<string, []>("tensor_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_7_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_pre_encode_conv_2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3904))), scale = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6272))), zero_point = tensor<int8, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2432)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6848)))];
+            tensor<fp16, [1, 256, 115, 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 = tensor<string, []>("tensor_7_cast_fp16")];
+            tensor<fp16, []> var_208_promoted_to_fp16 = const()[name = tensor<string, []>("op_208_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> var_209_cast_fp16 = add(x = current_lengths_3_cast_fp16, y = var_208_promoted_to_fp16)[name = tensor<string, []>("op_209_cast_fp16")];
+            tensor<fp16, []> var_210_promoted_to_fp16 = const()[name = tensor<string, []>("op_210_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> var_211_cast_fp16 = add(x = var_209_cast_fp16, y = var_210_promoted_to_fp16)[name = tensor<string, []>("op_211_cast_fp16")];
+            tensor<fp16, []> var_212_promoted_to_fp16 = const()[name = tensor<string, []>("op_212_promoted_to_fp16"), val = tensor<fp16, []>(0x1.8p+1)];
+            tensor<fp16, [1]> var_213_cast_fp16 = sub(x = var_211_cast_fp16, y = var_212_promoted_to_fp16)[name = tensor<string, []>("op_213_cast_fp16")];
+            tensor<fp16, []> var_56_promoted_1_to_fp16 = const()[name = tensor<string, []>("op_56_promoted_1_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> floor_div_1_cast_fp16 = floor_div(x = var_213_cast_fp16, y = var_56_promoted_1_to_fp16)[name = tensor<string, []>("floor_div_1_cast_fp16")];
+            tensor<fp16, []> var_215_promoted_to_fp16 = const()[name = tensor<string, []>("op_215_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_5_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_215_promoted_to_fp16)[name = tensor<string, []>("current_lengths_5_cast_fp16")];
+            tensor<string, []> cast_7_dtype_0 = const()[name = tensor<string, []>("cast_7_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 115]> expand_dims_2 = const()[name = tensor<string, []>("expand_dims_2"), val = tensor<int32, [1, 115]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114]])];
+            tensor<int32, [1]> var_224_axes_0 = const()[name = tensor<string, []>("op_224_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_5_cast_fp16_to_int32 = cast(dtype = cast_7_dtype_0, x = current_lengths_5_cast_fp16)[name = tensor<string, []>("cast_13")];
+            tensor<int32, [1, 1]> var_224 = expand_dims(axes = var_224_axes_0, x = current_lengths_5_cast_fp16_to_int32)[name = tensor<string, []>("op_224")];
+            tensor<bool, [1, 115]> time_mask_5 = less(x = expand_dims_2, y = var_224)[name = tensor<string, []>("time_mask_5")];
+            tensor<int32, [1]> var_226_axes_0 = const()[name = tensor<string, []>("op_226_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 115, 1]> var_226 = expand_dims(axes = var_226_axes_0, x = time_mask_5)[name = tensor<string, []>("op_226")];
+            tensor<int32, [3]> var_228_reps_0 = const()[name = tensor<string, []>("op_228_reps_0"), val = tensor<int32, [3]>([1, 1, 33])];
+            tensor<bool, [1, 115, 33]> var_228 = tile(reps = var_228_reps_0, x = var_226)[name = tensor<string, []>("op_228")];
+            tensor<int32, [1]> var_234_axes_0 = const()[name = tensor<string, []>("op_234_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<string, []> cast_8_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_8_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 115, 33]> var_228_to_fp16 = cast(dtype = cast_8_to_fp16_dtype_0, x = var_228)[name = tensor<string, []>("cast_12")];
+            tensor<fp16, [1, 1, 115, 33]> var_234_cast_fp16 = expand_dims(axes = var_234_axes_0, x = var_228_to_fp16)[name = tensor<string, []>("op_234_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_7_reps_0 = const()[name = tensor<string, []>("expanded_mask_7_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 115, 33]> expanded_mask_7_cast_fp16 = tile(reps = expanded_mask_7_reps_0, x = var_234_cast_fp16)[name = tensor<string, []>("expanded_mask_7_cast_fp16")];
+            tensor<fp16, [1, 256, 115, 33]> input_11_cast_fp16 = mul(x = tensor_7_cast_fp16, y = expanded_mask_7_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<string, []> tensor_9_pad_type_0 = const()[name = tensor<string, []>("tensor_9_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_9_strides_0 = const()[name = tensor<string, []>("tensor_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_9_pad_0 = const()[name = tensor<string, []>("tensor_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_9_dilations_0 = const()[name = tensor<string, []>("tensor_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> tensor_9_groups_0 = const()[name = tensor<string, []>("tensor_9_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_3_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_pre_encode_conv_3_weight_to_fp16_quantized"), quantized_data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7424))), scale = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73024))), zero_point = tensor<int8, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2432)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73600)))];
+            tensor<fp16, [1, 256, 115, 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 = tensor<string, []>("tensor_9_cast_fp16")];
+            tensor<fp16, [1, 256, 115, 33]> input_13_cast_fp16 = mul(x = tensor_9_cast_fp16, y = expanded_mask_7_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
+            tensor<fp16, [1, 256, 115, 33]> tensor_11_cast_fp16 = relu(x = input_13_cast_fp16)[name = tensor<string, []>("tensor_11_cast_fp16")];
+            tensor<fp16, [1, 256, 115, 33]> input_15_cast_fp16 = mul(x = tensor_11_cast_fp16, y = expanded_mask_7_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
+            tensor<int32, [8]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_17_mode_0 = const()[name = tensor<string, []>("input_17_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_41_to_fp16 = const()[name = tensor<string, []>("const_41_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 256, 118, 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 = tensor<string, []>("input_17_cast_fp16")];
+            tensor<string, []> tensor_13_pad_type_0 = const()[name = tensor<string, []>("tensor_13_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_13_strides_0 = const()[name = tensor<string, []>("tensor_13_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, []> tensor_13_groups_0 = const()[name = tensor<string, []>("tensor_13_groups_0"), val = tensor<int32, []>(256)];
+            tensor<int32, [4]> tensor_13_pad_0 = const()[name = tensor<string, []>("tensor_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_13_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_pre_encode_conv_5_weight_to_fp16_quantized"), quantized_data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74176))), scale = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76544))), zero_point = tensor<int8, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2432)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77120)))];
+            tensor<fp16, [1, 256, 58, 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 = tensor<string, []>("tensor_13_cast_fp16")];
+            tensor<fp16, []> var_271_promoted_to_fp16 = const()[name = tensor<string, []>("op_271_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> var_272_cast_fp16 = add(x = current_lengths_5_cast_fp16, y = var_271_promoted_to_fp16)[name = tensor<string, []>("op_272_cast_fp16")];
+            tensor<fp16, []> var_273_promoted_to_fp16 = const()[name = tensor<string, []>("op_273_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> var_274_cast_fp16 = add(x = var_272_cast_fp16, y = var_273_promoted_to_fp16)[name = tensor<string, []>("op_274_cast_fp16")];
+            tensor<fp16, []> var_275_promoted_to_fp16 = const()[name = tensor<string, []>("op_275_promoted_to_fp16"), val = tensor<fp16, []>(0x1.8p+1)];
+            tensor<fp16, [1]> var_276_cast_fp16 = sub(x = var_274_cast_fp16, y = var_275_promoted_to_fp16)[name = tensor<string, []>("op_276_cast_fp16")];
+            tensor<fp16, []> var_56_promoted_2_to_fp16 = const()[name = tensor<string, []>("op_56_promoted_2_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> floor_div_2_cast_fp16 = floor_div(x = var_276_cast_fp16, y = var_56_promoted_2_to_fp16)[name = tensor<string, []>("floor_div_2_cast_fp16")];
+            tensor<fp16, []> var_278_promoted_to_fp16 = const()[name = tensor<string, []>("op_278_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_278_promoted_to_fp16)[name = tensor<string, []>("current_lengths_cast_fp16")];
+            tensor<string, []> cast_9_dtype_0 = const()[name = tensor<string, []>("cast_9_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 58]> expand_dims_3 = const()[name = tensor<string, []>("expand_dims_3"), val = tensor<int32, [1, 58]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57]])];
+            tensor<int32, [1]> var_287_axes_0 = const()[name = tensor<string, []>("op_287_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_cast_fp16_to_int32 = cast(dtype = cast_9_dtype_0, x = current_lengths_cast_fp16)[name = tensor<string, []>("cast_11")];
+            tensor<int32, [1, 1]> var_287 = expand_dims(axes = var_287_axes_0, x = current_lengths_cast_fp16_to_int32)[name = tensor<string, []>("op_287")];
+            tensor<bool, [1, 58]> time_mask = less(x = expand_dims_3, y = var_287)[name = tensor<string, []>("time_mask")];
+            tensor<int32, [1]> var_289_axes_0 = const()[name = tensor<string, []>("op_289_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 58, 1]> var_289 = expand_dims(axes = var_289_axes_0, x = time_mask)[name = tensor<string, []>("op_289")];
+            tensor<int32, [3]> var_291_reps_0 = const()[name = tensor<string, []>("op_291_reps_0"), val = tensor<int32, [3]>([1, 1, 17])];
+            tensor<bool, [1, 58, 17]> var_291 = tile(reps = var_291_reps_0, x = var_289)[name = tensor<string, []>("op_291")];
+            tensor<int32, [1]> var_297_axes_0 = const()[name = tensor<string, []>("op_297_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<string, []> cast_10_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_10_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 58, 17]> var_291_to_fp16 = cast(dtype = cast_10_to_fp16_dtype_0, x = var_291)[name = tensor<string, []>("cast_10")];
+            tensor<fp16, [1, 1, 58, 17]> var_297_cast_fp16 = expand_dims(axes = var_297_axes_0, x = var_291_to_fp16)[name = tensor<string, []>("op_297_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_13_reps_0 = const()[name = tensor<string, []>("expanded_mask_13_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 58, 17]> expanded_mask_13_cast_fp16 = tile(reps = expanded_mask_13_reps_0, x = var_297_cast_fp16)[name = tensor<string, []>("expanded_mask_13_cast_fp16")];
+            tensor<fp16, [1, 256, 58, 17]> input_19_cast_fp16 = mul(x = tensor_13_cast_fp16, y = expanded_mask_13_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
+            tensor<string, []> tensor_15_pad_type_0 = const()[name = tensor<string, []>("tensor_15_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_15_strides_0 = const()[name = tensor<string, []>("tensor_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_15_pad_0 = const()[name = tensor<string, []>("tensor_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_15_dilations_0 = const()[name = tensor<string, []>("tensor_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> tensor_15_groups_0 = const()[name = tensor<string, []>("tensor_15_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_6_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_pre_encode_conv_6_weight_to_fp16_quantized"), quantized_data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77696))), scale = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143296))), zero_point = tensor<int8, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2432)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143872)))];
+            tensor<fp16, [1, 256, 58, 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 = tensor<string, []>("tensor_15_cast_fp16")];
+            tensor<fp16, [1, 256, 58, 17]> input_21_cast_fp16 = mul(x = tensor_15_cast_fp16, y = expanded_mask_13_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
+            tensor<fp16, [1, 256, 58, 17]> tensor_cast_fp16 = relu(x = input_21_cast_fp16)[name = tensor<string, []>("tensor_cast_fp16")];
+            tensor<fp16, [1, 256, 58, 17]> x_3_cast_fp16 = mul(x = tensor_cast_fp16, y = expanded_mask_13_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
+            tensor<int32, [4]> var_331_perm_0 = const()[name = tensor<string, []>("op_331_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_332 = const()[name = tensor<string, []>("op_332"), val = tensor<int32, [3]>([1, 58, -1])];
+            tensor<fp16, [1, 58, 256, 17]> var_331_cast_fp16 = transpose(perm = var_331_perm_0, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_366")];
+            tensor<fp16, [1, 58, 4352]> input_23_cast_fp16 = reshape(shape = var_332, x = var_331_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
+            tensor<fp16, [1024, 4352]> encoder_pre_encode_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_pre_encode_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4352]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144448))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4602048))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_pre_encode_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4604160)))];
+            tensor<fp16, [1, 58, 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 = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<int32, [3]> var_342_begin_0 = const()[name = tensor<string, []>("op_342_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_342_end_0 = const()[name = tensor<string, []>("op_342_end_0"), val = tensor<int32, [3]>([1, 58, 1024])];
+            tensor<bool, [3]> var_342_end_mask_0 = const()[name = tensor<string, []>("op_342_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 56, 1024]> var_342_cast_fp16 = slice_by_index(begin = var_342_begin_0, end = var_342_end_0, end_mask = var_342_end_mask_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_342_cast_fp16")];
+            tensor<int32, []> var_344 = const()[name = tensor<string, []>("op_344"), val = tensor<int32, []>(2)];
+            tensor<int32, [1]> var_345 = sub(x = current_lengths_cast_fp16_to_int32, y = var_344)[name = tensor<string, []>("op_345")];
+            tensor<string, []> var_345_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_345_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> var_62_promoted_to_fp16 = const()[name = tensor<string, []>("op_62_promoted_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, []> const_61_to_fp16 = const()[name = tensor<string, []>("const_61_to_fp16"), val = tensor<fp16, []>(inf)];
+            tensor<fp16, [1]> var_345_to_fp16 = cast(dtype = var_345_promoted_to_fp16_dtype_0, x = var_345)[name = tensor<string, []>("cast_9")];
+            tensor<fp16, [1]> clip_0_cast_fp16 = clip(alpha = var_62_promoted_to_fp16, beta = const_61_to_fp16, x = var_345_to_fp16)[name = tensor<string, []>("clip_0_cast_fp16")];
+            tensor<int32, [1]> max_audio_length_1 = const()[name = tensor<string, []>("max_audio_length_1"), val = tensor<int32, [1]>([56])];
+            tensor<fp16, []> var_361_promoted_to_fp16 = const()[name = tensor<string, []>("op_361_promoted_to_fp16"), val = tensor<fp16, []>(0x1.5p+5)];
+            tensor<fp16, [1]> padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_361_promoted_to_fp16)[name = tensor<string, []>("padding_length_cast_fp16")];
+            tensor<int32, []> const_63 = const()[name = tensor<string, []>("const_63"), val = tensor<int32, []>(-1)];
+            tensor<int32, [1]> var_363 = mul(x = cache_len, y = const_63)[name = tensor<string, []>("op_363")];
+            tensor<int32, []> var_364 = const()[name = tensor<string, []>("op_364"), val = tensor<int32, []>(42)];
+            tensor<int32, [1]> offset = add(x = var_363, y = var_364)[name = tensor<string, []>("offset")];
+            tensor<int32, [1]> var_404_axes_0 = const()[name = tensor<string, []>("op_404_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1]> var_404_cast_fp16 = expand_dims(axes = var_404_axes_0, x = padding_length_cast_fp16)[name = tensor<string, []>("op_404_cast_fp16")];
+            tensor<fp16, [1, 98]> var_403_promoted_to_fp16 = const()[name = tensor<string, []>("op_403_promoted_to_fp16"), val = tensor<fp16, [1, 98]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4606272)))];
+            tensor<bool, [1, 98]> pad_mask_1_cast_fp16 = less(x = var_403_promoted_to_fp16, y = var_404_cast_fp16)[name = tensor<string, []>("pad_mask_1_cast_fp16")];
+            tensor<int32, [1, 98]> expand_dims_5 = const()[name = tensor<string, []>("expand_dims_5"), val = tensor<int32, [1, 98]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97]])];
+            tensor<int32, [1]> var_410_axes_0 = const()[name = tensor<string, []>("op_410_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 1]> var_410 = expand_dims(axes = var_410_axes_0, x = offset)[name = tensor<string, []>("op_410")];
+            tensor<bool, [1, 98]> pad_mask_off = greater_equal(x = expand_dims_5, y = var_410)[name = tensor<string, []>("pad_mask_off")];
+            tensor<bool, [1, 98]> pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = tensor<string, []>("pad_mask_3")];
+            tensor<int32, [1]> var_413_axes_0 = const()[name = tensor<string, []>("op_413_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 98]> var_413 = expand_dims(axes = var_413_axes_0, x = pad_mask_3)[name = tensor<string, []>("op_413")];
+            tensor<int32, [3]> var_414 = const()[name = tensor<string, []>("op_414"), val = tensor<int32, [3]>([1, 98, 1])];
+            tensor<bool, [1, 98, 98]> pad_mask_for_att_mask_1 = tile(reps = var_414, x = var_413)[name = tensor<string, []>("pad_mask_for_att_mask_1")];
+            tensor<int32, [3]> var_416_perm_0 = const()[name = tensor<string, []>("op_416_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<bool, [1, 98, 98]> var_416 = transpose(perm = var_416_perm_0, x = pad_mask_for_att_mask_1)[name = tensor<string, []>("transpose_365")];
+            tensor<bool, [1, 98, 98]> pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_416)[name = tensor<string, []>("pad_mask_for_att_mask")];
+            tensor<bool, [1, 98, 98]> const_71 = const()[name = tensor<string, []>("const_71"), val = tensor<bool, [1, 98, 98]>([[[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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, 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, true, true, true, true, 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, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, 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, true, true, true, true, 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, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, 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, true, true, true, true, 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, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, 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, true, true, true, true, 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, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, 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, true, true, true, true, 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, false, false, false, false, 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, 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, true, true, true, true, 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], [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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, 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, true, true, true, true, 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], [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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, 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, true, true, true, true, 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], [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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, 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, true, true, true, true, 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], [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, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, 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, true, true, true, true, 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], [false, false, false, false, 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, 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, true, true, true, true, 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, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, 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, 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, true, true, true, true, 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, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, 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, 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, true, true, true, true, 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, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, 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, 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, true, true, true, true, 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, 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, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, 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, 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, true, true, true, true, 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, false, false, false, false, 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, 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, true, true, true, true, 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, 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], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, 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, true, true, true, true, 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, 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], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, 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, true, true, true, true, 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, 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], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, 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, true, true, true, true, 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, 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], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, 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, 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, true, true, true, true, 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, 98, 98]> att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = tensor<string, []>("att_mask_9")];
+            tensor<bool, [1, 98, 98]> att_mask = logical_not(x = att_mask_9)[name = tensor<string, []>("att_mask")];
+            tensor<bool, [1, 98]> pad_mask_5 = logical_not(x = pad_mask_3)[name = tensor<string, []>("pad_mask_5")];
+            tensor<int32, [2]> pad_mask_begin_0 = const()[name = tensor<string, []>("pad_mask_begin_0"), val = tensor<int32, [2]>([0, 42])];
+            tensor<int32, [2]> pad_mask_end_0 = const()[name = tensor<string, []>("pad_mask_end_0"), val = tensor<int32, [2]>([1, 98])];
+            tensor<bool, [2]> pad_mask_end_mask_0 = const()[name = tensor<string, []>("pad_mask_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<bool, [1, 56]> 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 = tensor<string, []>("pad_mask")];
+            tensor<int32, [3]> mask_9_begin_0 = const()[name = tensor<string, []>("mask_9_begin_0"), val = tensor<int32, [3]>([0, 42, 0])];
+            tensor<int32, [3]> mask_9_end_0 = const()[name = tensor<string, []>("mask_9_end_0"), val = tensor<int32, [3]>([1, 98, 98])];
+            tensor<bool, [3]> mask_9_end_mask_0 = const()[name = tensor<string, []>("mask_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<bool, [1, 56, 98]> 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 = tensor<string, []>("mask_9")];
+            tensor<int32, [4]> cache_1_begin_0 = const()[name = tensor<string, []>("cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_1_end_0 = const()[name = tensor<string, []>("cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 42, 1024])];
+            tensor<bool, [4]> cache_1_end_mask_0 = const()[name = tensor<string, []>("cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_1_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cast_8")];
+            tensor<fp16, [24, 1, 42, 1024]> value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = cache_channel_to_fp16)[name = tensor<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 = tensor<string, []>("cache_1_cast_fp16")];
+            tensor<int32, [4]> cache_3_begin_0 = const()[name = tensor<string, []>("cache_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_3_end_0 = const()[name = tensor<string, []>("cache_3_end_0"), val = tensor<int32, [4]>([1, 1, 1024, 8])];
+            tensor<bool, [4]> cache_3_end_mask_0 = const()[name = tensor<string, []>("cache_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_3_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cast_7")];
+            tensor<fp16, [24, 1, 1024, 8]> value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = cache_time_to_fp16)[name = tensor<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 = tensor<string, []>("cache_3_cast_fp16")];
+            tensor<int32, [1]> input_27_axes_0 = const()[name = tensor<string, []>("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4606592)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4608704)))];
+            tensor<fp16, []> var_42_to_fp16 = const()[name = tensor<string, []>("op_42_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 56, 1024]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_342_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4610816))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8809344))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805184)))];
+            tensor<fp16, [4096]> encoder_layers_0_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8817600)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8825856))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13020224))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_0_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13022336)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<fp16, []> var_455_to_fp16 = const()[name = tensor<string, []>("op_455_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_456_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_455_to_fp16)[name = tensor<string, []>("op_456_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_37_cast_fp16 = add(x = var_342_cast_fp16, y = var_456_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
+            tensor<int32, [1]> key_1_axes_0 = const()[name = tensor<string, []>("key_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13024448)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13026560)))];
+            tensor<fp16, [1, 56, 1024]> key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")];
+            tensor<bool, []> input_39_interleave_0 = const()[name = tensor<string, []>("input_39_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_39_cast_fp16 = concat(axis = var_68, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = tensor<string, []>("input_39_cast_fp16")];
+            tensor<bool, []> var_484_interleave_0 = const()[name = tensor<string, []>("op_484_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_484_cast_fp16 = concat(axis = var_68, interleave = var_484_interleave_0, values = key_1_cast_fp16)[name = tensor<string, []>("op_484_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13028672))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14077312))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14079424)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_3_cast_fp16")];
+            tensor<int32, [4]> var_489 = const()[name = tensor<string, []>("op_489"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_1_cast_fp16 = reshape(shape = var_489, x = linear_3_cast_fp16)[name = tensor<string, []>("q_1_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14081536))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15130176))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15132288)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_4_cast_fp16")];
+            tensor<int32, [4]> var_494 = const()[name = tensor<string, []>("op_494"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_1_cast_fp16 = reshape(shape = var_494, x = linear_4_cast_fp16)[name = tensor<string, []>("k_1_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15134400))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16183040))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16185152)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_5_cast_fp16")];
+            tensor<int32, [4]> var_499 = const()[name = tensor<string, []>("op_499"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_1_cast_fp16 = reshape(shape = var_499, x = linear_5_cast_fp16)[name = tensor<string, []>("v_1_cast_fp16")];
+            tensor<int32, [4]> value_9_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16187264)))];
+            tensor<fp16, [1, 56, 8, 128]> var_512_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_512_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16189376)))];
+            tensor<fp16, [1, 56, 8, 128]> var_514_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_514_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_1_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_1_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_7_transpose_x_0 = const()[name = tensor<string, []>("x_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_7_transpose_y_0 = const()[name = tensor<string, []>("x_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_516_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_516_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16191488))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391552))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_514_cast_fp16)[name = tensor<string, []>("transpose_362")];
+            tensor<fp16, [1, 8, 56, 195]> 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_516_to_fp16_quantized)[name = tensor<string, []>("x_7_cast_fp16")];
+            tensor<int32, [8]> x_9_pad_0 = const()[name = tensor<string, []>("x_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_9_mode_0 = const()[name = tensor<string, []>("x_9_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_79_to_fp16 = const()[name = tensor<string, []>("const_79_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_9_cast_fp16")];
+            tensor<int32, [4]> var_524 = const()[name = tensor<string, []>("op_524"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_11_cast_fp16 = reshape(shape = var_524, x = x_9_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
+            tensor<int32, [4]> var_528_begin_0 = const()[name = tensor<string, []>("op_528_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_528_end_0 = const()[name = tensor<string, []>("op_528_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_528_end_mask_0 = const()[name = tensor<string, []>("op_528_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_528_cast_fp16 = slice_by_index(begin = var_528_begin_0, end = var_528_end_0, end_mask = var_528_end_mask_0, x = x_11_cast_fp16)[name = tensor<string, []>("op_528_cast_fp16")];
+            tensor<int32, [4]> var_529 = const()[name = tensor<string, []>("op_529"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_1_cast_fp16 = reshape(shape = var_529, x = var_528_cast_fp16)[name = tensor<string, []>("matrix_bd_1_cast_fp16")];
+            tensor<bool, []> matrix_ac_1_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_1_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_96_perm_0 = const()[name = tensor<string, []>("transpose_96_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_97_perm_0 = const()[name = tensor<string, []>("transpose_97_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1_cast_fp16)[name = tensor<string, []>("transpose_360")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_96 = transpose(perm = transpose_96_perm_0, x = var_512_cast_fp16)[name = tensor<string, []>("transpose_361")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_1_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_3_begin_0 = const()[name = tensor<string, []>("matrix_bd_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_3_end_0 = const()[name = tensor<string, []>("matrix_bd_3_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_3_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_3_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_538_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = tensor<string, []>("op_538_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_1_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_1_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_1_cast_fp16 = mul(x = var_538_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_1_cast_fp16")];
+            tensor<int32, [1]> mask_11_axes_0 = const()[name = tensor<string, []>("mask_11_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 56, 98]> mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = tensor<string, []>("mask_11")];
+            tensor<fp16, []> var_45_to_fp16 = const()[name = tensor<string, []>("op_45_to_fp16"), val = tensor<fp16, []>(-0x1.388p+13)];
+            tensor<fp16, [1, 8, 56, 98]> scores_3_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_3_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_544_cast_fp16 = softmax(axis = var_59, x = scores_3_cast_fp16)[name = tensor<string, []>("op_544_cast_fp16")];
+            tensor<fp16, []> var_44_to_fp16 = const()[name = tensor<string, []>("op_44_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 98]> input_41_cast_fp16 = select(a = var_44_to_fp16, b = var_544_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_41_cast_fp16")];
+            tensor<bool, []> x_13_transpose_x_0 = const()[name = tensor<string, []>("x_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_13_transpose_y_0 = const()[name = tensor<string, []>("x_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_1_cast_fp16)[name = tensor<string, []>("transpose_359")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_13_cast_fp16")];
+            tensor<int32, [4]> var_548_perm_0 = const()[name = tensor<string, []>("op_548_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_549 = const()[name = tensor<string, []>("op_549"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_548_cast_fp16 = transpose(perm = var_548_perm_0, x = x_13_cast_fp16)[name = tensor<string, []>("transpose_358")];
+            tensor<fp16, [1, 56, 1024]> input_43_cast_fp16 = reshape(shape = var_549, x = var_548_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16392064))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17440704))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17442816)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_7_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
+            tensor<int32, [1]> x_17_axes_0 = const()[name = tensor<string, []>("x_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17444928)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17447040)))];
+            tensor<fp16, [1, 56, 1024]> x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("x_17_cast_fp16")];
+            tensor<int32, [3]> input_49_perm_0 = const()[name = tensor<string, []>("input_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_51_pad_type_0 = const()[name = tensor<string, []>("input_51_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_51_strides_0 = const()[name = tensor<string, []>("input_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_51_pad_0 = const()[name = tensor<string, []>("input_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_51_dilations_0 = const()[name = tensor<string, []>("input_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_51_groups_0 = const()[name = tensor<string, []>("input_51_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17449152))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19548480))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = tensor<string, []>("transpose_357")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_51_cast_fp16")];
+            tensor<int32, []> x_19_split_num_splits_0 = const()[name = tensor<string, []>("x_19_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_19_split_axis_0 = const()[name = tensor<string, []>("x_19_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_19_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_19_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = tensor<string, []>("x_19_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")];
+            tensor<int32, [1]> var_575_axes_0 = const()[name = tensor<string, []>("op_575_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 56]> var_575 = expand_dims(axes = var_575_axes_0, x = pad_mask)[name = tensor<string, []>("op_575")];
+            tensor<fp16, [1, 1024, 56]> input_53_cast_fp16 = select(a = var_44_to_fp16, b = x_19_cast_fp16, cond = var_575)[name = tensor<string, []>("input_53_cast_fp16")];
+            tensor<bool, []> new_x_3_interleave_0 = const()[name = tensor<string, []>("new_x_3_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_3_cast_fp16 = concat(axis = var_59, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = tensor<string, []>("new_x_3_cast_fp16")];
+            tensor<int32, [3]> var_588_begin_0 = const()[name = tensor<string, []>("op_588_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_588_end_0 = const()[name = tensor<string, []>("op_588_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_588_end_mask_0 = const()[name = tensor<string, []>("op_588_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_588_cast_fp16 = slice_by_index(begin = var_588_begin_0, end = var_588_end_0, end_mask = var_588_end_mask_0, x = new_x_3_cast_fp16)[name = tensor<string, []>("op_588_cast_fp16")];
+            tensor<string, []> x_21_pad_type_0 = const()[name = tensor<string, []>("x_21_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_21_groups_0 = const()[name = tensor<string, []>("x_21_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_21_strides_0 = const()[name = tensor<string, []>("x_21_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_21_pad_0 = const()[name = tensor<string, []>("x_21_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_21_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19552640))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19561920))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_21_cast_fp16")];
+            tensor<int32, [3]> input_55_perm_0 = const()[name = tensor<string, []>("input_55_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_23_axes_0 = const()[name = tensor<string, []>("x_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19564032)))];
+            tensor<fp16, [1024]> encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19566144)))];
+            tensor<fp16, [1, 56, 1024]> input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = tensor<string, []>("transpose_356")];
+            tensor<fp16, [1, 56, 1024]> x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
+            tensor<int32, [3]> input_57_perm_0 = const()[name = tensor<string, []>("input_57_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = tensor<string, []>("transpose_355")];
+            tensor<fp16, [1, 1024, 56]> input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
+            tensor<string, []> x_25_pad_type_0 = const()[name = tensor<string, []>("x_25_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_25_strides_0 = const()[name = tensor<string, []>("x_25_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_25_pad_0 = const()[name = tensor<string, []>("x_25_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_25_dilations_0 = const()[name = tensor<string, []>("x_25_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_25_groups_0 = const()[name = tensor<string, []>("x_25_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19568256))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20616896))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_25_cast_fp16")];
+            tensor<int32, [3]> input_61_perm_0 = const()[name = tensor<string, []>("input_61_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = tensor<string, []>("transpose_354")];
+            tensor<fp16, [1, 56, 1024]> input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
+            tensor<int32, [1]> input_65_axes_0 = const()[name = tensor<string, []>("input_65_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20619008)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20621120)))];
+            tensor<fp16, [1, 56, 1024]> input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20623232))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24817600))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805184)))];
+            tensor<fp16, [4096]> encoder_layers_0_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24825856)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_8_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24834112))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29028480))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_0_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29030592)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_9_cast_fp16")];
+            tensor<fp16, []> var_631_to_fp16 = const()[name = tensor<string, []>("op_631_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_632_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_631_to_fp16)[name = tensor<string, []>("op_632_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_632_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
+            tensor<int32, [1]> input_77_axes_0 = const()[name = tensor<string, []>("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29032704)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29034816)))];
+            tensor<fp16, [1, 56, 1024]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
+            tensor<int32, [4]> cache_5_begin_0 = const()[name = tensor<string, []>("cache_5_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_5_end_0 = const()[name = tensor<string, []>("cache_5_end_0"), val = tensor<int32, [4]>([2, 1, 42, 1024])];
+            tensor<bool, [4]> cache_5_end_mask_0 = const()[name = tensor<string, []>("cache_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_5_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_5_cast_fp16")];
+            tensor<int32, [4]> cache_7_begin_0 = const()[name = tensor<string, []>("cache_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_7_end_0 = const()[name = tensor<string, []>("cache_7_end_0"), val = tensor<int32, [4]>([2, 1, 1024, 8])];
+            tensor<bool, [4]> cache_7_end_mask_0 = const()[name = tensor<string, []>("cache_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_7_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_7_cast_fp16")];
+            tensor<int32, [1]> input_79_axes_0 = const()[name = tensor<string, []>("input_79_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29036928)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29039040)))];
+            tensor<fp16, [1, 56, 1024]> input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29041152))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33235520))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805184)))];
+            tensor<fp16, [4096]> encoder_layers_1_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33243776)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_10_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33252032))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37446400))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_1_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37448512)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_11_cast_fp16")];
+            tensor<fp16, []> var_668_to_fp16 = const()[name = tensor<string, []>("op_668_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_669_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_668_to_fp16)[name = tensor<string, []>("op_669_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_669_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
+            tensor<int32, [1]> key_3_axes_0 = const()[name = tensor<string, []>("key_3_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37450624)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37452736)))];
+            tensor<fp16, [1, 56, 1024]> key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("key_3_cast_fp16")];
+            tensor<bool, []> input_91_interleave_0 = const()[name = tensor<string, []>("input_91_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_91_cast_fp16 = concat(axis = var_68, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = tensor<string, []>("input_91_cast_fp16")];
+            tensor<bool, []> var_697_interleave_0 = const()[name = tensor<string, []>("op_697_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_697_cast_fp16 = concat(axis = var_68, interleave = var_697_interleave_0, values = key_3_cast_fp16)[name = tensor<string, []>("op_697_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37454848))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38503488))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38505600)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_12_cast_fp16")];
+            tensor<int32, [4]> var_702 = const()[name = tensor<string, []>("op_702"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_7_cast_fp16 = reshape(shape = var_702, x = linear_12_cast_fp16)[name = tensor<string, []>("q_7_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38507712))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39556352))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39558464)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_13_cast_fp16")];
+            tensor<int32, [4]> var_707 = const()[name = tensor<string, []>("op_707"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_5_cast_fp16 = reshape(shape = var_707, x = linear_13_cast_fp16)[name = tensor<string, []>("k_5_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39560576))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40609216))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40611328)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_14_cast_fp16")];
+            tensor<int32, [4]> var_712 = const()[name = tensor<string, []>("op_712"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_3_cast_fp16 = reshape(shape = var_712, x = linear_14_cast_fp16)[name = tensor<string, []>("v_3_cast_fp16")];
+            tensor<int32, [4]> value_11_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40613440)))];
+            tensor<fp16, [1, 56, 8, 128]> var_725_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_725_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40615552)))];
+            tensor<fp16, [1, 56, 8, 128]> var_727_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_727_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_3_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_33_transpose_x_0 = const()[name = tensor<string, []>("x_33_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_33_transpose_y_0 = const()[name = tensor<string, []>("x_33_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_729_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_729_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40617664))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40817408))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_727_cast_fp16)[name = tensor<string, []>("transpose_353")];
+            tensor<fp16, [1, 8, 56, 195]> 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_729_to_fp16_quantized)[name = tensor<string, []>("x_33_cast_fp16")];
+            tensor<int32, [8]> x_35_pad_0 = const()[name = tensor<string, []>("x_35_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_35_mode_0 = const()[name = tensor<string, []>("x_35_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_92_to_fp16 = const()[name = tensor<string, []>("const_92_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_35_cast_fp16")];
+            tensor<int32, [4]> var_737 = const()[name = tensor<string, []>("op_737"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_37_cast_fp16 = reshape(shape = var_737, x = x_35_cast_fp16)[name = tensor<string, []>("x_37_cast_fp16")];
+            tensor<int32, [4]> var_741_begin_0 = const()[name = tensor<string, []>("op_741_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_741_end_0 = const()[name = tensor<string, []>("op_741_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_741_end_mask_0 = const()[name = tensor<string, []>("op_741_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_741_cast_fp16 = slice_by_index(begin = var_741_begin_0, end = var_741_end_0, end_mask = var_741_end_mask_0, x = x_37_cast_fp16)[name = tensor<string, []>("op_741_cast_fp16")];
+            tensor<int32, [4]> var_742 = const()[name = tensor<string, []>("op_742"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_5_cast_fp16 = reshape(shape = var_742, x = var_741_cast_fp16)[name = tensor<string, []>("matrix_bd_5_cast_fp16")];
+            tensor<bool, []> matrix_ac_3_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_3_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_3_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_3_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_98_perm_0 = const()[name = tensor<string, []>("transpose_98_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_99_perm_0 = const()[name = tensor<string, []>("transpose_99_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5_cast_fp16)[name = tensor<string, []>("transpose_351")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_98 = transpose(perm = transpose_98_perm_0, x = var_725_cast_fp16)[name = tensor<string, []>("transpose_352")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_3_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_7_begin_0 = const()[name = tensor<string, []>("matrix_bd_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_7_end_0 = const()[name = tensor<string, []>("matrix_bd_7_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_7_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_7_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_751_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = tensor<string, []>("op_751_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_5_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_5_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_5_cast_fp16 = mul(x = var_751_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_5_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_7_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_7_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_757_cast_fp16 = softmax(axis = var_59, x = scores_7_cast_fp16)[name = tensor<string, []>("op_757_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_93_cast_fp16 = select(a = var_44_to_fp16, b = var_757_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_93_cast_fp16")];
+            tensor<bool, []> x_39_transpose_x_0 = const()[name = tensor<string, []>("x_39_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_39_transpose_y_0 = const()[name = tensor<string, []>("x_39_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_3_cast_fp16)[name = tensor<string, []>("transpose_350")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_39_cast_fp16")];
+            tensor<int32, [4]> var_761_perm_0 = const()[name = tensor<string, []>("op_761_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_762 = const()[name = tensor<string, []>("op_762"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_761_cast_fp16 = transpose(perm = var_761_perm_0, x = x_39_cast_fp16)[name = tensor<string, []>("transpose_349")];
+            tensor<fp16, [1, 56, 1024]> input_95_cast_fp16 = reshape(shape = var_762, x = var_761_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40817920))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41866560))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41868672)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_16_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
+            tensor<int32, [1]> x_43_axes_0 = const()[name = tensor<string, []>("x_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41870784)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41872896)))];
+            tensor<fp16, [1, 56, 1024]> x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")];
+            tensor<int32, [3]> input_101_perm_0 = const()[name = tensor<string, []>("input_101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_103_pad_type_0 = const()[name = tensor<string, []>("input_103_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_103_strides_0 = const()[name = tensor<string, []>("input_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_103_pad_0 = const()[name = tensor<string, []>("input_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_103_dilations_0 = const()[name = tensor<string, []>("input_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_103_groups_0 = const()[name = tensor<string, []>("input_103_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41875008))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43972224))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = tensor<string, []>("transpose_348")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_103_cast_fp16")];
+            tensor<int32, []> x_45_split_num_splits_0 = const()[name = tensor<string, []>("x_45_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_45_split_axis_0 = const()[name = tensor<string, []>("x_45_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_45_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_45_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = tensor<string, []>("x_45_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_45_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_105_cast_fp16 = select(a = var_44_to_fp16, b = x_45_cast_fp16, cond = var_575)[name = tensor<string, []>("input_105_cast_fp16")];
+            tensor<bool, []> new_x_7_interleave_0 = const()[name = tensor<string, []>("new_x_7_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_7_cast_fp16 = concat(axis = var_59, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = tensor<string, []>("new_x_7_cast_fp16")];
+            tensor<int32, [3]> var_801_begin_0 = const()[name = tensor<string, []>("op_801_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_801_end_0 = const()[name = tensor<string, []>("op_801_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_801_end_mask_0 = const()[name = tensor<string, []>("op_801_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_801_cast_fp16 = slice_by_index(begin = var_801_begin_0, end = var_801_end_0, end_mask = var_801_end_mask_0, x = new_x_7_cast_fp16)[name = tensor<string, []>("op_801_cast_fp16")];
+            tensor<string, []> x_47_pad_type_0 = const()[name = tensor<string, []>("x_47_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_47_groups_0 = const()[name = tensor<string, []>("x_47_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_47_strides_0 = const()[name = tensor<string, []>("x_47_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_47_pad_0 = const()[name = tensor<string, []>("x_47_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_47_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43976384))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43985664))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_47_cast_fp16")];
+            tensor<int32, [3]> input_107_perm_0 = const()[name = tensor<string, []>("input_107_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_49_axes_0 = const()[name = tensor<string, []>("x_49_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43987776)))];
+            tensor<fp16, [1024]> encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43989888)))];
+            tensor<fp16, [1, 56, 1024]> input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = tensor<string, []>("transpose_347")];
+            tensor<fp16, [1, 56, 1024]> x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("x_49_cast_fp16")];
+            tensor<int32, [3]> input_109_perm_0 = const()[name = tensor<string, []>("input_109_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = tensor<string, []>("transpose_346")];
+            tensor<fp16, [1, 1024, 56]> input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
+            tensor<string, []> x_51_pad_type_0 = const()[name = tensor<string, []>("x_51_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_51_strides_0 = const()[name = tensor<string, []>("x_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_51_pad_0 = const()[name = tensor<string, []>("x_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_51_dilations_0 = const()[name = tensor<string, []>("x_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_51_groups_0 = const()[name = tensor<string, []>("x_51_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43992000))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45040640))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_51_cast_fp16")];
+            tensor<int32, [3]> input_113_perm_0 = const()[name = tensor<string, []>("input_113_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = tensor<string, []>("transpose_345")];
+            tensor<fp16, [1, 56, 1024]> input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
+            tensor<int32, [1]> input_117_axes_0 = const()[name = tensor<string, []>("input_117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45042752)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45044864)))];
+            tensor<fp16, [1, 56, 1024]> input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45046976))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49241344))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805184)))];
+            tensor<fp16, [4096]> encoder_layers_1_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49249600)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_17_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49257856))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53452224))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_1_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53454336)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_18_cast_fp16")];
+            tensor<fp16, []> var_844_to_fp16 = const()[name = tensor<string, []>("op_844_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_845_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_844_to_fp16)[name = tensor<string, []>("op_845_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_845_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
+            tensor<int32, [1]> input_129_axes_0 = const()[name = tensor<string, []>("input_129_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53456448)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53458560)))];
+            tensor<fp16, [1, 56, 1024]> input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
+            tensor<int32, [4]> cache_9_begin_0 = const()[name = tensor<string, []>("cache_9_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_9_end_0 = const()[name = tensor<string, []>("cache_9_end_0"), val = tensor<int32, [4]>([3, 1, 42, 1024])];
+            tensor<bool, [4]> cache_9_end_mask_0 = const()[name = tensor<string, []>("cache_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_9_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_9_cast_fp16")];
+            tensor<int32, [4]> cache_11_begin_0 = const()[name = tensor<string, []>("cache_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_11_end_0 = const()[name = tensor<string, []>("cache_11_end_0"), val = tensor<int32, [4]>([3, 1, 1024, 8])];
+            tensor<bool, [4]> cache_11_end_mask_0 = const()[name = tensor<string, []>("cache_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_11_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_11_cast_fp16")];
+            tensor<int32, [1]> input_131_axes_0 = const()[name = tensor<string, []>("input_131_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53460672)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53462784)))];
+            tensor<fp16, [1, 56, 1024]> input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53464896))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57659264))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805184)))];
+            tensor<fp16, [4096]> encoder_layers_2_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57667520)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_19_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57675776))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61870144))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_2_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61872256)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_20_cast_fp16")];
+            tensor<fp16, []> var_881_to_fp16 = const()[name = tensor<string, []>("op_881_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_882_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_881_to_fp16)[name = tensor<string, []>("op_882_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_882_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
+            tensor<int32, [1]> key_5_axes_0 = const()[name = tensor<string, []>("key_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61874368)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61876480)))];
+            tensor<fp16, [1, 56, 1024]> key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")];
+            tensor<bool, []> input_143_interleave_0 = const()[name = tensor<string, []>("input_143_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_143_cast_fp16 = concat(axis = var_68, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = tensor<string, []>("input_143_cast_fp16")];
+            tensor<bool, []> var_910_interleave_0 = const()[name = tensor<string, []>("op_910_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_910_cast_fp16 = concat(axis = var_68, interleave = var_910_interleave_0, values = key_5_cast_fp16)[name = tensor<string, []>("op_910_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61878592))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62927232))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62929344)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_21_cast_fp16")];
+            tensor<int32, [4]> var_915 = const()[name = tensor<string, []>("op_915"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_13_cast_fp16 = reshape(shape = var_915, x = linear_21_cast_fp16)[name = tensor<string, []>("q_13_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62931456))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63980096))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63982208)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_22_cast_fp16")];
+            tensor<int32, [4]> var_920 = const()[name = tensor<string, []>("op_920"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_9_cast_fp16 = reshape(shape = var_920, x = linear_22_cast_fp16)[name = tensor<string, []>("k_9_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63984320))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65032960))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65035072)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_23_cast_fp16")];
+            tensor<int32, [4]> var_925 = const()[name = tensor<string, []>("op_925"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_5_cast_fp16 = reshape(shape = var_925, x = linear_23_cast_fp16)[name = tensor<string, []>("v_5_cast_fp16")];
+            tensor<int32, [4]> value_13_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65037184)))];
+            tensor<fp16, [1, 56, 8, 128]> var_938_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_938_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65039296)))];
+            tensor<fp16, [1, 56, 8, 128]> var_940_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_940_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_5_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_5_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_59_transpose_x_0 = const()[name = tensor<string, []>("x_59_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_59_transpose_y_0 = const()[name = tensor<string, []>("x_59_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_942_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_942_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65041408))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65241152))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_940_cast_fp16)[name = tensor<string, []>("transpose_344")];
+            tensor<fp16, [1, 8, 56, 195]> 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_942_to_fp16_quantized)[name = tensor<string, []>("x_59_cast_fp16")];
+            tensor<int32, [8]> x_61_pad_0 = const()[name = tensor<string, []>("x_61_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_61_mode_0 = const()[name = tensor<string, []>("x_61_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_105_to_fp16 = const()[name = tensor<string, []>("const_105_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_61_cast_fp16")];
+            tensor<int32, [4]> var_950 = const()[name = tensor<string, []>("op_950"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_63_cast_fp16 = reshape(shape = var_950, x = x_61_cast_fp16)[name = tensor<string, []>("x_63_cast_fp16")];
+            tensor<int32, [4]> var_954_begin_0 = const()[name = tensor<string, []>("op_954_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_954_end_0 = const()[name = tensor<string, []>("op_954_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_954_end_mask_0 = const()[name = tensor<string, []>("op_954_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_954_cast_fp16 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, x = x_63_cast_fp16)[name = tensor<string, []>("op_954_cast_fp16")];
+            tensor<int32, [4]> var_955 = const()[name = tensor<string, []>("op_955"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_9_cast_fp16 = reshape(shape = var_955, x = var_954_cast_fp16)[name = tensor<string, []>("matrix_bd_9_cast_fp16")];
+            tensor<bool, []> matrix_ac_5_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_5_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_5_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_5_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_100_perm_0 = const()[name = tensor<string, []>("transpose_100_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_101_perm_0 = const()[name = tensor<string, []>("transpose_101_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9_cast_fp16)[name = tensor<string, []>("transpose_342")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_100 = transpose(perm = transpose_100_perm_0, x = var_938_cast_fp16)[name = tensor<string, []>("transpose_343")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_5_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_11_begin_0 = const()[name = tensor<string, []>("matrix_bd_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_11_end_0 = const()[name = tensor<string, []>("matrix_bd_11_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_11_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_11_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_964_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = tensor<string, []>("op_964_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_9_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_9_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_9_cast_fp16 = mul(x = var_964_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_9_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_11_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_11_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_970_cast_fp16 = softmax(axis = var_59, x = scores_11_cast_fp16)[name = tensor<string, []>("op_970_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_145_cast_fp16 = select(a = var_44_to_fp16, b = var_970_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_145_cast_fp16")];
+            tensor<bool, []> x_65_transpose_x_0 = const()[name = tensor<string, []>("x_65_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_65_transpose_y_0 = const()[name = tensor<string, []>("x_65_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_5_cast_fp16)[name = tensor<string, []>("transpose_341")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_65_cast_fp16")];
+            tensor<int32, [4]> var_974_perm_0 = const()[name = tensor<string, []>("op_974_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_975 = const()[name = tensor<string, []>("op_975"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_974_cast_fp16 = transpose(perm = var_974_perm_0, x = x_65_cast_fp16)[name = tensor<string, []>("transpose_340")];
+            tensor<fp16, [1, 56, 1024]> input_147_cast_fp16 = reshape(shape = var_975, x = var_974_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65241664))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66028160))), name = tensor<string, []>("encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66028352)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_25_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")];
+            tensor<int32, [1]> x_69_axes_0 = const()[name = tensor<string, []>("x_69_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66030464)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66032576)))];
+            tensor<fp16, [1, 56, 1024]> x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = tensor<string, []>("x_69_cast_fp16")];
+            tensor<int32, [3]> input_153_perm_0 = const()[name = tensor<string, []>("input_153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_155_pad_type_0 = const()[name = tensor<string, []>("input_155_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_155_strides_0 = const()[name = tensor<string, []>("input_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_155_pad_0 = const()[name = tensor<string, []>("input_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_155_dilations_0 = const()[name = tensor<string, []>("input_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_155_groups_0 = const()[name = tensor<string, []>("input_155_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66034688))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68131904))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = tensor<string, []>("transpose_339")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_155_cast_fp16")];
+            tensor<int32, []> x_71_split_num_splits_0 = const()[name = tensor<string, []>("x_71_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_71_split_axis_0 = const()[name = tensor<string, []>("x_71_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_71_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_71_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = tensor<string, []>("x_71_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_71_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_157_cast_fp16 = select(a = var_44_to_fp16, b = x_71_cast_fp16, cond = var_575)[name = tensor<string, []>("input_157_cast_fp16")];
+            tensor<bool, []> new_x_11_interleave_0 = const()[name = tensor<string, []>("new_x_11_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_11_cast_fp16 = concat(axis = var_59, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = tensor<string, []>("new_x_11_cast_fp16")];
+            tensor<int32, [3]> var_1014_begin_0 = const()[name = tensor<string, []>("op_1014_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_1014_end_0 = const()[name = tensor<string, []>("op_1014_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_1014_end_mask_0 = const()[name = tensor<string, []>("op_1014_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1014_cast_fp16 = slice_by_index(begin = var_1014_begin_0, end = var_1014_end_0, end_mask = var_1014_end_mask_0, x = new_x_11_cast_fp16)[name = tensor<string, []>("op_1014_cast_fp16")];
+            tensor<string, []> x_73_pad_type_0 = const()[name = tensor<string, []>("x_73_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_73_groups_0 = const()[name = tensor<string, []>("x_73_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_73_strides_0 = const()[name = tensor<string, []>("x_73_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_73_pad_0 = const()[name = tensor<string, []>("x_73_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_73_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68136064))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68145344))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_73_cast_fp16")];
+            tensor<int32, [3]> input_159_perm_0 = const()[name = tensor<string, []>("input_159_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_75_axes_0 = const()[name = tensor<string, []>("x_75_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68147456)))];
+            tensor<fp16, [1024]> encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68149568)))];
+            tensor<fp16, [1, 56, 1024]> input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = tensor<string, []>("transpose_338")];
+            tensor<fp16, [1, 56, 1024]> x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = tensor<string, []>("x_75_cast_fp16")];
+            tensor<int32, [3]> input_161_perm_0 = const()[name = tensor<string, []>("input_161_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = tensor<string, []>("transpose_337")];
+            tensor<fp16, [1, 1024, 56]> input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
+            tensor<string, []> x_77_pad_type_0 = const()[name = tensor<string, []>("x_77_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_77_strides_0 = const()[name = tensor<string, []>("x_77_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_77_pad_0 = const()[name = tensor<string, []>("x_77_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_77_dilations_0 = const()[name = tensor<string, []>("x_77_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_77_groups_0 = const()[name = tensor<string, []>("x_77_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68151680))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69200320))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_77_cast_fp16")];
+            tensor<int32, [3]> input_165_perm_0 = const()[name = tensor<string, []>("input_165_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = tensor<string, []>("transpose_336")];
+            tensor<fp16, [1, 56, 1024]> input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
+            tensor<int32, [1]> input_169_axes_0 = const()[name = tensor<string, []>("input_169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69202432)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69204544)))];
+            tensor<fp16, [1, 56, 1024]> input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69206656))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72352448))), name = tensor<string, []>("encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_2_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72352640)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_26_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72360896))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75506688))), name = tensor<string, []>("encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_2_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75506880)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_27_cast_fp16")];
+            tensor<fp16, []> var_1057_to_fp16 = const()[name = tensor<string, []>("op_1057_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1058_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1057_to_fp16)[name = tensor<string, []>("op_1058_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_1058_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
+            tensor<int32, [1]> input_181_axes_0 = const()[name = tensor<string, []>("input_181_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75508992)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75511104)))];
+            tensor<fp16, [1, 56, 1024]> input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
+            tensor<int32, [4]> cache_13_begin_0 = const()[name = tensor<string, []>("cache_13_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_13_end_0 = const()[name = tensor<string, []>("cache_13_end_0"), val = tensor<int32, [4]>([4, 1, 42, 1024])];
+            tensor<bool, [4]> cache_13_end_mask_0 = const()[name = tensor<string, []>("cache_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_13_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_13_cast_fp16")];
+            tensor<int32, [4]> cache_15_begin_0 = const()[name = tensor<string, []>("cache_15_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_15_end_0 = const()[name = tensor<string, []>("cache_15_end_0"), val = tensor<int32, [4]>([4, 1, 1024, 8])];
+            tensor<bool, [4]> cache_15_end_mask_0 = const()[name = tensor<string, []>("cache_15_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_15_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_15_cast_fp16")];
+            tensor<int32, [1]> input_183_axes_0 = const()[name = tensor<string, []>("input_183_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75513216)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75515328)))];
+            tensor<fp16, [1, 56, 1024]> input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75517440))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78663232))), name = tensor<string, []>("encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_3_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78663424)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_28_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78671680))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81817472))), name = tensor<string, []>("encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_3_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81817664)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_29_cast_fp16")];
+            tensor<fp16, []> var_1094_to_fp16 = const()[name = tensor<string, []>("op_1094_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1095_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1094_to_fp16)[name = tensor<string, []>("op_1095_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_1095_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
+            tensor<int32, [1]> key_7_axes_0 = const()[name = tensor<string, []>("key_7_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81819776)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81821888)))];
+            tensor<fp16, [1, 56, 1024]> key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = tensor<string, []>("key_7_cast_fp16")];
+            tensor<bool, []> input_195_interleave_0 = const()[name = tensor<string, []>("input_195_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_195_cast_fp16 = concat(axis = var_68, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = tensor<string, []>("input_195_cast_fp16")];
+            tensor<bool, []> var_1123_interleave_0 = const()[name = tensor<string, []>("op_1123_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_1123_cast_fp16 = concat(axis = var_68, interleave = var_1123_interleave_0, values = key_7_cast_fp16)[name = tensor<string, []>("op_1123_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81824000))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82610496))), name = tensor<string, []>("encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82610688)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_30_cast_fp16")];
+            tensor<int32, [4]> var_1128 = const()[name = tensor<string, []>("op_1128"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_19_cast_fp16 = reshape(shape = var_1128, x = linear_30_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82612800))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83399296))), name = tensor<string, []>("encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83399488)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_31_cast_fp16")];
+            tensor<int32, [4]> var_1133 = const()[name = tensor<string, []>("op_1133"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_13_cast_fp16 = reshape(shape = var_1133, x = linear_31_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83401600))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84188096))), name = tensor<string, []>("encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84188288)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_32_cast_fp16")];
+            tensor<int32, [4]> var_1138 = const()[name = tensor<string, []>("op_1138"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_7_cast_fp16 = reshape(shape = var_1138, x = linear_32_cast_fp16)[name = tensor<string, []>("v_7_cast_fp16")];
+            tensor<int32, [4]> value_15_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84190400)))];
+            tensor<fp16, [1, 56, 8, 128]> var_1151_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1151_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84192512)))];
+            tensor<fp16, [1, 56, 8, 128]> var_1153_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1153_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_7_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_85_transpose_x_0 = const()[name = tensor<string, []>("x_85_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_85_transpose_y_0 = const()[name = tensor<string, []>("x_85_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_1155_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1155_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84194624))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84394368))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1153_cast_fp16)[name = tensor<string, []>("transpose_335")];
+            tensor<fp16, [1, 8, 56, 195]> 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_1155_to_fp16_quantized)[name = tensor<string, []>("x_85_cast_fp16")];
+            tensor<int32, [8]> x_87_pad_0 = const()[name = tensor<string, []>("x_87_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_87_mode_0 = const()[name = tensor<string, []>("x_87_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_118_to_fp16 = const()[name = tensor<string, []>("const_118_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_87_cast_fp16")];
+            tensor<int32, [4]> var_1163 = const()[name = tensor<string, []>("op_1163"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_89_cast_fp16 = reshape(shape = var_1163, x = x_87_cast_fp16)[name = tensor<string, []>("x_89_cast_fp16")];
+            tensor<int32, [4]> var_1167_begin_0 = const()[name = tensor<string, []>("op_1167_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1167_end_0 = const()[name = tensor<string, []>("op_1167_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_1167_end_mask_0 = const()[name = tensor<string, []>("op_1167_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_1167_cast_fp16 = slice_by_index(begin = var_1167_begin_0, end = var_1167_end_0, end_mask = var_1167_end_mask_0, x = x_89_cast_fp16)[name = tensor<string, []>("op_1167_cast_fp16")];
+            tensor<int32, [4]> var_1168 = const()[name = tensor<string, []>("op_1168"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_13_cast_fp16 = reshape(shape = var_1168, x = var_1167_cast_fp16)[name = tensor<string, []>("matrix_bd_13_cast_fp16")];
+            tensor<bool, []> matrix_ac_7_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_7_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_102_perm_0 = const()[name = tensor<string, []>("transpose_102_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_103_perm_0 = const()[name = tensor<string, []>("transpose_103_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13_cast_fp16)[name = tensor<string, []>("transpose_333")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1151_cast_fp16)[name = tensor<string, []>("transpose_334")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_7_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_15_begin_0 = const()[name = tensor<string, []>("matrix_bd_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_15_end_0 = const()[name = tensor<string, []>("matrix_bd_15_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_15_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_15_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_1177_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = tensor<string, []>("op_1177_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_13_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_13_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_13_cast_fp16 = mul(x = var_1177_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_13_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_15_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_15_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_1183_cast_fp16 = softmax(axis = var_59, x = scores_15_cast_fp16)[name = tensor<string, []>("op_1183_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_197_cast_fp16 = select(a = var_44_to_fp16, b = var_1183_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_197_cast_fp16")];
+            tensor<bool, []> x_91_transpose_x_0 = const()[name = tensor<string, []>("x_91_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_91_transpose_y_0 = const()[name = tensor<string, []>("x_91_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_7_cast_fp16)[name = tensor<string, []>("transpose_332")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_91_cast_fp16")];
+            tensor<int32, [4]> var_1187_perm_0 = const()[name = tensor<string, []>("op_1187_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1188 = const()[name = tensor<string, []>("op_1188"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_1187_cast_fp16 = transpose(perm = var_1187_perm_0, x = x_91_cast_fp16)[name = tensor<string, []>("transpose_331")];
+            tensor<fp16, [1, 56, 1024]> input_199_cast_fp16 = reshape(shape = var_1188, x = var_1187_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84394880))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85181376))), name = tensor<string, []>("encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85181568)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_34_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
+            tensor<int32, [1]> x_95_axes_0 = const()[name = tensor<string, []>("x_95_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85183680)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85185792)))];
+            tensor<fp16, [1, 56, 1024]> x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = tensor<string, []>("x_95_cast_fp16")];
+            tensor<int32, [3]> input_205_perm_0 = const()[name = tensor<string, []>("input_205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_207_pad_type_0 = const()[name = tensor<string, []>("input_207_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_207_strides_0 = const()[name = tensor<string, []>("input_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_207_pad_0 = const()[name = tensor<string, []>("input_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_207_dilations_0 = const()[name = tensor<string, []>("input_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_207_groups_0 = const()[name = tensor<string, []>("input_207_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85187904))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87285120))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = tensor<string, []>("transpose_330")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_207_cast_fp16")];
+            tensor<int32, []> x_97_split_num_splits_0 = const()[name = tensor<string, []>("x_97_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_97_split_axis_0 = const()[name = tensor<string, []>("x_97_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_97_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_97_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = tensor<string, []>("x_97_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_97_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_209_cast_fp16 = select(a = var_44_to_fp16, b = x_97_cast_fp16, cond = var_575)[name = tensor<string, []>("input_209_cast_fp16")];
+            tensor<bool, []> new_x_15_interleave_0 = const()[name = tensor<string, []>("new_x_15_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_15_cast_fp16 = concat(axis = var_59, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = tensor<string, []>("new_x_15_cast_fp16")];
+            tensor<int32, [3]> var_1227_begin_0 = const()[name = tensor<string, []>("op_1227_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_1227_end_0 = const()[name = tensor<string, []>("op_1227_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_1227_end_mask_0 = const()[name = tensor<string, []>("op_1227_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1227_cast_fp16 = slice_by_index(begin = var_1227_begin_0, end = var_1227_end_0, end_mask = var_1227_end_mask_0, x = new_x_15_cast_fp16)[name = tensor<string, []>("op_1227_cast_fp16")];
+            tensor<string, []> x_99_pad_type_0 = const()[name = tensor<string, []>("x_99_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_99_groups_0 = const()[name = tensor<string, []>("x_99_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_99_strides_0 = const()[name = tensor<string, []>("x_99_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_99_pad_0 = const()[name = tensor<string, []>("x_99_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_99_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87289280))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87298560))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_99_cast_fp16")];
+            tensor<int32, [3]> input_211_perm_0 = const()[name = tensor<string, []>("input_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_101_axes_0 = const()[name = tensor<string, []>("x_101_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87300672)))];
+            tensor<fp16, [1024]> encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87302784)))];
+            tensor<fp16, [1, 56, 1024]> input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = tensor<string, []>("transpose_329")];
+            tensor<fp16, [1, 56, 1024]> x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = tensor<string, []>("x_101_cast_fp16")];
+            tensor<int32, [3]> input_213_perm_0 = const()[name = tensor<string, []>("input_213_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = tensor<string, []>("transpose_328")];
+            tensor<fp16, [1, 1024, 56]> input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")];
+            tensor<string, []> x_103_pad_type_0 = const()[name = tensor<string, []>("x_103_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_103_strides_0 = const()[name = tensor<string, []>("x_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_103_pad_0 = const()[name = tensor<string, []>("x_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_103_dilations_0 = const()[name = tensor<string, []>("x_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_103_groups_0 = const()[name = tensor<string, []>("x_103_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87304896))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88353536))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_103_cast_fp16")];
+            tensor<int32, [3]> input_217_perm_0 = const()[name = tensor<string, []>("input_217_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = tensor<string, []>("transpose_327")];
+            tensor<fp16, [1, 56, 1024]> input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
+            tensor<int32, [1]> input_221_axes_0 = const()[name = tensor<string, []>("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88355648)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88357760)))];
+            tensor<fp16, [1, 56, 1024]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88359872))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91505664))), name = tensor<string, []>("encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_3_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91505856)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_35_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91514112))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94659904))), name = tensor<string, []>("encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_3_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94660096)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_36_cast_fp16")];
+            tensor<fp16, []> var_1270_to_fp16 = const()[name = tensor<string, []>("op_1270_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1271_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1270_to_fp16)[name = tensor<string, []>("op_1271_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1271_cast_fp16)[name = tensor<string, []>("input_231_cast_fp16")];
+            tensor<int32, [1]> input_233_axes_0 = const()[name = tensor<string, []>("input_233_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94662208)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94664320)))];
+            tensor<fp16, [1, 56, 1024]> input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")];
+            tensor<int32, [4]> cache_17_begin_0 = const()[name = tensor<string, []>("cache_17_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_17_end_0 = const()[name = tensor<string, []>("cache_17_end_0"), val = tensor<int32, [4]>([5, 1, 42, 1024])];
+            tensor<bool, [4]> cache_17_end_mask_0 = const()[name = tensor<string, []>("cache_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_17_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_17_cast_fp16")];
+            tensor<int32, [4]> cache_19_begin_0 = const()[name = tensor<string, []>("cache_19_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_19_end_0 = const()[name = tensor<string, []>("cache_19_end_0"), val = tensor<int32, [4]>([5, 1, 1024, 8])];
+            tensor<bool, [4]> cache_19_end_mask_0 = const()[name = tensor<string, []>("cache_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_19_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_19_cast_fp16")];
+            tensor<int32, [1]> input_235_axes_0 = const()[name = tensor<string, []>("input_235_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94666432)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94668544)))];
+            tensor<fp16, [1, 56, 1024]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94670656))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97816448))), name = tensor<string, []>("encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_4_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97816640)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_37_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97824896))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100970688))), name = tensor<string, []>("encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_4_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100970880)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_38_cast_fp16")];
+            tensor<fp16, []> var_1307_to_fp16 = const()[name = tensor<string, []>("op_1307_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1308_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1307_to_fp16)[name = tensor<string, []>("op_1308_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1308_cast_fp16)[name = tensor<string, []>("input_245_cast_fp16")];
+            tensor<int32, [1]> key_9_axes_0 = const()[name = tensor<string, []>("key_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100972992)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100975104)))];
+            tensor<fp16, [1, 56, 1024]> key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")];
+            tensor<bool, []> input_247_interleave_0 = const()[name = tensor<string, []>("input_247_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_247_cast_fp16 = concat(axis = var_68, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = tensor<string, []>("input_247_cast_fp16")];
+            tensor<bool, []> var_1336_interleave_0 = const()[name = tensor<string, []>("op_1336_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_1336_cast_fp16 = concat(axis = var_68, interleave = var_1336_interleave_0, values = key_9_cast_fp16)[name = tensor<string, []>("op_1336_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100977216))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101763712))), name = tensor<string, []>("encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101763904)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_39_cast_fp16")];
+            tensor<int32, [4]> var_1341 = const()[name = tensor<string, []>("op_1341"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_25_cast_fp16 = reshape(shape = var_1341, x = linear_39_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101766016))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102552512))), name = tensor<string, []>("encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102552704)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_40_cast_fp16")];
+            tensor<int32, [4]> var_1346 = const()[name = tensor<string, []>("op_1346"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_17_cast_fp16 = reshape(shape = var_1346, x = linear_40_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102554816))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103341312))), name = tensor<string, []>("encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103341504)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_41_cast_fp16")];
+            tensor<int32, [4]> var_1351 = const()[name = tensor<string, []>("op_1351"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_9_cast_fp16 = reshape(shape = var_1351, x = linear_41_cast_fp16)[name = tensor<string, []>("v_9_cast_fp16")];
+            tensor<int32, [4]> value_17_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103343616)))];
+            tensor<fp16, [1, 56, 8, 128]> var_1364_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1364_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103345728)))];
+            tensor<fp16, [1, 56, 8, 128]> var_1366_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1366_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_9_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_111_transpose_x_0 = const()[name = tensor<string, []>("x_111_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_111_transpose_y_0 = const()[name = tensor<string, []>("x_111_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_1368_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1368_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103347840))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103547584))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1366_cast_fp16)[name = tensor<string, []>("transpose_326")];
+            tensor<fp16, [1, 8, 56, 195]> 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_1368_to_fp16_quantized)[name = tensor<string, []>("x_111_cast_fp16")];
+            tensor<int32, [8]> x_113_pad_0 = const()[name = tensor<string, []>("x_113_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_113_mode_0 = const()[name = tensor<string, []>("x_113_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_131_to_fp16 = const()[name = tensor<string, []>("const_131_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_113_cast_fp16")];
+            tensor<int32, [4]> var_1376 = const()[name = tensor<string, []>("op_1376"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_115_cast_fp16 = reshape(shape = var_1376, x = x_113_cast_fp16)[name = tensor<string, []>("x_115_cast_fp16")];
+            tensor<int32, [4]> var_1380_begin_0 = const()[name = tensor<string, []>("op_1380_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1380_end_0 = const()[name = tensor<string, []>("op_1380_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_1380_end_mask_0 = const()[name = tensor<string, []>("op_1380_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_1380_cast_fp16 = slice_by_index(begin = var_1380_begin_0, end = var_1380_end_0, end_mask = var_1380_end_mask_0, x = x_115_cast_fp16)[name = tensor<string, []>("op_1380_cast_fp16")];
+            tensor<int32, [4]> var_1381 = const()[name = tensor<string, []>("op_1381"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_17_cast_fp16 = reshape(shape = var_1381, x = var_1380_cast_fp16)[name = tensor<string, []>("matrix_bd_17_cast_fp16")];
+            tensor<bool, []> matrix_ac_9_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_9_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_9_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_9_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_104_perm_0 = const()[name = tensor<string, []>("transpose_104_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_105_perm_0 = const()[name = tensor<string, []>("transpose_105_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17_cast_fp16)[name = tensor<string, []>("transpose_324")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1364_cast_fp16)[name = tensor<string, []>("transpose_325")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_9_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_19_begin_0 = const()[name = tensor<string, []>("matrix_bd_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_19_end_0 = const()[name = tensor<string, []>("matrix_bd_19_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_19_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_19_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_1390_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = tensor<string, []>("op_1390_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_17_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_17_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_17_cast_fp16 = mul(x = var_1390_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_17_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_19_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_19_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_1396_cast_fp16 = softmax(axis = var_59, x = scores_19_cast_fp16)[name = tensor<string, []>("op_1396_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_249_cast_fp16 = select(a = var_44_to_fp16, b = var_1396_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_249_cast_fp16")];
+            tensor<bool, []> x_117_transpose_x_0 = const()[name = tensor<string, []>("x_117_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_117_transpose_y_0 = const()[name = tensor<string, []>("x_117_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_9_cast_fp16)[name = tensor<string, []>("transpose_323")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_117_cast_fp16")];
+            tensor<int32, [4]> var_1400_perm_0 = const()[name = tensor<string, []>("op_1400_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1401 = const()[name = tensor<string, []>("op_1401"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_1400_cast_fp16 = transpose(perm = var_1400_perm_0, x = x_117_cast_fp16)[name = tensor<string, []>("transpose_322")];
+            tensor<fp16, [1, 56, 1024]> input_251_cast_fp16 = reshape(shape = var_1401, x = var_1400_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103548096))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104334592))), name = tensor<string, []>("encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104334784)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_43_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = tensor<string, []>("input_255_cast_fp16")];
+            tensor<int32, [1]> x_121_axes_0 = const()[name = tensor<string, []>("x_121_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104336896)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104339008)))];
+            tensor<fp16, [1, 56, 1024]> x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = tensor<string, []>("x_121_cast_fp16")];
+            tensor<int32, [3]> input_257_perm_0 = const()[name = tensor<string, []>("input_257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_259_pad_type_0 = const()[name = tensor<string, []>("input_259_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_259_strides_0 = const()[name = tensor<string, []>("input_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_259_pad_0 = const()[name = tensor<string, []>("input_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_259_dilations_0 = const()[name = tensor<string, []>("input_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_259_groups_0 = const()[name = tensor<string, []>("input_259_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104341120))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106438336))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = tensor<string, []>("transpose_321")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_259_cast_fp16")];
+            tensor<int32, []> x_123_split_num_splits_0 = const()[name = tensor<string, []>("x_123_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_123_split_axis_0 = const()[name = tensor<string, []>("x_123_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_123_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_123_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = tensor<string, []>("x_123_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_123_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_261_cast_fp16 = select(a = var_44_to_fp16, b = x_123_cast_fp16, cond = var_575)[name = tensor<string, []>("input_261_cast_fp16")];
+            tensor<bool, []> new_x_19_interleave_0 = const()[name = tensor<string, []>("new_x_19_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_19_cast_fp16 = concat(axis = var_59, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = tensor<string, []>("new_x_19_cast_fp16")];
+            tensor<int32, [3]> var_1440_begin_0 = const()[name = tensor<string, []>("op_1440_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_1440_end_0 = const()[name = tensor<string, []>("op_1440_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_1440_end_mask_0 = const()[name = tensor<string, []>("op_1440_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1440_cast_fp16 = slice_by_index(begin = var_1440_begin_0, end = var_1440_end_0, end_mask = var_1440_end_mask_0, x = new_x_19_cast_fp16)[name = tensor<string, []>("op_1440_cast_fp16")];
+            tensor<string, []> x_125_pad_type_0 = const()[name = tensor<string, []>("x_125_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_125_groups_0 = const()[name = tensor<string, []>("x_125_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_125_strides_0 = const()[name = tensor<string, []>("x_125_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_125_pad_0 = const()[name = tensor<string, []>("x_125_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_125_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106442496))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106451776))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_125_cast_fp16")];
+            tensor<int32, [3]> input_263_perm_0 = const()[name = tensor<string, []>("input_263_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_127_axes_0 = const()[name = tensor<string, []>("x_127_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106453888)))];
+            tensor<fp16, [1024]> encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106456000)))];
+            tensor<fp16, [1, 56, 1024]> input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = tensor<string, []>("transpose_320")];
+            tensor<fp16, [1, 56, 1024]> x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = tensor<string, []>("x_127_cast_fp16")];
+            tensor<int32, [3]> input_265_perm_0 = const()[name = tensor<string, []>("input_265_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = tensor<string, []>("transpose_319")];
+            tensor<fp16, [1, 1024, 56]> input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = tensor<string, []>("input_267_cast_fp16")];
+            tensor<string, []> x_129_pad_type_0 = const()[name = tensor<string, []>("x_129_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_129_strides_0 = const()[name = tensor<string, []>("x_129_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_129_pad_0 = const()[name = tensor<string, []>("x_129_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_129_dilations_0 = const()[name = tensor<string, []>("x_129_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_129_groups_0 = const()[name = tensor<string, []>("x_129_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106458112))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107506752))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_129_cast_fp16")];
+            tensor<int32, [3]> input_269_perm_0 = const()[name = tensor<string, []>("input_269_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = tensor<string, []>("transpose_318")];
+            tensor<fp16, [1, 56, 1024]> input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = tensor<string, []>("input_271_cast_fp16")];
+            tensor<int32, [1]> input_273_axes_0 = const()[name = tensor<string, []>("input_273_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107508864)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107510976)))];
+            tensor<fp16, [1, 56, 1024]> input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107513088))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110658880))), name = tensor<string, []>("encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_4_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110659072)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_44_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110667328))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113813120))), name = tensor<string, []>("encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_4_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113813312)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_45_cast_fp16")];
+            tensor<fp16, []> var_1483_to_fp16 = const()[name = tensor<string, []>("op_1483_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1484_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1483_to_fp16)[name = tensor<string, []>("op_1484_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1484_cast_fp16)[name = tensor<string, []>("input_283_cast_fp16")];
+            tensor<int32, [1]> input_285_axes_0 = const()[name = tensor<string, []>("input_285_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113815424)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113817536)))];
+            tensor<fp16, [1, 56, 1024]> input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = tensor<string, []>("input_285_cast_fp16")];
+            tensor<int32, [4]> cache_21_begin_0 = const()[name = tensor<string, []>("cache_21_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_21_end_0 = const()[name = tensor<string, []>("cache_21_end_0"), val = tensor<int32, [4]>([6, 1, 42, 1024])];
+            tensor<bool, [4]> cache_21_end_mask_0 = const()[name = tensor<string, []>("cache_21_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_21_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_21_cast_fp16")];
+            tensor<int32, [4]> cache_23_begin_0 = const()[name = tensor<string, []>("cache_23_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_23_end_0 = const()[name = tensor<string, []>("cache_23_end_0"), val = tensor<int32, [4]>([6, 1, 1024, 8])];
+            tensor<bool, [4]> cache_23_end_mask_0 = const()[name = tensor<string, []>("cache_23_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_23_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_23_cast_fp16")];
+            tensor<int32, [1]> input_287_axes_0 = const()[name = tensor<string, []>("input_287_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113819648)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113821760)))];
+            tensor<fp16, [1, 56, 1024]> input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113823872))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116969664))), name = tensor<string, []>("encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_5_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116969856)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_46_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116978112))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120123904))), name = tensor<string, []>("encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_5_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120124096)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_47_cast_fp16")];
+            tensor<fp16, []> var_1520_to_fp16 = const()[name = tensor<string, []>("op_1520_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1521_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1520_to_fp16)[name = tensor<string, []>("op_1521_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1521_cast_fp16)[name = tensor<string, []>("input_297_cast_fp16")];
+            tensor<int32, [1]> key_11_axes_0 = const()[name = tensor<string, []>("key_11_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120126208)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120128320)))];
+            tensor<fp16, [1, 56, 1024]> key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = tensor<string, []>("key_11_cast_fp16")];
+            tensor<bool, []> input_299_interleave_0 = const()[name = tensor<string, []>("input_299_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_299_cast_fp16 = concat(axis = var_68, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = tensor<string, []>("input_299_cast_fp16")];
+            tensor<bool, []> var_1549_interleave_0 = const()[name = tensor<string, []>("op_1549_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_1549_cast_fp16 = concat(axis = var_68, interleave = var_1549_interleave_0, values = key_11_cast_fp16)[name = tensor<string, []>("op_1549_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120130432))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120916928))), name = tensor<string, []>("encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120917120)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_48_cast_fp16")];
+            tensor<int32, [4]> var_1554 = const()[name = tensor<string, []>("op_1554"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_31_cast_fp16 = reshape(shape = var_1554, x = linear_48_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120919232))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121705728))), name = tensor<string, []>("encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121705920)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_49_cast_fp16")];
+            tensor<int32, [4]> var_1559 = const()[name = tensor<string, []>("op_1559"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_21_cast_fp16 = reshape(shape = var_1559, x = linear_49_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121708032))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122494528))), name = tensor<string, []>("encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122494720)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_50_cast_fp16")];
+            tensor<int32, [4]> var_1564 = const()[name = tensor<string, []>("op_1564"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_11_cast_fp16 = reshape(shape = var_1564, x = linear_50_cast_fp16)[name = tensor<string, []>("v_11_cast_fp16")];
+            tensor<int32, [4]> value_19_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122496832)))];
+            tensor<fp16, [1, 56, 8, 128]> var_1577_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1577_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122498944)))];
+            tensor<fp16, [1, 56, 8, 128]> var_1579_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1579_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_11_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_137_transpose_x_0 = const()[name = tensor<string, []>("x_137_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_137_transpose_y_0 = const()[name = tensor<string, []>("x_137_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_1581_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1581_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122501056))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122700800))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1579_cast_fp16)[name = tensor<string, []>("transpose_317")];
+            tensor<fp16, [1, 8, 56, 195]> 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_1581_to_fp16_quantized)[name = tensor<string, []>("x_137_cast_fp16")];
+            tensor<int32, [8]> x_139_pad_0 = const()[name = tensor<string, []>("x_139_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_139_mode_0 = const()[name = tensor<string, []>("x_139_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_144_to_fp16 = const()[name = tensor<string, []>("const_144_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_139_cast_fp16")];
+            tensor<int32, [4]> var_1589 = const()[name = tensor<string, []>("op_1589"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_141_cast_fp16 = reshape(shape = var_1589, x = x_139_cast_fp16)[name = tensor<string, []>("x_141_cast_fp16")];
+            tensor<int32, [4]> var_1593_begin_0 = const()[name = tensor<string, []>("op_1593_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1593_end_0 = const()[name = tensor<string, []>("op_1593_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_1593_end_mask_0 = const()[name = tensor<string, []>("op_1593_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_1593_cast_fp16 = slice_by_index(begin = var_1593_begin_0, end = var_1593_end_0, end_mask = var_1593_end_mask_0, x = x_141_cast_fp16)[name = tensor<string, []>("op_1593_cast_fp16")];
+            tensor<int32, [4]> var_1594 = const()[name = tensor<string, []>("op_1594"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_21_cast_fp16 = reshape(shape = var_1594, x = var_1593_cast_fp16)[name = tensor<string, []>("matrix_bd_21_cast_fp16")];
+            tensor<bool, []> matrix_ac_11_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_11_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_11_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_11_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_106_perm_0 = const()[name = tensor<string, []>("transpose_106_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_107_perm_0 = const()[name = tensor<string, []>("transpose_107_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21_cast_fp16)[name = tensor<string, []>("transpose_315")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1577_cast_fp16)[name = tensor<string, []>("transpose_316")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_11_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_23_begin_0 = const()[name = tensor<string, []>("matrix_bd_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_23_end_0 = const()[name = tensor<string, []>("matrix_bd_23_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_23_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_23_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_1603_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = tensor<string, []>("op_1603_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_21_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_21_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_21_cast_fp16 = mul(x = var_1603_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_21_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_23_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_23_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_1609_cast_fp16 = softmax(axis = var_59, x = scores_23_cast_fp16)[name = tensor<string, []>("op_1609_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_301_cast_fp16 = select(a = var_44_to_fp16, b = var_1609_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_301_cast_fp16")];
+            tensor<bool, []> x_143_transpose_x_0 = const()[name = tensor<string, []>("x_143_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_143_transpose_y_0 = const()[name = tensor<string, []>("x_143_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_11_cast_fp16)[name = tensor<string, []>("transpose_314")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_143_cast_fp16")];
+            tensor<int32, [4]> var_1613_perm_0 = const()[name = tensor<string, []>("op_1613_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1614 = const()[name = tensor<string, []>("op_1614"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_1613_cast_fp16 = transpose(perm = var_1613_perm_0, x = x_143_cast_fp16)[name = tensor<string, []>("transpose_313")];
+            tensor<fp16, [1, 56, 1024]> input_303_cast_fp16 = reshape(shape = var_1614, x = var_1613_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122701312))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123487808))), name = tensor<string, []>("encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123488000)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_52_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = tensor<string, []>("input_307_cast_fp16")];
+            tensor<int32, [1]> x_147_axes_0 = const()[name = tensor<string, []>("x_147_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123490112)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123492224)))];
+            tensor<fp16, [1, 56, 1024]> x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = tensor<string, []>("x_147_cast_fp16")];
+            tensor<int32, [3]> input_309_perm_0 = const()[name = tensor<string, []>("input_309_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_311_pad_type_0 = const()[name = tensor<string, []>("input_311_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_311_strides_0 = const()[name = tensor<string, []>("input_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_311_pad_0 = const()[name = tensor<string, []>("input_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_311_dilations_0 = const()[name = tensor<string, []>("input_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_311_groups_0 = const()[name = tensor<string, []>("input_311_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123494336))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125591552))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = tensor<string, []>("transpose_312")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_311_cast_fp16")];
+            tensor<int32, []> x_149_split_num_splits_0 = const()[name = tensor<string, []>("x_149_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_149_split_axis_0 = const()[name = tensor<string, []>("x_149_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_149_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_149_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = tensor<string, []>("x_149_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_149_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_313_cast_fp16 = select(a = var_44_to_fp16, b = x_149_cast_fp16, cond = var_575)[name = tensor<string, []>("input_313_cast_fp16")];
+            tensor<bool, []> new_x_23_interleave_0 = const()[name = tensor<string, []>("new_x_23_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_23_cast_fp16 = concat(axis = var_59, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = tensor<string, []>("new_x_23_cast_fp16")];
+            tensor<int32, [3]> var_1653_begin_0 = const()[name = tensor<string, []>("op_1653_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_1653_end_0 = const()[name = tensor<string, []>("op_1653_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_1653_end_mask_0 = const()[name = tensor<string, []>("op_1653_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1653_cast_fp16 = slice_by_index(begin = var_1653_begin_0, end = var_1653_end_0, end_mask = var_1653_end_mask_0, x = new_x_23_cast_fp16)[name = tensor<string, []>("op_1653_cast_fp16")];
+            tensor<string, []> x_151_pad_type_0 = const()[name = tensor<string, []>("x_151_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_151_groups_0 = const()[name = tensor<string, []>("x_151_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_151_strides_0 = const()[name = tensor<string, []>("x_151_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_151_pad_0 = const()[name = tensor<string, []>("x_151_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_151_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125595712))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125604992))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_151_cast_fp16")];
+            tensor<int32, [3]> input_315_perm_0 = const()[name = tensor<string, []>("input_315_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_153_axes_0 = const()[name = tensor<string, []>("x_153_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125607104)))];
+            tensor<fp16, [1024]> encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125609216)))];
+            tensor<fp16, [1, 56, 1024]> input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = tensor<string, []>("transpose_311")];
+            tensor<fp16, [1, 56, 1024]> x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = tensor<string, []>("x_153_cast_fp16")];
+            tensor<int32, [3]> input_317_perm_0 = const()[name = tensor<string, []>("input_317_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = tensor<string, []>("transpose_310")];
+            tensor<fp16, [1, 1024, 56]> input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = tensor<string, []>("input_319_cast_fp16")];
+            tensor<string, []> x_155_pad_type_0 = const()[name = tensor<string, []>("x_155_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_155_strides_0 = const()[name = tensor<string, []>("x_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_155_pad_0 = const()[name = tensor<string, []>("x_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_155_dilations_0 = const()[name = tensor<string, []>("x_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_155_groups_0 = const()[name = tensor<string, []>("x_155_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125611328))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126659968))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_155_cast_fp16")];
+            tensor<int32, [3]> input_321_perm_0 = const()[name = tensor<string, []>("input_321_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = tensor<string, []>("transpose_309")];
+            tensor<fp16, [1, 56, 1024]> input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = tensor<string, []>("input_323_cast_fp16")];
+            tensor<int32, [1]> input_325_axes_0 = const()[name = tensor<string, []>("input_325_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126662080)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126664192)))];
+            tensor<fp16, [1, 56, 1024]> input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126666304))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129812096))), name = tensor<string, []>("encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_5_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129812288)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_53_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129820544))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132966336))), name = tensor<string, []>("encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_5_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132966528)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_54_cast_fp16")];
+            tensor<fp16, []> var_1696_to_fp16 = const()[name = tensor<string, []>("op_1696_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1697_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1696_to_fp16)[name = tensor<string, []>("op_1697_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1697_cast_fp16)[name = tensor<string, []>("input_335_cast_fp16")];
+            tensor<int32, [1]> input_337_axes_0 = const()[name = tensor<string, []>("input_337_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132968640)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132970752)))];
+            tensor<fp16, [1, 56, 1024]> input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = tensor<string, []>("input_337_cast_fp16")];
+            tensor<int32, [4]> cache_25_begin_0 = const()[name = tensor<string, []>("cache_25_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_25_end_0 = const()[name = tensor<string, []>("cache_25_end_0"), val = tensor<int32, [4]>([7, 1, 42, 1024])];
+            tensor<bool, [4]> cache_25_end_mask_0 = const()[name = tensor<string, []>("cache_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_25_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_25_cast_fp16")];
+            tensor<int32, [4]> cache_27_begin_0 = const()[name = tensor<string, []>("cache_27_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_27_end_0 = const()[name = tensor<string, []>("cache_27_end_0"), val = tensor<int32, [4]>([7, 1, 1024, 8])];
+            tensor<bool, [4]> cache_27_end_mask_0 = const()[name = tensor<string, []>("cache_27_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_27_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_27_cast_fp16")];
+            tensor<int32, [1]> input_339_axes_0 = const()[name = tensor<string, []>("input_339_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132972864)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132974976)))];
+            tensor<fp16, [1, 56, 1024]> input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132977088))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136122880))), name = tensor<string, []>("encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_6_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136123072)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_55_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136131328))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139277120))), name = tensor<string, []>("encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_6_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139277312)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_56_cast_fp16")];
+            tensor<fp16, []> var_1733_to_fp16 = const()[name = tensor<string, []>("op_1733_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1734_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1733_to_fp16)[name = tensor<string, []>("op_1734_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1734_cast_fp16)[name = tensor<string, []>("input_349_cast_fp16")];
+            tensor<int32, [1]> key_13_axes_0 = const()[name = tensor<string, []>("key_13_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139279424)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139281536)))];
+            tensor<fp16, [1, 56, 1024]> key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")];
+            tensor<bool, []> input_351_interleave_0 = const()[name = tensor<string, []>("input_351_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_351_cast_fp16 = concat(axis = var_68, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = tensor<string, []>("input_351_cast_fp16")];
+            tensor<bool, []> var_1762_interleave_0 = const()[name = tensor<string, []>("op_1762_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_1762_cast_fp16 = concat(axis = var_68, interleave = var_1762_interleave_0, values = key_13_cast_fp16)[name = tensor<string, []>("op_1762_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139283648))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140070144))), name = tensor<string, []>("encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140070336)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_57_cast_fp16")];
+            tensor<int32, [4]> var_1767 = const()[name = tensor<string, []>("op_1767"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_37_cast_fp16 = reshape(shape = var_1767, x = linear_57_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140072448))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140858944))), name = tensor<string, []>("encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140859136)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_58_cast_fp16")];
+            tensor<int32, [4]> var_1772 = const()[name = tensor<string, []>("op_1772"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_25_cast_fp16 = reshape(shape = var_1772, x = linear_58_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140861248))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141647744))), name = tensor<string, []>("encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141647936)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_59_cast_fp16")];
+            tensor<int32, [4]> var_1777 = const()[name = tensor<string, []>("op_1777"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_13_cast_fp16 = reshape(shape = var_1777, x = linear_59_cast_fp16)[name = tensor<string, []>("v_13_cast_fp16")];
+            tensor<int32, [4]> value_21_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141650048)))];
+            tensor<fp16, [1, 56, 8, 128]> var_1790_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1790_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141652160)))];
+            tensor<fp16, [1, 56, 8, 128]> var_1792_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1792_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_13_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_163_transpose_x_0 = const()[name = tensor<string, []>("x_163_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_163_transpose_y_0 = const()[name = tensor<string, []>("x_163_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_1794_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1794_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141654272))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141854016))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1792_cast_fp16)[name = tensor<string, []>("transpose_308")];
+            tensor<fp16, [1, 8, 56, 195]> 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_1794_to_fp16_quantized)[name = tensor<string, []>("x_163_cast_fp16")];
+            tensor<int32, [8]> x_165_pad_0 = const()[name = tensor<string, []>("x_165_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_165_mode_0 = const()[name = tensor<string, []>("x_165_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_157_to_fp16 = const()[name = tensor<string, []>("const_157_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_165_cast_fp16")];
+            tensor<int32, [4]> var_1802 = const()[name = tensor<string, []>("op_1802"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_167_cast_fp16 = reshape(shape = var_1802, x = x_165_cast_fp16)[name = tensor<string, []>("x_167_cast_fp16")];
+            tensor<int32, [4]> var_1806_begin_0 = const()[name = tensor<string, []>("op_1806_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1806_end_0 = const()[name = tensor<string, []>("op_1806_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_1806_end_mask_0 = const()[name = tensor<string, []>("op_1806_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_1806_cast_fp16 = slice_by_index(begin = var_1806_begin_0, end = var_1806_end_0, end_mask = var_1806_end_mask_0, x = x_167_cast_fp16)[name = tensor<string, []>("op_1806_cast_fp16")];
+            tensor<int32, [4]> var_1807 = const()[name = tensor<string, []>("op_1807"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_25_cast_fp16 = reshape(shape = var_1807, x = var_1806_cast_fp16)[name = tensor<string, []>("matrix_bd_25_cast_fp16")];
+            tensor<bool, []> matrix_ac_13_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_13_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_108_perm_0 = const()[name = tensor<string, []>("transpose_108_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_109_perm_0 = const()[name = tensor<string, []>("transpose_109_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25_cast_fp16)[name = tensor<string, []>("transpose_306")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1790_cast_fp16)[name = tensor<string, []>("transpose_307")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_13_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_27_begin_0 = const()[name = tensor<string, []>("matrix_bd_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_27_end_0 = const()[name = tensor<string, []>("matrix_bd_27_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_27_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_27_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_1816_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = tensor<string, []>("op_1816_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_25_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_25_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_25_cast_fp16 = mul(x = var_1816_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_25_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_27_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_27_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_1822_cast_fp16 = softmax(axis = var_59, x = scores_27_cast_fp16)[name = tensor<string, []>("op_1822_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_353_cast_fp16 = select(a = var_44_to_fp16, b = var_1822_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_353_cast_fp16")];
+            tensor<bool, []> x_169_transpose_x_0 = const()[name = tensor<string, []>("x_169_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_169_transpose_y_0 = const()[name = tensor<string, []>("x_169_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_13_cast_fp16)[name = tensor<string, []>("transpose_305")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_169_cast_fp16")];
+            tensor<int32, [4]> var_1826_perm_0 = const()[name = tensor<string, []>("op_1826_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1827 = const()[name = tensor<string, []>("op_1827"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_1826_cast_fp16 = transpose(perm = var_1826_perm_0, x = x_169_cast_fp16)[name = tensor<string, []>("transpose_304")];
+            tensor<fp16, [1, 56, 1024]> input_355_cast_fp16 = reshape(shape = var_1827, x = var_1826_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141854528))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142641024))), name = tensor<string, []>("encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142641216)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_61_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = tensor<string, []>("input_359_cast_fp16")];
+            tensor<int32, [1]> x_173_axes_0 = const()[name = tensor<string, []>("x_173_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142643328)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142645440)))];
+            tensor<fp16, [1, 56, 1024]> x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = tensor<string, []>("x_173_cast_fp16")];
+            tensor<int32, [3]> input_361_perm_0 = const()[name = tensor<string, []>("input_361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_363_pad_type_0 = const()[name = tensor<string, []>("input_363_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_363_strides_0 = const()[name = tensor<string, []>("input_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_363_pad_0 = const()[name = tensor<string, []>("input_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_363_dilations_0 = const()[name = tensor<string, []>("input_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_363_groups_0 = const()[name = tensor<string, []>("input_363_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142647552))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144744768))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = tensor<string, []>("transpose_303")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_363_cast_fp16")];
+            tensor<int32, []> x_175_split_num_splits_0 = const()[name = tensor<string, []>("x_175_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_175_split_axis_0 = const()[name = tensor<string, []>("x_175_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_175_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_175_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = tensor<string, []>("x_175_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_175_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_365_cast_fp16 = select(a = var_44_to_fp16, b = x_175_cast_fp16, cond = var_575)[name = tensor<string, []>("input_365_cast_fp16")];
+            tensor<bool, []> new_x_27_interleave_0 = const()[name = tensor<string, []>("new_x_27_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_27_cast_fp16 = concat(axis = var_59, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = tensor<string, []>("new_x_27_cast_fp16")];
+            tensor<int32, [3]> var_1866_begin_0 = const()[name = tensor<string, []>("op_1866_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_1866_end_0 = const()[name = tensor<string, []>("op_1866_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_1866_end_mask_0 = const()[name = tensor<string, []>("op_1866_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1866_cast_fp16 = slice_by_index(begin = var_1866_begin_0, end = var_1866_end_0, end_mask = var_1866_end_mask_0, x = new_x_27_cast_fp16)[name = tensor<string, []>("op_1866_cast_fp16")];
+            tensor<string, []> x_177_pad_type_0 = const()[name = tensor<string, []>("x_177_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_177_groups_0 = const()[name = tensor<string, []>("x_177_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_177_strides_0 = const()[name = tensor<string, []>("x_177_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_177_pad_0 = const()[name = tensor<string, []>("x_177_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_177_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144748928))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144758208))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_177_cast_fp16")];
+            tensor<int32, [3]> input_367_perm_0 = const()[name = tensor<string, []>("input_367_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_179_axes_0 = const()[name = tensor<string, []>("x_179_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144760320)))];
+            tensor<fp16, [1024]> encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144762432)))];
+            tensor<fp16, [1, 56, 1024]> input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = tensor<string, []>("transpose_302")];
+            tensor<fp16, [1, 56, 1024]> x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = tensor<string, []>("x_179_cast_fp16")];
+            tensor<int32, [3]> input_369_perm_0 = const()[name = tensor<string, []>("input_369_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = tensor<string, []>("transpose_301")];
+            tensor<fp16, [1, 1024, 56]> input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = tensor<string, []>("input_371_cast_fp16")];
+            tensor<string, []> x_181_pad_type_0 = const()[name = tensor<string, []>("x_181_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_181_strides_0 = const()[name = tensor<string, []>("x_181_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_181_pad_0 = const()[name = tensor<string, []>("x_181_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_181_dilations_0 = const()[name = tensor<string, []>("x_181_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_181_groups_0 = const()[name = tensor<string, []>("x_181_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144764544))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145813184))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_181_cast_fp16")];
+            tensor<int32, [3]> input_373_perm_0 = const()[name = tensor<string, []>("input_373_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = tensor<string, []>("transpose_300")];
+            tensor<fp16, [1, 56, 1024]> input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = tensor<string, []>("input_375_cast_fp16")];
+            tensor<int32, [1]> input_377_axes_0 = const()[name = tensor<string, []>("input_377_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145815296)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145817408)))];
+            tensor<fp16, [1, 56, 1024]> input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145819520))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148965312))), name = tensor<string, []>("encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_6_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148965504)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_62_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148973760))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152119552))), name = tensor<string, []>("encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_6_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152119744)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_63_cast_fp16")];
+            tensor<fp16, []> var_1909_to_fp16 = const()[name = tensor<string, []>("op_1909_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1910_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1909_to_fp16)[name = tensor<string, []>("op_1910_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1910_cast_fp16)[name = tensor<string, []>("input_387_cast_fp16")];
+            tensor<int32, [1]> input_389_axes_0 = const()[name = tensor<string, []>("input_389_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152121856)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152123968)))];
+            tensor<fp16, [1, 56, 1024]> input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = tensor<string, []>("input_389_cast_fp16")];
+            tensor<int32, [4]> cache_29_begin_0 = const()[name = tensor<string, []>("cache_29_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_29_end_0 = const()[name = tensor<string, []>("cache_29_end_0"), val = tensor<int32, [4]>([8, 1, 42, 1024])];
+            tensor<bool, [4]> cache_29_end_mask_0 = const()[name = tensor<string, []>("cache_29_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_29_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_29_cast_fp16")];
+            tensor<int32, [4]> cache_31_begin_0 = const()[name = tensor<string, []>("cache_31_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_31_end_0 = const()[name = tensor<string, []>("cache_31_end_0"), val = tensor<int32, [4]>([8, 1, 1024, 8])];
+            tensor<bool, [4]> cache_31_end_mask_0 = const()[name = tensor<string, []>("cache_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_31_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_31_cast_fp16")];
+            tensor<int32, [1]> input_391_axes_0 = const()[name = tensor<string, []>("input_391_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152126080)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152128192)))];
+            tensor<fp16, [1, 56, 1024]> input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152130304))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155276096))), name = tensor<string, []>("encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_7_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155276288)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_64_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155284544))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158430336))), name = tensor<string, []>("encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_7_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158430528)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_65_cast_fp16")];
+            tensor<fp16, []> var_1946_to_fp16 = const()[name = tensor<string, []>("op_1946_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1947_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1946_to_fp16)[name = tensor<string, []>("op_1947_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1947_cast_fp16)[name = tensor<string, []>("input_401_cast_fp16")];
+            tensor<int32, [1]> key_15_axes_0 = const()[name = tensor<string, []>("key_15_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158432640)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158434752)))];
+            tensor<fp16, [1, 56, 1024]> key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = tensor<string, []>("key_15_cast_fp16")];
+            tensor<bool, []> input_403_interleave_0 = const()[name = tensor<string, []>("input_403_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_403_cast_fp16 = concat(axis = var_68, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = tensor<string, []>("input_403_cast_fp16")];
+            tensor<bool, []> var_1975_interleave_0 = const()[name = tensor<string, []>("op_1975_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_1975_cast_fp16 = concat(axis = var_68, interleave = var_1975_interleave_0, values = key_15_cast_fp16)[name = tensor<string, []>("op_1975_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158436864))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159223360))), name = tensor<string, []>("encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159223552)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_66_cast_fp16")];
+            tensor<int32, [4]> var_1980 = const()[name = tensor<string, []>("op_1980"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_43_cast_fp16 = reshape(shape = var_1980, x = linear_66_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159225664))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160012160))), name = tensor<string, []>("encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160012352)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_67_cast_fp16")];
+            tensor<int32, [4]> var_1985 = const()[name = tensor<string, []>("op_1985"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_29_cast_fp16 = reshape(shape = var_1985, x = linear_67_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160014464))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160800960))), name = tensor<string, []>("encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160801152)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_68_cast_fp16")];
+            tensor<int32, [4]> var_1990 = const()[name = tensor<string, []>("op_1990"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_15_cast_fp16 = reshape(shape = var_1990, x = linear_68_cast_fp16)[name = tensor<string, []>("v_15_cast_fp16")];
+            tensor<int32, [4]> value_23_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160803264)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2003_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2003_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160805376)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2005_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2005_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_15_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_189_transpose_x_0 = const()[name = tensor<string, []>("x_189_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_189_transpose_y_0 = const()[name = tensor<string, []>("x_189_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_2007_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2007_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160807488))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161007232))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_2005_cast_fp16)[name = tensor<string, []>("transpose_299")];
+            tensor<fp16, [1, 8, 56, 195]> 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_2007_to_fp16_quantized)[name = tensor<string, []>("x_189_cast_fp16")];
+            tensor<int32, [8]> x_191_pad_0 = const()[name = tensor<string, []>("x_191_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_191_mode_0 = const()[name = tensor<string, []>("x_191_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_170_to_fp16 = const()[name = tensor<string, []>("const_170_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_191_cast_fp16")];
+            tensor<int32, [4]> var_2015 = const()[name = tensor<string, []>("op_2015"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_193_cast_fp16 = reshape(shape = var_2015, x = x_191_cast_fp16)[name = tensor<string, []>("x_193_cast_fp16")];
+            tensor<int32, [4]> var_2019_begin_0 = const()[name = tensor<string, []>("op_2019_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2019_end_0 = const()[name = tensor<string, []>("op_2019_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_2019_end_mask_0 = const()[name = tensor<string, []>("op_2019_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_2019_cast_fp16 = slice_by_index(begin = var_2019_begin_0, end = var_2019_end_0, end_mask = var_2019_end_mask_0, x = x_193_cast_fp16)[name = tensor<string, []>("op_2019_cast_fp16")];
+            tensor<int32, [4]> var_2020 = const()[name = tensor<string, []>("op_2020"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_29_cast_fp16 = reshape(shape = var_2020, x = var_2019_cast_fp16)[name = tensor<string, []>("matrix_bd_29_cast_fp16")];
+            tensor<bool, []> matrix_ac_15_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_15_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_15_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_15_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_110_perm_0 = const()[name = tensor<string, []>("transpose_110_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_111_perm_0 = const()[name = tensor<string, []>("transpose_111_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29_cast_fp16)[name = tensor<string, []>("transpose_297")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_110 = transpose(perm = transpose_110_perm_0, x = var_2003_cast_fp16)[name = tensor<string, []>("transpose_298")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_15_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_31_begin_0 = const()[name = tensor<string, []>("matrix_bd_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_31_end_0 = const()[name = tensor<string, []>("matrix_bd_31_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_31_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_31_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2029_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = tensor<string, []>("op_2029_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_29_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_29_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_29_cast_fp16 = mul(x = var_2029_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_29_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_31_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_31_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2035_cast_fp16 = softmax(axis = var_59, x = scores_31_cast_fp16)[name = tensor<string, []>("op_2035_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_405_cast_fp16 = select(a = var_44_to_fp16, b = var_2035_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_405_cast_fp16")];
+            tensor<bool, []> x_195_transpose_x_0 = const()[name = tensor<string, []>("x_195_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_195_transpose_y_0 = const()[name = tensor<string, []>("x_195_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_15_cast_fp16)[name = tensor<string, []>("transpose_296")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_195_cast_fp16")];
+            tensor<int32, [4]> var_2039_perm_0 = const()[name = tensor<string, []>("op_2039_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2040 = const()[name = tensor<string, []>("op_2040"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_2039_cast_fp16 = transpose(perm = var_2039_perm_0, x = x_195_cast_fp16)[name = tensor<string, []>("transpose_295")];
+            tensor<fp16, [1, 56, 1024]> input_407_cast_fp16 = reshape(shape = var_2040, x = var_2039_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161007744))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161794240))), name = tensor<string, []>("encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161794432)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_70_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = tensor<string, []>("input_411_cast_fp16")];
+            tensor<int32, [1]> x_199_axes_0 = const()[name = tensor<string, []>("x_199_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161796544)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161798656)))];
+            tensor<fp16, [1, 56, 1024]> x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = tensor<string, []>("x_199_cast_fp16")];
+            tensor<int32, [3]> input_413_perm_0 = const()[name = tensor<string, []>("input_413_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_415_pad_type_0 = const()[name = tensor<string, []>("input_415_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_415_strides_0 = const()[name = tensor<string, []>("input_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_415_pad_0 = const()[name = tensor<string, []>("input_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_415_dilations_0 = const()[name = tensor<string, []>("input_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_415_groups_0 = const()[name = tensor<string, []>("input_415_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161800768))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163897984))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = tensor<string, []>("transpose_294")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_415_cast_fp16")];
+            tensor<int32, []> x_201_split_num_splits_0 = const()[name = tensor<string, []>("x_201_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_201_split_axis_0 = const()[name = tensor<string, []>("x_201_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_201_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_201_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = tensor<string, []>("x_201_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_201_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_417_cast_fp16 = select(a = var_44_to_fp16, b = x_201_cast_fp16, cond = var_575)[name = tensor<string, []>("input_417_cast_fp16")];
+            tensor<bool, []> new_x_31_interleave_0 = const()[name = tensor<string, []>("new_x_31_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_31_cast_fp16 = concat(axis = var_59, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = tensor<string, []>("new_x_31_cast_fp16")];
+            tensor<int32, [3]> var_2079_begin_0 = const()[name = tensor<string, []>("op_2079_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_2079_end_0 = const()[name = tensor<string, []>("op_2079_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_2079_end_mask_0 = const()[name = tensor<string, []>("op_2079_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2079_cast_fp16 = slice_by_index(begin = var_2079_begin_0, end = var_2079_end_0, end_mask = var_2079_end_mask_0, x = new_x_31_cast_fp16)[name = tensor<string, []>("op_2079_cast_fp16")];
+            tensor<string, []> x_203_pad_type_0 = const()[name = tensor<string, []>("x_203_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_203_groups_0 = const()[name = tensor<string, []>("x_203_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_203_strides_0 = const()[name = tensor<string, []>("x_203_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_203_pad_0 = const()[name = tensor<string, []>("x_203_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_203_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163902144))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163911424))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_203_cast_fp16")];
+            tensor<int32, [3]> input_419_perm_0 = const()[name = tensor<string, []>("input_419_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_205_axes_0 = const()[name = tensor<string, []>("x_205_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163913536)))];
+            tensor<fp16, [1024]> encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163915648)))];
+            tensor<fp16, [1, 56, 1024]> input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = tensor<string, []>("transpose_293")];
+            tensor<fp16, [1, 56, 1024]> x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = tensor<string, []>("x_205_cast_fp16")];
+            tensor<int32, [3]> input_421_perm_0 = const()[name = tensor<string, []>("input_421_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = tensor<string, []>("transpose_292")];
+            tensor<fp16, [1, 1024, 56]> input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor<string, []>("input_423_cast_fp16")];
+            tensor<string, []> x_207_pad_type_0 = const()[name = tensor<string, []>("x_207_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_207_strides_0 = const()[name = tensor<string, []>("x_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_207_pad_0 = const()[name = tensor<string, []>("x_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_207_dilations_0 = const()[name = tensor<string, []>("x_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_207_groups_0 = const()[name = tensor<string, []>("x_207_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163917760))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164966400))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_207_cast_fp16")];
+            tensor<int32, [3]> input_425_perm_0 = const()[name = tensor<string, []>("input_425_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = tensor<string, []>("transpose_291")];
+            tensor<fp16, [1, 56, 1024]> input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = tensor<string, []>("input_427_cast_fp16")];
+            tensor<int32, [1]> input_429_axes_0 = const()[name = tensor<string, []>("input_429_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164968512)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164970624)))];
+            tensor<fp16, [1, 56, 1024]> input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164972736))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168118528))), name = tensor<string, []>("encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_7_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168118720)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_71_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168126976))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171272768))), name = tensor<string, []>("encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_7_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171272960)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_72_cast_fp16")];
+            tensor<fp16, []> var_2122_to_fp16 = const()[name = tensor<string, []>("op_2122_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2123_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2122_to_fp16)[name = tensor<string, []>("op_2123_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_2123_cast_fp16)[name = tensor<string, []>("input_439_cast_fp16")];
+            tensor<int32, [1]> input_441_axes_0 = const()[name = tensor<string, []>("input_441_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171275072)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171277184)))];
+            tensor<fp16, [1, 56, 1024]> input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = tensor<string, []>("input_441_cast_fp16")];
+            tensor<int32, [4]> cache_33_begin_0 = const()[name = tensor<string, []>("cache_33_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_33_end_0 = const()[name = tensor<string, []>("cache_33_end_0"), val = tensor<int32, [4]>([9, 1, 42, 1024])];
+            tensor<bool, [4]> cache_33_end_mask_0 = const()[name = tensor<string, []>("cache_33_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_33_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_33_cast_fp16")];
+            tensor<int32, [4]> cache_35_begin_0 = const()[name = tensor<string, []>("cache_35_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_35_end_0 = const()[name = tensor<string, []>("cache_35_end_0"), val = tensor<int32, [4]>([9, 1, 1024, 8])];
+            tensor<bool, [4]> cache_35_end_mask_0 = const()[name = tensor<string, []>("cache_35_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_35_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_35_cast_fp16")];
+            tensor<int32, [1]> input_443_axes_0 = const()[name = tensor<string, []>("input_443_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171279296)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171281408)))];
+            tensor<fp16, [1, 56, 1024]> input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171283520))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174429312))), name = tensor<string, []>("encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_8_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174429504)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_73_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174437760))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177583552))), name = tensor<string, []>("encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_8_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177583744)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_74_cast_fp16")];
+            tensor<fp16, []> var_2159_to_fp16 = const()[name = tensor<string, []>("op_2159_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2160_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2159_to_fp16)[name = tensor<string, []>("op_2160_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_2160_cast_fp16)[name = tensor<string, []>("input_453_cast_fp16")];
+            tensor<int32, [1]> key_17_axes_0 = const()[name = tensor<string, []>("key_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177585856)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177587968)))];
+            tensor<fp16, [1, 56, 1024]> key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = tensor<string, []>("key_17_cast_fp16")];
+            tensor<bool, []> input_455_interleave_0 = const()[name = tensor<string, []>("input_455_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_455_cast_fp16 = concat(axis = var_68, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = tensor<string, []>("input_455_cast_fp16")];
+            tensor<bool, []> var_2188_interleave_0 = const()[name = tensor<string, []>("op_2188_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_2188_cast_fp16 = concat(axis = var_68, interleave = var_2188_interleave_0, values = key_17_cast_fp16)[name = tensor<string, []>("op_2188_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177590080))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178376576))), name = tensor<string, []>("encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178376768)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_75_cast_fp16")];
+            tensor<int32, [4]> var_2193 = const()[name = tensor<string, []>("op_2193"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_49_cast_fp16 = reshape(shape = var_2193, x = linear_75_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178378880))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179165376))), name = tensor<string, []>("encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179165568)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_76_cast_fp16")];
+            tensor<int32, [4]> var_2198 = const()[name = tensor<string, []>("op_2198"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_33_cast_fp16 = reshape(shape = var_2198, x = linear_76_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179167680))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179954176))), name = tensor<string, []>("encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179954368)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_77_cast_fp16")];
+            tensor<int32, [4]> var_2203 = const()[name = tensor<string, []>("op_2203"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_17_cast_fp16 = reshape(shape = var_2203, x = linear_77_cast_fp16)[name = tensor<string, []>("v_17_cast_fp16")];
+            tensor<int32, [4]> value_25_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179956480)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2216_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2216_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179958592)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2218_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2218_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_17_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_215_transpose_x_0 = const()[name = tensor<string, []>("x_215_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_215_transpose_y_0 = const()[name = tensor<string, []>("x_215_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_2220_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2220_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179960704))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180160448))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2218_cast_fp16)[name = tensor<string, []>("transpose_290")];
+            tensor<fp16, [1, 8, 56, 195]> 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_2220_to_fp16_quantized)[name = tensor<string, []>("x_215_cast_fp16")];
+            tensor<int32, [8]> x_217_pad_0 = const()[name = tensor<string, []>("x_217_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_217_mode_0 = const()[name = tensor<string, []>("x_217_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_183_to_fp16 = const()[name = tensor<string, []>("const_183_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_217_cast_fp16")];
+            tensor<int32, [4]> var_2228 = const()[name = tensor<string, []>("op_2228"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_219_cast_fp16 = reshape(shape = var_2228, x = x_217_cast_fp16)[name = tensor<string, []>("x_219_cast_fp16")];
+            tensor<int32, [4]> var_2232_begin_0 = const()[name = tensor<string, []>("op_2232_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2232_end_0 = const()[name = tensor<string, []>("op_2232_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_2232_end_mask_0 = const()[name = tensor<string, []>("op_2232_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_2232_cast_fp16 = slice_by_index(begin = var_2232_begin_0, end = var_2232_end_0, end_mask = var_2232_end_mask_0, x = x_219_cast_fp16)[name = tensor<string, []>("op_2232_cast_fp16")];
+            tensor<int32, [4]> var_2233 = const()[name = tensor<string, []>("op_2233"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_33_cast_fp16 = reshape(shape = var_2233, x = var_2232_cast_fp16)[name = tensor<string, []>("matrix_bd_33_cast_fp16")];
+            tensor<bool, []> matrix_ac_17_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_17_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_17_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_17_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_112_perm_0 = const()[name = tensor<string, []>("transpose_112_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_113_perm_0 = const()[name = tensor<string, []>("transpose_113_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33_cast_fp16)[name = tensor<string, []>("transpose_288")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_112 = transpose(perm = transpose_112_perm_0, x = var_2216_cast_fp16)[name = tensor<string, []>("transpose_289")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_17_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_35_begin_0 = const()[name = tensor<string, []>("matrix_bd_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_35_end_0 = const()[name = tensor<string, []>("matrix_bd_35_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_35_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_35_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2242_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = tensor<string, []>("op_2242_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_33_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_33_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_33_cast_fp16 = mul(x = var_2242_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_33_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_35_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_35_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2248_cast_fp16 = softmax(axis = var_59, x = scores_35_cast_fp16)[name = tensor<string, []>("op_2248_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_457_cast_fp16 = select(a = var_44_to_fp16, b = var_2248_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_457_cast_fp16")];
+            tensor<bool, []> x_221_transpose_x_0 = const()[name = tensor<string, []>("x_221_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_221_transpose_y_0 = const()[name = tensor<string, []>("x_221_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_17_cast_fp16)[name = tensor<string, []>("transpose_287")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_221_cast_fp16")];
+            tensor<int32, [4]> var_2252_perm_0 = const()[name = tensor<string, []>("op_2252_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2253 = const()[name = tensor<string, []>("op_2253"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_2252_cast_fp16 = transpose(perm = var_2252_perm_0, x = x_221_cast_fp16)[name = tensor<string, []>("transpose_286")];
+            tensor<fp16, [1, 56, 1024]> input_459_cast_fp16 = reshape(shape = var_2253, x = var_2252_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180160960))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180947456))), name = tensor<string, []>("encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180947648)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_79_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = tensor<string, []>("input_463_cast_fp16")];
+            tensor<int32, [1]> x_225_axes_0 = const()[name = tensor<string, []>("x_225_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180949760)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180951872)))];
+            tensor<fp16, [1, 56, 1024]> x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = tensor<string, []>("x_225_cast_fp16")];
+            tensor<int32, [3]> input_465_perm_0 = const()[name = tensor<string, []>("input_465_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_467_pad_type_0 = const()[name = tensor<string, []>("input_467_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_467_strides_0 = const()[name = tensor<string, []>("input_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_467_pad_0 = const()[name = tensor<string, []>("input_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_467_dilations_0 = const()[name = tensor<string, []>("input_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_467_groups_0 = const()[name = tensor<string, []>("input_467_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180953984))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183051200))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = tensor<string, []>("transpose_285")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_467_cast_fp16")];
+            tensor<int32, []> x_227_split_num_splits_0 = const()[name = tensor<string, []>("x_227_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_227_split_axis_0 = const()[name = tensor<string, []>("x_227_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_227_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_227_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = tensor<string, []>("x_227_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_227_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_469_cast_fp16 = select(a = var_44_to_fp16, b = x_227_cast_fp16, cond = var_575)[name = tensor<string, []>("input_469_cast_fp16")];
+            tensor<bool, []> new_x_35_interleave_0 = const()[name = tensor<string, []>("new_x_35_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_35_cast_fp16 = concat(axis = var_59, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = tensor<string, []>("new_x_35_cast_fp16")];
+            tensor<int32, [3]> var_2292_begin_0 = const()[name = tensor<string, []>("op_2292_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_2292_end_0 = const()[name = tensor<string, []>("op_2292_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_2292_end_mask_0 = const()[name = tensor<string, []>("op_2292_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2292_cast_fp16 = slice_by_index(begin = var_2292_begin_0, end = var_2292_end_0, end_mask = var_2292_end_mask_0, x = new_x_35_cast_fp16)[name = tensor<string, []>("op_2292_cast_fp16")];
+            tensor<string, []> x_229_pad_type_0 = const()[name = tensor<string, []>("x_229_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_229_groups_0 = const()[name = tensor<string, []>("x_229_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_229_strides_0 = const()[name = tensor<string, []>("x_229_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_229_pad_0 = const()[name = tensor<string, []>("x_229_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_229_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183055360))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183064640))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_229_cast_fp16")];
+            tensor<int32, [3]> input_471_perm_0 = const()[name = tensor<string, []>("input_471_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_231_axes_0 = const()[name = tensor<string, []>("x_231_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183066752)))];
+            tensor<fp16, [1024]> encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183068864)))];
+            tensor<fp16, [1, 56, 1024]> input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = tensor<string, []>("transpose_284")];
+            tensor<fp16, [1, 56, 1024]> x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = tensor<string, []>("x_231_cast_fp16")];
+            tensor<int32, [3]> input_473_perm_0 = const()[name = tensor<string, []>("input_473_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = tensor<string, []>("transpose_283")];
+            tensor<fp16, [1, 1024, 56]> input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = tensor<string, []>("input_475_cast_fp16")];
+            tensor<string, []> x_233_pad_type_0 = const()[name = tensor<string, []>("x_233_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_233_strides_0 = const()[name = tensor<string, []>("x_233_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_233_pad_0 = const()[name = tensor<string, []>("x_233_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_233_dilations_0 = const()[name = tensor<string, []>("x_233_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_233_groups_0 = const()[name = tensor<string, []>("x_233_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183070976))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184119616))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_233_cast_fp16")];
+            tensor<int32, [3]> input_477_perm_0 = const()[name = tensor<string, []>("input_477_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = tensor<string, []>("transpose_282")];
+            tensor<fp16, [1, 56, 1024]> input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = tensor<string, []>("input_479_cast_fp16")];
+            tensor<int32, [1]> input_481_axes_0 = const()[name = tensor<string, []>("input_481_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184121728)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184123840)))];
+            tensor<fp16, [1, 56, 1024]> input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184125952))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187271744))), name = tensor<string, []>("encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_8_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187271936)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_80_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187280192))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190425984))), name = tensor<string, []>("encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_8_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190426176)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_81_cast_fp16")];
+            tensor<fp16, []> var_2335_to_fp16 = const()[name = tensor<string, []>("op_2335_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2336_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2335_to_fp16)[name = tensor<string, []>("op_2336_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2336_cast_fp16)[name = tensor<string, []>("input_491_cast_fp16")];
+            tensor<int32, [1]> input_493_axes_0 = const()[name = tensor<string, []>("input_493_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190428288)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190430400)))];
+            tensor<fp16, [1, 56, 1024]> input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = tensor<string, []>("input_493_cast_fp16")];
+            tensor<int32, [4]> cache_37_begin_0 = const()[name = tensor<string, []>("cache_37_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_37_end_0 = const()[name = tensor<string, []>("cache_37_end_0"), val = tensor<int32, [4]>([10, 1, 42, 1024])];
+            tensor<bool, [4]> cache_37_end_mask_0 = const()[name = tensor<string, []>("cache_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_37_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_37_cast_fp16")];
+            tensor<int32, [4]> cache_39_begin_0 = const()[name = tensor<string, []>("cache_39_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_39_end_0 = const()[name = tensor<string, []>("cache_39_end_0"), val = tensor<int32, [4]>([10, 1, 1024, 8])];
+            tensor<bool, [4]> cache_39_end_mask_0 = const()[name = tensor<string, []>("cache_39_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_39_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_39_cast_fp16")];
+            tensor<int32, [1]> input_495_axes_0 = const()[name = tensor<string, []>("input_495_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190432512)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190434624)))];
+            tensor<fp16, [1, 56, 1024]> input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190436736))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193582528))), name = tensor<string, []>("encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_9_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193582720)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_82_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193590976))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196736768))), name = tensor<string, []>("encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_9_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196736960)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_83_cast_fp16")];
+            tensor<fp16, []> var_2372_to_fp16 = const()[name = tensor<string, []>("op_2372_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2373_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2372_to_fp16)[name = tensor<string, []>("op_2373_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2373_cast_fp16)[name = tensor<string, []>("input_505_cast_fp16")];
+            tensor<int32, [1]> key_19_axes_0 = const()[name = tensor<string, []>("key_19_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196739072)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196741184)))];
+            tensor<fp16, [1, 56, 1024]> key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = tensor<string, []>("key_19_cast_fp16")];
+            tensor<bool, []> input_507_interleave_0 = const()[name = tensor<string, []>("input_507_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_507_cast_fp16 = concat(axis = var_68, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = tensor<string, []>("input_507_cast_fp16")];
+            tensor<bool, []> var_2401_interleave_0 = const()[name = tensor<string, []>("op_2401_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_2401_cast_fp16 = concat(axis = var_68, interleave = var_2401_interleave_0, values = key_19_cast_fp16)[name = tensor<string, []>("op_2401_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196743296))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197529792))), name = tensor<string, []>("encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197529984)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_84_cast_fp16")];
+            tensor<int32, [4]> var_2406 = const()[name = tensor<string, []>("op_2406"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_55_cast_fp16 = reshape(shape = var_2406, x = linear_84_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197532096))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198318592))), name = tensor<string, []>("encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198318784)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_85_cast_fp16")];
+            tensor<int32, [4]> var_2411 = const()[name = tensor<string, []>("op_2411"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_37_cast_fp16 = reshape(shape = var_2411, x = linear_85_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198320896))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199107392))), name = tensor<string, []>("encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199107584)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_86_cast_fp16")];
+            tensor<int32, [4]> var_2416 = const()[name = tensor<string, []>("op_2416"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_19_cast_fp16 = reshape(shape = var_2416, x = linear_86_cast_fp16)[name = tensor<string, []>("v_19_cast_fp16")];
+            tensor<int32, [4]> value_27_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199109696)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2429_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2429_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199111808)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2431_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2431_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_19_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_241_transpose_x_0 = const()[name = tensor<string, []>("x_241_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_241_transpose_y_0 = const()[name = tensor<string, []>("x_241_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_2433_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2433_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199113920))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199313664))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2431_cast_fp16)[name = tensor<string, []>("transpose_281")];
+            tensor<fp16, [1, 8, 56, 195]> 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_2433_to_fp16_quantized)[name = tensor<string, []>("x_241_cast_fp16")];
+            tensor<int32, [8]> x_243_pad_0 = const()[name = tensor<string, []>("x_243_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_243_mode_0 = const()[name = tensor<string, []>("x_243_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_196_to_fp16 = const()[name = tensor<string, []>("const_196_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_243_cast_fp16")];
+            tensor<int32, [4]> var_2441 = const()[name = tensor<string, []>("op_2441"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_245_cast_fp16 = reshape(shape = var_2441, x = x_243_cast_fp16)[name = tensor<string, []>("x_245_cast_fp16")];
+            tensor<int32, [4]> var_2445_begin_0 = const()[name = tensor<string, []>("op_2445_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2445_end_0 = const()[name = tensor<string, []>("op_2445_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_2445_end_mask_0 = const()[name = tensor<string, []>("op_2445_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_2445_cast_fp16 = slice_by_index(begin = var_2445_begin_0, end = var_2445_end_0, end_mask = var_2445_end_mask_0, x = x_245_cast_fp16)[name = tensor<string, []>("op_2445_cast_fp16")];
+            tensor<int32, [4]> var_2446 = const()[name = tensor<string, []>("op_2446"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_37_cast_fp16 = reshape(shape = var_2446, x = var_2445_cast_fp16)[name = tensor<string, []>("matrix_bd_37_cast_fp16")];
+            tensor<bool, []> matrix_ac_19_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_19_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_19_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_19_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_114_perm_0 = const()[name = tensor<string, []>("transpose_114_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_115_perm_0 = const()[name = tensor<string, []>("transpose_115_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37_cast_fp16)[name = tensor<string, []>("transpose_279")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_114 = transpose(perm = transpose_114_perm_0, x = var_2429_cast_fp16)[name = tensor<string, []>("transpose_280")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_19_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_39_begin_0 = const()[name = tensor<string, []>("matrix_bd_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_39_end_0 = const()[name = tensor<string, []>("matrix_bd_39_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_39_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_39_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2455_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = tensor<string, []>("op_2455_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_37_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_37_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_37_cast_fp16 = mul(x = var_2455_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_37_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_39_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_39_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2461_cast_fp16 = softmax(axis = var_59, x = scores_39_cast_fp16)[name = tensor<string, []>("op_2461_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_509_cast_fp16 = select(a = var_44_to_fp16, b = var_2461_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_509_cast_fp16")];
+            tensor<bool, []> x_247_transpose_x_0 = const()[name = tensor<string, []>("x_247_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_247_transpose_y_0 = const()[name = tensor<string, []>("x_247_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_19_cast_fp16)[name = tensor<string, []>("transpose_278")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_247_cast_fp16")];
+            tensor<int32, [4]> var_2465_perm_0 = const()[name = tensor<string, []>("op_2465_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2466 = const()[name = tensor<string, []>("op_2466"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_2465_cast_fp16 = transpose(perm = var_2465_perm_0, x = x_247_cast_fp16)[name = tensor<string, []>("transpose_277")];
+            tensor<fp16, [1, 56, 1024]> input_511_cast_fp16 = reshape(shape = var_2466, x = var_2465_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199314176))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200100672))), name = tensor<string, []>("encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200100864)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_88_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = tensor<string, []>("input_515_cast_fp16")];
+            tensor<int32, [1]> x_251_axes_0 = const()[name = tensor<string, []>("x_251_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200102976)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200105088)))];
+            tensor<fp16, [1, 56, 1024]> x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = tensor<string, []>("x_251_cast_fp16")];
+            tensor<int32, [3]> input_517_perm_0 = const()[name = tensor<string, []>("input_517_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_519_pad_type_0 = const()[name = tensor<string, []>("input_519_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_519_strides_0 = const()[name = tensor<string, []>("input_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_519_pad_0 = const()[name = tensor<string, []>("input_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_519_dilations_0 = const()[name = tensor<string, []>("input_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_519_groups_0 = const()[name = tensor<string, []>("input_519_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200107200))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202204416))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = tensor<string, []>("transpose_276")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_519_cast_fp16")];
+            tensor<int32, []> x_253_split_num_splits_0 = const()[name = tensor<string, []>("x_253_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_253_split_axis_0 = const()[name = tensor<string, []>("x_253_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_253_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_253_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = tensor<string, []>("x_253_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_253_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_521_cast_fp16 = select(a = var_44_to_fp16, b = x_253_cast_fp16, cond = var_575)[name = tensor<string, []>("input_521_cast_fp16")];
+            tensor<bool, []> new_x_39_interleave_0 = const()[name = tensor<string, []>("new_x_39_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_39_cast_fp16 = concat(axis = var_59, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = tensor<string, []>("new_x_39_cast_fp16")];
+            tensor<int32, [3]> var_2505_begin_0 = const()[name = tensor<string, []>("op_2505_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_2505_end_0 = const()[name = tensor<string, []>("op_2505_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_2505_end_mask_0 = const()[name = tensor<string, []>("op_2505_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2505_cast_fp16 = slice_by_index(begin = var_2505_begin_0, end = var_2505_end_0, end_mask = var_2505_end_mask_0, x = new_x_39_cast_fp16)[name = tensor<string, []>("op_2505_cast_fp16")];
+            tensor<string, []> x_255_pad_type_0 = const()[name = tensor<string, []>("x_255_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_255_groups_0 = const()[name = tensor<string, []>("x_255_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_255_strides_0 = const()[name = tensor<string, []>("x_255_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_255_pad_0 = const()[name = tensor<string, []>("x_255_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_255_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202208576))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202217856))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_255_cast_fp16")];
+            tensor<int32, [3]> input_523_perm_0 = const()[name = tensor<string, []>("input_523_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_257_axes_0 = const()[name = tensor<string, []>("x_257_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202219968)))];
+            tensor<fp16, [1024]> encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202222080)))];
+            tensor<fp16, [1, 56, 1024]> input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = tensor<string, []>("transpose_275")];
+            tensor<fp16, [1, 56, 1024]> x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = tensor<string, []>("x_257_cast_fp16")];
+            tensor<int32, [3]> input_525_perm_0 = const()[name = tensor<string, []>("input_525_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = tensor<string, []>("transpose_274")];
+            tensor<fp16, [1, 1024, 56]> input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = tensor<string, []>("input_527_cast_fp16")];
+            tensor<string, []> x_259_pad_type_0 = const()[name = tensor<string, []>("x_259_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_259_strides_0 = const()[name = tensor<string, []>("x_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_259_pad_0 = const()[name = tensor<string, []>("x_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_259_dilations_0 = const()[name = tensor<string, []>("x_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_259_groups_0 = const()[name = tensor<string, []>("x_259_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202224192))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203272832))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_259_cast_fp16")];
+            tensor<int32, [3]> input_529_perm_0 = const()[name = tensor<string, []>("input_529_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = tensor<string, []>("transpose_273")];
+            tensor<fp16, [1, 56, 1024]> input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = tensor<string, []>("input_531_cast_fp16")];
+            tensor<int32, [1]> input_533_axes_0 = const()[name = tensor<string, []>("input_533_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203274944)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203277056)))];
+            tensor<fp16, [1, 56, 1024]> input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203279168))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206424960))), name = tensor<string, []>("encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_9_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206425152)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_89_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206433408))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209579200))), name = tensor<string, []>("encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_9_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209579392)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_90_cast_fp16")];
+            tensor<fp16, []> var_2548_to_fp16 = const()[name = tensor<string, []>("op_2548_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2549_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2548_to_fp16)[name = tensor<string, []>("op_2549_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2549_cast_fp16)[name = tensor<string, []>("input_543_cast_fp16")];
+            tensor<int32, [1]> input_545_axes_0 = const()[name = tensor<string, []>("input_545_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209581504)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209583616)))];
+            tensor<fp16, [1, 56, 1024]> input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = tensor<string, []>("input_545_cast_fp16")];
+            tensor<int32, [4]> cache_41_begin_0 = const()[name = tensor<string, []>("cache_41_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_41_end_0 = const()[name = tensor<string, []>("cache_41_end_0"), val = tensor<int32, [4]>([11, 1, 42, 1024])];
+            tensor<bool, [4]> cache_41_end_mask_0 = const()[name = tensor<string, []>("cache_41_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_41_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_41_cast_fp16")];
+            tensor<int32, [4]> cache_43_begin_0 = const()[name = tensor<string, []>("cache_43_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_43_end_0 = const()[name = tensor<string, []>("cache_43_end_0"), val = tensor<int32, [4]>([11, 1, 1024, 8])];
+            tensor<bool, [4]> cache_43_end_mask_0 = const()[name = tensor<string, []>("cache_43_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_43_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_43_cast_fp16")];
+            tensor<int32, [1]> input_547_axes_0 = const()[name = tensor<string, []>("input_547_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209585728)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209587840)))];
+            tensor<fp16, [1, 56, 1024]> input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209589952))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212735744))), name = tensor<string, []>("encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_10_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212735936)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_91_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212744192))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215889984))), name = tensor<string, []>("encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_10_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215890176)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_92_cast_fp16")];
+            tensor<fp16, []> var_2585_to_fp16 = const()[name = tensor<string, []>("op_2585_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2586_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2585_to_fp16)[name = tensor<string, []>("op_2586_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2586_cast_fp16)[name = tensor<string, []>("input_557_cast_fp16")];
+            tensor<int32, [1]> key_21_axes_0 = const()[name = tensor<string, []>("key_21_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215892288)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215894400)))];
+            tensor<fp16, [1, 56, 1024]> key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = tensor<string, []>("key_21_cast_fp16")];
+            tensor<bool, []> input_559_interleave_0 = const()[name = tensor<string, []>("input_559_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_559_cast_fp16 = concat(axis = var_68, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = tensor<string, []>("input_559_cast_fp16")];
+            tensor<bool, []> var_2614_interleave_0 = const()[name = tensor<string, []>("op_2614_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_2614_cast_fp16 = concat(axis = var_68, interleave = var_2614_interleave_0, values = key_21_cast_fp16)[name = tensor<string, []>("op_2614_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215896512))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216683008))), name = tensor<string, []>("encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216683200)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_93_cast_fp16")];
+            tensor<int32, [4]> var_2619 = const()[name = tensor<string, []>("op_2619"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_61_cast_fp16 = reshape(shape = var_2619, x = linear_93_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216685312))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217471808))), name = tensor<string, []>("encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217472000)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_94_cast_fp16")];
+            tensor<int32, [4]> var_2624 = const()[name = tensor<string, []>("op_2624"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_41_cast_fp16 = reshape(shape = var_2624, x = linear_94_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217474112))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218260608))), name = tensor<string, []>("encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218260800)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_95_cast_fp16")];
+            tensor<int32, [4]> var_2629 = const()[name = tensor<string, []>("op_2629"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_21_cast_fp16 = reshape(shape = var_2629, x = linear_95_cast_fp16)[name = tensor<string, []>("v_21_cast_fp16")];
+            tensor<int32, [4]> value_29_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218262912)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2642_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2642_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218265024)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2644_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2644_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_21_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_267_transpose_x_0 = const()[name = tensor<string, []>("x_267_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_267_transpose_y_0 = const()[name = tensor<string, []>("x_267_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_2646_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2646_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218267136))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218466880))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2644_cast_fp16)[name = tensor<string, []>("transpose_272")];
+            tensor<fp16, [1, 8, 56, 195]> 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_2646_to_fp16_quantized)[name = tensor<string, []>("x_267_cast_fp16")];
+            tensor<int32, [8]> x_269_pad_0 = const()[name = tensor<string, []>("x_269_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_269_mode_0 = const()[name = tensor<string, []>("x_269_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_209_to_fp16 = const()[name = tensor<string, []>("const_209_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_269_cast_fp16")];
+            tensor<int32, [4]> var_2654 = const()[name = tensor<string, []>("op_2654"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_271_cast_fp16 = reshape(shape = var_2654, x = x_269_cast_fp16)[name = tensor<string, []>("x_271_cast_fp16")];
+            tensor<int32, [4]> var_2658_begin_0 = const()[name = tensor<string, []>("op_2658_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2658_end_0 = const()[name = tensor<string, []>("op_2658_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_2658_end_mask_0 = const()[name = tensor<string, []>("op_2658_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_2658_cast_fp16 = slice_by_index(begin = var_2658_begin_0, end = var_2658_end_0, end_mask = var_2658_end_mask_0, x = x_271_cast_fp16)[name = tensor<string, []>("op_2658_cast_fp16")];
+            tensor<int32, [4]> var_2659 = const()[name = tensor<string, []>("op_2659"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_41_cast_fp16 = reshape(shape = var_2659, x = var_2658_cast_fp16)[name = tensor<string, []>("matrix_bd_41_cast_fp16")];
+            tensor<bool, []> matrix_ac_21_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_21_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_21_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_21_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_116_perm_0 = const()[name = tensor<string, []>("transpose_116_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_117_perm_0 = const()[name = tensor<string, []>("transpose_117_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41_cast_fp16)[name = tensor<string, []>("transpose_270")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_116 = transpose(perm = transpose_116_perm_0, x = var_2642_cast_fp16)[name = tensor<string, []>("transpose_271")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_21_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_43_begin_0 = const()[name = tensor<string, []>("matrix_bd_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_43_end_0 = const()[name = tensor<string, []>("matrix_bd_43_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_43_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_43_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2668_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = tensor<string, []>("op_2668_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_41_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_41_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_41_cast_fp16 = mul(x = var_2668_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_41_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_43_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_43_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2674_cast_fp16 = softmax(axis = var_59, x = scores_43_cast_fp16)[name = tensor<string, []>("op_2674_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_561_cast_fp16 = select(a = var_44_to_fp16, b = var_2674_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_561_cast_fp16")];
+            tensor<bool, []> x_273_transpose_x_0 = const()[name = tensor<string, []>("x_273_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_273_transpose_y_0 = const()[name = tensor<string, []>("x_273_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_21_cast_fp16)[name = tensor<string, []>("transpose_269")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_273_cast_fp16")];
+            tensor<int32, [4]> var_2678_perm_0 = const()[name = tensor<string, []>("op_2678_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2679 = const()[name = tensor<string, []>("op_2679"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_2678_cast_fp16 = transpose(perm = var_2678_perm_0, x = x_273_cast_fp16)[name = tensor<string, []>("transpose_268")];
+            tensor<fp16, [1, 56, 1024]> input_563_cast_fp16 = reshape(shape = var_2679, x = var_2678_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218467392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219253888))), name = tensor<string, []>("encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219254080)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_97_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = tensor<string, []>("input_567_cast_fp16")];
+            tensor<int32, [1]> x_277_axes_0 = const()[name = tensor<string, []>("x_277_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219256192)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219258304)))];
+            tensor<fp16, [1, 56, 1024]> x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = tensor<string, []>("x_277_cast_fp16")];
+            tensor<int32, [3]> input_569_perm_0 = const()[name = tensor<string, []>("input_569_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_571_pad_type_0 = const()[name = tensor<string, []>("input_571_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_571_strides_0 = const()[name = tensor<string, []>("input_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_571_pad_0 = const()[name = tensor<string, []>("input_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_571_dilations_0 = const()[name = tensor<string, []>("input_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_571_groups_0 = const()[name = tensor<string, []>("input_571_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219260416))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221357632))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = tensor<string, []>("transpose_267")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_571_cast_fp16")];
+            tensor<int32, []> x_279_split_num_splits_0 = const()[name = tensor<string, []>("x_279_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_279_split_axis_0 = const()[name = tensor<string, []>("x_279_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_279_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_279_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = tensor<string, []>("x_279_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_279_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_573_cast_fp16 = select(a = var_44_to_fp16, b = x_279_cast_fp16, cond = var_575)[name = tensor<string, []>("input_573_cast_fp16")];
+            tensor<bool, []> new_x_43_interleave_0 = const()[name = tensor<string, []>("new_x_43_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_43_cast_fp16 = concat(axis = var_59, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = tensor<string, []>("new_x_43_cast_fp16")];
+            tensor<int32, [3]> var_2718_begin_0 = const()[name = tensor<string, []>("op_2718_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_2718_end_0 = const()[name = tensor<string, []>("op_2718_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_2718_end_mask_0 = const()[name = tensor<string, []>("op_2718_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2718_cast_fp16 = slice_by_index(begin = var_2718_begin_0, end = var_2718_end_0, end_mask = var_2718_end_mask_0, x = new_x_43_cast_fp16)[name = tensor<string, []>("op_2718_cast_fp16")];
+            tensor<string, []> x_281_pad_type_0 = const()[name = tensor<string, []>("x_281_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_281_groups_0 = const()[name = tensor<string, []>("x_281_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_281_strides_0 = const()[name = tensor<string, []>("x_281_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_281_pad_0 = const()[name = tensor<string, []>("x_281_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_281_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221361792))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221371072))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_281_cast_fp16")];
+            tensor<int32, [3]> input_575_perm_0 = const()[name = tensor<string, []>("input_575_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_283_axes_0 = const()[name = tensor<string, []>("x_283_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221373184)))];
+            tensor<fp16, [1024]> encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221375296)))];
+            tensor<fp16, [1, 56, 1024]> input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = tensor<string, []>("transpose_266")];
+            tensor<fp16, [1, 56, 1024]> x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = tensor<string, []>("x_283_cast_fp16")];
+            tensor<int32, [3]> input_577_perm_0 = const()[name = tensor<string, []>("input_577_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = tensor<string, []>("transpose_265")];
+            tensor<fp16, [1, 1024, 56]> input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = tensor<string, []>("input_579_cast_fp16")];
+            tensor<string, []> x_285_pad_type_0 = const()[name = tensor<string, []>("x_285_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_285_strides_0 = const()[name = tensor<string, []>("x_285_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_285_pad_0 = const()[name = tensor<string, []>("x_285_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_285_dilations_0 = const()[name = tensor<string, []>("x_285_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_285_groups_0 = const()[name = tensor<string, []>("x_285_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221377408))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222426048))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_285_cast_fp16")];
+            tensor<int32, [3]> input_581_perm_0 = const()[name = tensor<string, []>("input_581_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = tensor<string, []>("transpose_264")];
+            tensor<fp16, [1, 56, 1024]> input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = tensor<string, []>("input_583_cast_fp16")];
+            tensor<int32, [1]> input_585_axes_0 = const()[name = tensor<string, []>("input_585_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222428160)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222430272)))];
+            tensor<fp16, [1, 56, 1024]> input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222432384))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(225578176))), name = tensor<string, []>("encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_10_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(225578368)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_98_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(225586624))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228732416))), name = tensor<string, []>("encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_10_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228732608)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_99_cast_fp16")];
+            tensor<fp16, []> var_2761_to_fp16 = const()[name = tensor<string, []>("op_2761_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2762_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2761_to_fp16)[name = tensor<string, []>("op_2762_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2762_cast_fp16)[name = tensor<string, []>("input_595_cast_fp16")];
+            tensor<int32, [1]> input_597_axes_0 = const()[name = tensor<string, []>("input_597_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228734720)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228736832)))];
+            tensor<fp16, [1, 56, 1024]> input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = tensor<string, []>("input_597_cast_fp16")];
+            tensor<int32, [4]> cache_45_begin_0 = const()[name = tensor<string, []>("cache_45_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_45_end_0 = const()[name = tensor<string, []>("cache_45_end_0"), val = tensor<int32, [4]>([12, 1, 42, 1024])];
+            tensor<bool, [4]> cache_45_end_mask_0 = const()[name = tensor<string, []>("cache_45_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_45_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_45_cast_fp16")];
+            tensor<int32, [4]> cache_47_begin_0 = const()[name = tensor<string, []>("cache_47_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_47_end_0 = const()[name = tensor<string, []>("cache_47_end_0"), val = tensor<int32, [4]>([12, 1, 1024, 8])];
+            tensor<bool, [4]> cache_47_end_mask_0 = const()[name = tensor<string, []>("cache_47_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_47_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_47_cast_fp16")];
+            tensor<int32, [1]> input_599_axes_0 = const()[name = tensor<string, []>("input_599_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228738944)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228741056)))];
+            tensor<fp16, [1, 56, 1024]> input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228743168))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231888960))), name = tensor<string, []>("encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_11_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231889152)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_100_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231897408))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235043200))), name = tensor<string, []>("encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_11_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235043392)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_101_cast_fp16")];
+            tensor<fp16, []> var_2798_to_fp16 = const()[name = tensor<string, []>("op_2798_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2799_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2798_to_fp16)[name = tensor<string, []>("op_2799_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2799_cast_fp16)[name = tensor<string, []>("input_609_cast_fp16")];
+            tensor<int32, [1]> key_23_axes_0 = const()[name = tensor<string, []>("key_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235045504)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235047616)))];
+            tensor<fp16, [1, 56, 1024]> key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = tensor<string, []>("key_23_cast_fp16")];
+            tensor<bool, []> input_611_interleave_0 = const()[name = tensor<string, []>("input_611_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_611_cast_fp16 = concat(axis = var_68, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = tensor<string, []>("input_611_cast_fp16")];
+            tensor<bool, []> var_2827_interleave_0 = const()[name = tensor<string, []>("op_2827_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_2827_cast_fp16 = concat(axis = var_68, interleave = var_2827_interleave_0, values = key_23_cast_fp16)[name = tensor<string, []>("op_2827_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235049728))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235836224))), name = tensor<string, []>("encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235836416)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_102_cast_fp16")];
+            tensor<int32, [4]> var_2832 = const()[name = tensor<string, []>("op_2832"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_67_cast_fp16 = reshape(shape = var_2832, x = linear_102_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235838528))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236625024))), name = tensor<string, []>("encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236625216)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_103_cast_fp16")];
+            tensor<int32, [4]> var_2837 = const()[name = tensor<string, []>("op_2837"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_45_cast_fp16 = reshape(shape = var_2837, x = linear_103_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236627328))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237413824))), name = tensor<string, []>("encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237414016)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_104_cast_fp16")];
+            tensor<int32, [4]> var_2842 = const()[name = tensor<string, []>("op_2842"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_23_cast_fp16 = reshape(shape = var_2842, x = linear_104_cast_fp16)[name = tensor<string, []>("v_23_cast_fp16")];
+            tensor<int32, [4]> value_31_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237416128)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2855_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2855_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237418240)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2857_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2857_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_23_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_293_transpose_x_0 = const()[name = tensor<string, []>("x_293_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_293_transpose_y_0 = const()[name = tensor<string, []>("x_293_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_2859_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2859_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237420352))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237620096))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2857_cast_fp16)[name = tensor<string, []>("transpose_263")];
+            tensor<fp16, [1, 8, 56, 195]> 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_2859_to_fp16_quantized)[name = tensor<string, []>("x_293_cast_fp16")];
+            tensor<int32, [8]> x_295_pad_0 = const()[name = tensor<string, []>("x_295_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_295_mode_0 = const()[name = tensor<string, []>("x_295_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_222_to_fp16 = const()[name = tensor<string, []>("const_222_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_295_cast_fp16")];
+            tensor<int32, [4]> var_2867 = const()[name = tensor<string, []>("op_2867"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_297_cast_fp16 = reshape(shape = var_2867, x = x_295_cast_fp16)[name = tensor<string, []>("x_297_cast_fp16")];
+            tensor<int32, [4]> var_2871_begin_0 = const()[name = tensor<string, []>("op_2871_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2871_end_0 = const()[name = tensor<string, []>("op_2871_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_2871_end_mask_0 = const()[name = tensor<string, []>("op_2871_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_2871_cast_fp16 = slice_by_index(begin = var_2871_begin_0, end = var_2871_end_0, end_mask = var_2871_end_mask_0, x = x_297_cast_fp16)[name = tensor<string, []>("op_2871_cast_fp16")];
+            tensor<int32, [4]> var_2872 = const()[name = tensor<string, []>("op_2872"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_45_cast_fp16 = reshape(shape = var_2872, x = var_2871_cast_fp16)[name = tensor<string, []>("matrix_bd_45_cast_fp16")];
+            tensor<bool, []> matrix_ac_23_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_23_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_23_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_23_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_118_perm_0 = const()[name = tensor<string, []>("transpose_118_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_119_perm_0 = const()[name = tensor<string, []>("transpose_119_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45_cast_fp16)[name = tensor<string, []>("transpose_261")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_118 = transpose(perm = transpose_118_perm_0, x = var_2855_cast_fp16)[name = tensor<string, []>("transpose_262")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_23_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_47_begin_0 = const()[name = tensor<string, []>("matrix_bd_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_47_end_0 = const()[name = tensor<string, []>("matrix_bd_47_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_47_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_47_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2881_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = tensor<string, []>("op_2881_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_45_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_45_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_45_cast_fp16 = mul(x = var_2881_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_45_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_47_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_47_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2887_cast_fp16 = softmax(axis = var_59, x = scores_47_cast_fp16)[name = tensor<string, []>("op_2887_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_613_cast_fp16 = select(a = var_44_to_fp16, b = var_2887_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_613_cast_fp16")];
+            tensor<bool, []> x_299_transpose_x_0 = const()[name = tensor<string, []>("x_299_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_299_transpose_y_0 = const()[name = tensor<string, []>("x_299_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_23_cast_fp16)[name = tensor<string, []>("transpose_260")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_299_cast_fp16")];
+            tensor<int32, [4]> var_2891_perm_0 = const()[name = tensor<string, []>("op_2891_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2892 = const()[name = tensor<string, []>("op_2892"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_2891_cast_fp16 = transpose(perm = var_2891_perm_0, x = x_299_cast_fp16)[name = tensor<string, []>("transpose_259")];
+            tensor<fp16, [1, 56, 1024]> input_615_cast_fp16 = reshape(shape = var_2892, x = var_2891_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237620608))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238407104))), name = tensor<string, []>("encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238407296)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_106_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = tensor<string, []>("input_619_cast_fp16")];
+            tensor<int32, [1]> x_303_axes_0 = const()[name = tensor<string, []>("x_303_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238409408)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238411520)))];
+            tensor<fp16, [1, 56, 1024]> x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = tensor<string, []>("x_303_cast_fp16")];
+            tensor<int32, [3]> input_621_perm_0 = const()[name = tensor<string, []>("input_621_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_623_pad_type_0 = const()[name = tensor<string, []>("input_623_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_623_strides_0 = const()[name = tensor<string, []>("input_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_623_pad_0 = const()[name = tensor<string, []>("input_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_623_dilations_0 = const()[name = tensor<string, []>("input_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_623_groups_0 = const()[name = tensor<string, []>("input_623_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238413632))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240510848))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = tensor<string, []>("transpose_258")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_623_cast_fp16")];
+            tensor<int32, []> x_305_split_num_splits_0 = const()[name = tensor<string, []>("x_305_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_305_split_axis_0 = const()[name = tensor<string, []>("x_305_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_305_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_305_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = tensor<string, []>("x_305_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_305_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_625_cast_fp16 = select(a = var_44_to_fp16, b = x_305_cast_fp16, cond = var_575)[name = tensor<string, []>("input_625_cast_fp16")];
+            tensor<bool, []> new_x_47_interleave_0 = const()[name = tensor<string, []>("new_x_47_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_47_cast_fp16 = concat(axis = var_59, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = tensor<string, []>("new_x_47_cast_fp16")];
+            tensor<int32, [3]> var_2931_begin_0 = const()[name = tensor<string, []>("op_2931_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_2931_end_0 = const()[name = tensor<string, []>("op_2931_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_2931_end_mask_0 = const()[name = tensor<string, []>("op_2931_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2931_cast_fp16 = slice_by_index(begin = var_2931_begin_0, end = var_2931_end_0, end_mask = var_2931_end_mask_0, x = new_x_47_cast_fp16)[name = tensor<string, []>("op_2931_cast_fp16")];
+            tensor<string, []> x_307_pad_type_0 = const()[name = tensor<string, []>("x_307_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_307_groups_0 = const()[name = tensor<string, []>("x_307_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_307_strides_0 = const()[name = tensor<string, []>("x_307_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_307_pad_0 = const()[name = tensor<string, []>("x_307_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_307_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240515008))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240524288))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_307_cast_fp16")];
+            tensor<int32, [3]> input_627_perm_0 = const()[name = tensor<string, []>("input_627_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_309_axes_0 = const()[name = tensor<string, []>("x_309_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240526400)))];
+            tensor<fp16, [1024]> encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240528512)))];
+            tensor<fp16, [1, 56, 1024]> input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = tensor<string, []>("transpose_257")];
+            tensor<fp16, [1, 56, 1024]> x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = tensor<string, []>("x_309_cast_fp16")];
+            tensor<int32, [3]> input_629_perm_0 = const()[name = tensor<string, []>("input_629_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = tensor<string, []>("transpose_256")];
+            tensor<fp16, [1, 1024, 56]> input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = tensor<string, []>("input_631_cast_fp16")];
+            tensor<string, []> x_311_pad_type_0 = const()[name = tensor<string, []>("x_311_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_311_strides_0 = const()[name = tensor<string, []>("x_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_311_pad_0 = const()[name = tensor<string, []>("x_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_311_dilations_0 = const()[name = tensor<string, []>("x_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_311_groups_0 = const()[name = tensor<string, []>("x_311_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240530624))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241579264))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_311_cast_fp16")];
+            tensor<int32, [3]> input_633_perm_0 = const()[name = tensor<string, []>("input_633_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = tensor<string, []>("transpose_255")];
+            tensor<fp16, [1, 56, 1024]> input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = tensor<string, []>("input_635_cast_fp16")];
+            tensor<int32, [1]> input_637_axes_0 = const()[name = tensor<string, []>("input_637_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241581376)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241583488)))];
+            tensor<fp16, [1, 56, 1024]> input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241585600))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244731392))), name = tensor<string, []>("encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_11_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244731584)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_107_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244739840))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247885632))), name = tensor<string, []>("encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_11_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247885824)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_108_cast_fp16")];
+            tensor<fp16, []> var_2974_to_fp16 = const()[name = tensor<string, []>("op_2974_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2975_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2974_to_fp16)[name = tensor<string, []>("op_2975_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2975_cast_fp16)[name = tensor<string, []>("input_647_cast_fp16")];
+            tensor<int32, [1]> input_649_axes_0 = const()[name = tensor<string, []>("input_649_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247887936)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247890048)))];
+            tensor<fp16, [1, 56, 1024]> input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = tensor<string, []>("input_649_cast_fp16")];
+            tensor<int32, [4]> cache_49_begin_0 = const()[name = tensor<string, []>("cache_49_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_49_end_0 = const()[name = tensor<string, []>("cache_49_end_0"), val = tensor<int32, [4]>([13, 1, 42, 1024])];
+            tensor<bool, [4]> cache_49_end_mask_0 = const()[name = tensor<string, []>("cache_49_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_49_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_49_cast_fp16")];
+            tensor<int32, [4]> cache_51_begin_0 = const()[name = tensor<string, []>("cache_51_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_51_end_0 = const()[name = tensor<string, []>("cache_51_end_0"), val = tensor<int32, [4]>([13, 1, 1024, 8])];
+            tensor<bool, [4]> cache_51_end_mask_0 = const()[name = tensor<string, []>("cache_51_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_51_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_51_cast_fp16")];
+            tensor<int32, [1]> input_651_axes_0 = const()[name = tensor<string, []>("input_651_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247892160)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247894272)))];
+            tensor<fp16, [1, 56, 1024]> input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247896384))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251042176))), name = tensor<string, []>("encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_12_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251042368)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_109_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251050624))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254196416))), name = tensor<string, []>("encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_12_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254196608)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_110_cast_fp16")];
+            tensor<fp16, []> var_3011_to_fp16 = const()[name = tensor<string, []>("op_3011_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3012_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_3011_to_fp16)[name = tensor<string, []>("op_3012_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_3012_cast_fp16)[name = tensor<string, []>("input_661_cast_fp16")];
+            tensor<int32, [1]> key_25_axes_0 = const()[name = tensor<string, []>("key_25_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254198720)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254200832)))];
+            tensor<fp16, [1, 56, 1024]> key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = tensor<string, []>("key_25_cast_fp16")];
+            tensor<bool, []> input_663_interleave_0 = const()[name = tensor<string, []>("input_663_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_663_cast_fp16 = concat(axis = var_68, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = tensor<string, []>("input_663_cast_fp16")];
+            tensor<bool, []> var_3040_interleave_0 = const()[name = tensor<string, []>("op_3040_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_3040_cast_fp16 = concat(axis = var_68, interleave = var_3040_interleave_0, values = key_25_cast_fp16)[name = tensor<string, []>("op_3040_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254202944))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254989440))), name = tensor<string, []>("encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254989632)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_111_cast_fp16")];
+            tensor<int32, [4]> var_3045 = const()[name = tensor<string, []>("op_3045"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_73_cast_fp16 = reshape(shape = var_3045, x = linear_111_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254991744))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255778240))), name = tensor<string, []>("encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255778432)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_112_cast_fp16")];
+            tensor<int32, [4]> var_3050 = const()[name = tensor<string, []>("op_3050"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_49_cast_fp16 = reshape(shape = var_3050, x = linear_112_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255780544))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256567040))), name = tensor<string, []>("encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256567232)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_113_cast_fp16")];
+            tensor<int32, [4]> var_3055 = const()[name = tensor<string, []>("op_3055"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_25_cast_fp16 = reshape(shape = var_3055, x = linear_113_cast_fp16)[name = tensor<string, []>("v_25_cast_fp16")];
+            tensor<int32, [4]> value_33_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256569344)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3068_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3068_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256571456)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3070_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3070_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_25_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_319_transpose_x_0 = const()[name = tensor<string, []>("x_319_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_319_transpose_y_0 = const()[name = tensor<string, []>("x_319_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_3072_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3072_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256573568))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256773312))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_3070_cast_fp16)[name = tensor<string, []>("transpose_254")];
+            tensor<fp16, [1, 8, 56, 195]> 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_3072_to_fp16_quantized)[name = tensor<string, []>("x_319_cast_fp16")];
+            tensor<int32, [8]> x_321_pad_0 = const()[name = tensor<string, []>("x_321_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_321_mode_0 = const()[name = tensor<string, []>("x_321_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_235_to_fp16 = const()[name = tensor<string, []>("const_235_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_321_cast_fp16")];
+            tensor<int32, [4]> var_3080 = const()[name = tensor<string, []>("op_3080"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_323_cast_fp16 = reshape(shape = var_3080, x = x_321_cast_fp16)[name = tensor<string, []>("x_323_cast_fp16")];
+            tensor<int32, [4]> var_3084_begin_0 = const()[name = tensor<string, []>("op_3084_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3084_end_0 = const()[name = tensor<string, []>("op_3084_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_3084_end_mask_0 = const()[name = tensor<string, []>("op_3084_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_3084_cast_fp16 = slice_by_index(begin = var_3084_begin_0, end = var_3084_end_0, end_mask = var_3084_end_mask_0, x = x_323_cast_fp16)[name = tensor<string, []>("op_3084_cast_fp16")];
+            tensor<int32, [4]> var_3085 = const()[name = tensor<string, []>("op_3085"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_49_cast_fp16 = reshape(shape = var_3085, x = var_3084_cast_fp16)[name = tensor<string, []>("matrix_bd_49_cast_fp16")];
+            tensor<bool, []> matrix_ac_25_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_25_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_25_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_25_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_120_perm_0 = const()[name = tensor<string, []>("transpose_120_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_121_perm_0 = const()[name = tensor<string, []>("transpose_121_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49_cast_fp16)[name = tensor<string, []>("transpose_252")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_120 = transpose(perm = transpose_120_perm_0, x = var_3068_cast_fp16)[name = tensor<string, []>("transpose_253")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_25_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_51_begin_0 = const()[name = tensor<string, []>("matrix_bd_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_51_end_0 = const()[name = tensor<string, []>("matrix_bd_51_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_51_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_51_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3094_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = tensor<string, []>("op_3094_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_49_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_49_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_49_cast_fp16 = mul(x = var_3094_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_49_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_51_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_51_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3100_cast_fp16 = softmax(axis = var_59, x = scores_51_cast_fp16)[name = tensor<string, []>("op_3100_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_665_cast_fp16 = select(a = var_44_to_fp16, b = var_3100_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_665_cast_fp16")];
+            tensor<bool, []> x_325_transpose_x_0 = const()[name = tensor<string, []>("x_325_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_325_transpose_y_0 = const()[name = tensor<string, []>("x_325_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_25_cast_fp16)[name = tensor<string, []>("transpose_251")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_325_cast_fp16")];
+            tensor<int32, [4]> var_3104_perm_0 = const()[name = tensor<string, []>("op_3104_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3105 = const()[name = tensor<string, []>("op_3105"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_3104_cast_fp16 = transpose(perm = var_3104_perm_0, x = x_325_cast_fp16)[name = tensor<string, []>("transpose_250")];
+            tensor<fp16, [1, 56, 1024]> input_667_cast_fp16 = reshape(shape = var_3105, x = var_3104_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256773824))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257560320))), name = tensor<string, []>("encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257560512)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_115_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = tensor<string, []>("input_671_cast_fp16")];
+            tensor<int32, [1]> x_329_axes_0 = const()[name = tensor<string, []>("x_329_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257562624)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257564736)))];
+            tensor<fp16, [1, 56, 1024]> x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = tensor<string, []>("x_329_cast_fp16")];
+            tensor<int32, [3]> input_673_perm_0 = const()[name = tensor<string, []>("input_673_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_675_pad_type_0 = const()[name = tensor<string, []>("input_675_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_675_strides_0 = const()[name = tensor<string, []>("input_675_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_675_pad_0 = const()[name = tensor<string, []>("input_675_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_675_dilations_0 = const()[name = tensor<string, []>("input_675_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_675_groups_0 = const()[name = tensor<string, []>("input_675_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257566848))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259664064))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = tensor<string, []>("transpose_249")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_675_cast_fp16")];
+            tensor<int32, []> x_331_split_num_splits_0 = const()[name = tensor<string, []>("x_331_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_331_split_axis_0 = const()[name = tensor<string, []>("x_331_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_331_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_331_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = tensor<string, []>("x_331_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_331_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_677_cast_fp16 = select(a = var_44_to_fp16, b = x_331_cast_fp16, cond = var_575)[name = tensor<string, []>("input_677_cast_fp16")];
+            tensor<bool, []> new_x_51_interleave_0 = const()[name = tensor<string, []>("new_x_51_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_51_cast_fp16 = concat(axis = var_59, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = tensor<string, []>("new_x_51_cast_fp16")];
+            tensor<int32, [3]> var_3144_begin_0 = const()[name = tensor<string, []>("op_3144_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_3144_end_0 = const()[name = tensor<string, []>("op_3144_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_3144_end_mask_0 = const()[name = tensor<string, []>("op_3144_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3144_cast_fp16 = slice_by_index(begin = var_3144_begin_0, end = var_3144_end_0, end_mask = var_3144_end_mask_0, x = new_x_51_cast_fp16)[name = tensor<string, []>("op_3144_cast_fp16")];
+            tensor<string, []> x_333_pad_type_0 = const()[name = tensor<string, []>("x_333_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_333_groups_0 = const()[name = tensor<string, []>("x_333_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_333_strides_0 = const()[name = tensor<string, []>("x_333_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_333_pad_0 = const()[name = tensor<string, []>("x_333_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_333_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259668224))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259677504))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_333_cast_fp16")];
+            tensor<int32, [3]> input_679_perm_0 = const()[name = tensor<string, []>("input_679_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_335_axes_0 = const()[name = tensor<string, []>("x_335_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259679616)))];
+            tensor<fp16, [1024]> encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259681728)))];
+            tensor<fp16, [1, 56, 1024]> input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = tensor<string, []>("transpose_248")];
+            tensor<fp16, [1, 56, 1024]> x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = tensor<string, []>("x_335_cast_fp16")];
+            tensor<int32, [3]> input_681_perm_0 = const()[name = tensor<string, []>("input_681_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = tensor<string, []>("transpose_247")];
+            tensor<fp16, [1, 1024, 56]> input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = tensor<string, []>("input_683_cast_fp16")];
+            tensor<string, []> x_337_pad_type_0 = const()[name = tensor<string, []>("x_337_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_337_strides_0 = const()[name = tensor<string, []>("x_337_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_337_pad_0 = const()[name = tensor<string, []>("x_337_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_337_dilations_0 = const()[name = tensor<string, []>("x_337_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_337_groups_0 = const()[name = tensor<string, []>("x_337_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259683840))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260732480))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_337_cast_fp16")];
+            tensor<int32, [3]> input_685_perm_0 = const()[name = tensor<string, []>("input_685_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = tensor<string, []>("transpose_246")];
+            tensor<fp16, [1, 56, 1024]> input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = tensor<string, []>("input_687_cast_fp16")];
+            tensor<int32, [1]> input_689_axes_0 = const()[name = tensor<string, []>("input_689_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260734592)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260736704)))];
+            tensor<fp16, [1, 56, 1024]> input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260738816))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263884608))), name = tensor<string, []>("encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_12_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263884800)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_116_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263893056))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267038848))), name = tensor<string, []>("encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_12_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267039040)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_117_cast_fp16")];
+            tensor<fp16, []> var_3187_to_fp16 = const()[name = tensor<string, []>("op_3187_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3188_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3187_to_fp16)[name = tensor<string, []>("op_3188_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_3188_cast_fp16)[name = tensor<string, []>("input_699_cast_fp16")];
+            tensor<int32, [1]> input_701_axes_0 = const()[name = tensor<string, []>("input_701_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267041152)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267043264)))];
+            tensor<fp16, [1, 56, 1024]> input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = tensor<string, []>("input_701_cast_fp16")];
+            tensor<int32, [4]> cache_53_begin_0 = const()[name = tensor<string, []>("cache_53_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_53_end_0 = const()[name = tensor<string, []>("cache_53_end_0"), val = tensor<int32, [4]>([14, 1, 42, 1024])];
+            tensor<bool, [4]> cache_53_end_mask_0 = const()[name = tensor<string, []>("cache_53_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_53_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_53_cast_fp16")];
+            tensor<int32, [4]> cache_55_begin_0 = const()[name = tensor<string, []>("cache_55_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_55_end_0 = const()[name = tensor<string, []>("cache_55_end_0"), val = tensor<int32, [4]>([14, 1, 1024, 8])];
+            tensor<bool, [4]> cache_55_end_mask_0 = const()[name = tensor<string, []>("cache_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_55_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_55_cast_fp16")];
+            tensor<int32, [1]> input_703_axes_0 = const()[name = tensor<string, []>("input_703_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267045376)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267047488)))];
+            tensor<fp16, [1, 56, 1024]> input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267049600))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270195392))), name = tensor<string, []>("encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_13_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270195584)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_118_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270203840))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273349632))), name = tensor<string, []>("encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_13_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273349824)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_119_cast_fp16")];
+            tensor<fp16, []> var_3224_to_fp16 = const()[name = tensor<string, []>("op_3224_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3225_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3224_to_fp16)[name = tensor<string, []>("op_3225_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_3225_cast_fp16)[name = tensor<string, []>("input_713_cast_fp16")];
+            tensor<int32, [1]> key_27_axes_0 = const()[name = tensor<string, []>("key_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273351936)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273354048)))];
+            tensor<fp16, [1, 56, 1024]> key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = tensor<string, []>("key_27_cast_fp16")];
+            tensor<bool, []> input_715_interleave_0 = const()[name = tensor<string, []>("input_715_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_715_cast_fp16 = concat(axis = var_68, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = tensor<string, []>("input_715_cast_fp16")];
+            tensor<bool, []> var_3253_interleave_0 = const()[name = tensor<string, []>("op_3253_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_3253_cast_fp16 = concat(axis = var_68, interleave = var_3253_interleave_0, values = key_27_cast_fp16)[name = tensor<string, []>("op_3253_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273356160))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274142656))), name = tensor<string, []>("encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274142848)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_120_cast_fp16")];
+            tensor<int32, [4]> var_3258 = const()[name = tensor<string, []>("op_3258"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_79_cast_fp16 = reshape(shape = var_3258, x = linear_120_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274144960))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274931456))), name = tensor<string, []>("encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274931648)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_121_cast_fp16")];
+            tensor<int32, [4]> var_3263 = const()[name = tensor<string, []>("op_3263"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_53_cast_fp16 = reshape(shape = var_3263, x = linear_121_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274933760))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275720256))), name = tensor<string, []>("encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275720448)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_122_cast_fp16")];
+            tensor<int32, [4]> var_3268 = const()[name = tensor<string, []>("op_3268"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_27_cast_fp16 = reshape(shape = var_3268, x = linear_122_cast_fp16)[name = tensor<string, []>("v_27_cast_fp16")];
+            tensor<int32, [4]> value_35_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275722560)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3281_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3281_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275724672)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3283_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3283_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_27_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_345_transpose_x_0 = const()[name = tensor<string, []>("x_345_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_345_transpose_y_0 = const()[name = tensor<string, []>("x_345_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_3285_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3285_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275726784))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275926528))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3283_cast_fp16)[name = tensor<string, []>("transpose_245")];
+            tensor<fp16, [1, 8, 56, 195]> 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_3285_to_fp16_quantized)[name = tensor<string, []>("x_345_cast_fp16")];
+            tensor<int32, [8]> x_347_pad_0 = const()[name = tensor<string, []>("x_347_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_347_mode_0 = const()[name = tensor<string, []>("x_347_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_248_to_fp16 = const()[name = tensor<string, []>("const_248_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_347_cast_fp16")];
+            tensor<int32, [4]> var_3293 = const()[name = tensor<string, []>("op_3293"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_349_cast_fp16 = reshape(shape = var_3293, x = x_347_cast_fp16)[name = tensor<string, []>("x_349_cast_fp16")];
+            tensor<int32, [4]> var_3297_begin_0 = const()[name = tensor<string, []>("op_3297_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3297_end_0 = const()[name = tensor<string, []>("op_3297_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_3297_end_mask_0 = const()[name = tensor<string, []>("op_3297_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_3297_cast_fp16 = slice_by_index(begin = var_3297_begin_0, end = var_3297_end_0, end_mask = var_3297_end_mask_0, x = x_349_cast_fp16)[name = tensor<string, []>("op_3297_cast_fp16")];
+            tensor<int32, [4]> var_3298 = const()[name = tensor<string, []>("op_3298"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_53_cast_fp16 = reshape(shape = var_3298, x = var_3297_cast_fp16)[name = tensor<string, []>("matrix_bd_53_cast_fp16")];
+            tensor<bool, []> matrix_ac_27_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_27_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_27_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_27_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_122_perm_0 = const()[name = tensor<string, []>("transpose_122_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_123_perm_0 = const()[name = tensor<string, []>("transpose_123_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53_cast_fp16)[name = tensor<string, []>("transpose_243")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_122 = transpose(perm = transpose_122_perm_0, x = var_3281_cast_fp16)[name = tensor<string, []>("transpose_244")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_27_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_55_begin_0 = const()[name = tensor<string, []>("matrix_bd_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_55_end_0 = const()[name = tensor<string, []>("matrix_bd_55_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_55_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_55_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3307_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = tensor<string, []>("op_3307_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_53_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_53_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_53_cast_fp16 = mul(x = var_3307_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_53_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_55_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_55_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3313_cast_fp16 = softmax(axis = var_59, x = scores_55_cast_fp16)[name = tensor<string, []>("op_3313_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_717_cast_fp16 = select(a = var_44_to_fp16, b = var_3313_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_717_cast_fp16")];
+            tensor<bool, []> x_351_transpose_x_0 = const()[name = tensor<string, []>("x_351_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_351_transpose_y_0 = const()[name = tensor<string, []>("x_351_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = v_27_cast_fp16)[name = tensor<string, []>("transpose_242")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_351_cast_fp16")];
+            tensor<int32, [4]> var_3317_perm_0 = const()[name = tensor<string, []>("op_3317_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3318 = const()[name = tensor<string, []>("op_3318"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_3317_cast_fp16 = transpose(perm = var_3317_perm_0, x = x_351_cast_fp16)[name = tensor<string, []>("transpose_241")];
+            tensor<fp16, [1, 56, 1024]> input_719_cast_fp16 = reshape(shape = var_3318, x = var_3317_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275927040))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(276713536))), name = tensor<string, []>("encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(276713728)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_124_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = tensor<string, []>("input_723_cast_fp16")];
+            tensor<int32, [1]> x_355_axes_0 = const()[name = tensor<string, []>("x_355_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(276715840)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(276717952)))];
+            tensor<fp16, [1, 56, 1024]> x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = tensor<string, []>("x_355_cast_fp16")];
+            tensor<int32, [3]> input_725_perm_0 = const()[name = tensor<string, []>("input_725_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_727_pad_type_0 = const()[name = tensor<string, []>("input_727_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_727_strides_0 = const()[name = tensor<string, []>("input_727_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_727_pad_0 = const()[name = tensor<string, []>("input_727_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_727_dilations_0 = const()[name = tensor<string, []>("input_727_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_727_groups_0 = const()[name = tensor<string, []>("input_727_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(276720064))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278817280))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = tensor<string, []>("transpose_240")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_727_cast_fp16")];
+            tensor<int32, []> x_357_split_num_splits_0 = const()[name = tensor<string, []>("x_357_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_357_split_axis_0 = const()[name = tensor<string, []>("x_357_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_357_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_357_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = tensor<string, []>("x_357_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_357_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_729_cast_fp16 = select(a = var_44_to_fp16, b = x_357_cast_fp16, cond = var_575)[name = tensor<string, []>("input_729_cast_fp16")];
+            tensor<bool, []> new_x_55_interleave_0 = const()[name = tensor<string, []>("new_x_55_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_55_cast_fp16 = concat(axis = var_59, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = tensor<string, []>("new_x_55_cast_fp16")];
+            tensor<int32, [3]> var_3357_begin_0 = const()[name = tensor<string, []>("op_3357_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_3357_end_0 = const()[name = tensor<string, []>("op_3357_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_3357_end_mask_0 = const()[name = tensor<string, []>("op_3357_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3357_cast_fp16 = slice_by_index(begin = var_3357_begin_0, end = var_3357_end_0, end_mask = var_3357_end_mask_0, x = new_x_55_cast_fp16)[name = tensor<string, []>("op_3357_cast_fp16")];
+            tensor<string, []> x_359_pad_type_0 = const()[name = tensor<string, []>("x_359_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_359_groups_0 = const()[name = tensor<string, []>("x_359_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_359_strides_0 = const()[name = tensor<string, []>("x_359_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_359_pad_0 = const()[name = tensor<string, []>("x_359_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_359_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278821440))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278830720))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_359_cast_fp16")];
+            tensor<int32, [3]> input_731_perm_0 = const()[name = tensor<string, []>("input_731_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_361_axes_0 = const()[name = tensor<string, []>("x_361_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278832832)))];
+            tensor<fp16, [1024]> encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278834944)))];
+            tensor<fp16, [1, 56, 1024]> input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = tensor<string, []>("transpose_239")];
+            tensor<fp16, [1, 56, 1024]> x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = tensor<string, []>("x_361_cast_fp16")];
+            tensor<int32, [3]> input_733_perm_0 = const()[name = tensor<string, []>("input_733_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = tensor<string, []>("transpose_238")];
+            tensor<fp16, [1, 1024, 56]> input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = tensor<string, []>("input_735_cast_fp16")];
+            tensor<string, []> x_363_pad_type_0 = const()[name = tensor<string, []>("x_363_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_363_strides_0 = const()[name = tensor<string, []>("x_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_363_pad_0 = const()[name = tensor<string, []>("x_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_363_dilations_0 = const()[name = tensor<string, []>("x_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_363_groups_0 = const()[name = tensor<string, []>("x_363_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278837056))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279885696))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_363_cast_fp16")];
+            tensor<int32, [3]> input_737_perm_0 = const()[name = tensor<string, []>("input_737_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = tensor<string, []>("transpose_237")];
+            tensor<fp16, [1, 56, 1024]> input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = tensor<string, []>("input_739_cast_fp16")];
+            tensor<int32, [1]> input_741_axes_0 = const()[name = tensor<string, []>("input_741_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279887808)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279889920)))];
+            tensor<fp16, [1, 56, 1024]> input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279892032))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283037824))), name = tensor<string, []>("encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_13_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283038016)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_125_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283046272))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286192064))), name = tensor<string, []>("encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_13_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286192256)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_126_cast_fp16")];
+            tensor<fp16, []> var_3400_to_fp16 = const()[name = tensor<string, []>("op_3400_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3401_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3400_to_fp16)[name = tensor<string, []>("op_3401_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3401_cast_fp16)[name = tensor<string, []>("input_751_cast_fp16")];
+            tensor<int32, [1]> input_753_axes_0 = const()[name = tensor<string, []>("input_753_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286194368)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286196480)))];
+            tensor<fp16, [1, 56, 1024]> input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = tensor<string, []>("input_753_cast_fp16")];
+            tensor<int32, [4]> cache_57_begin_0 = const()[name = tensor<string, []>("cache_57_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_57_end_0 = const()[name = tensor<string, []>("cache_57_end_0"), val = tensor<int32, [4]>([15, 1, 42, 1024])];
+            tensor<bool, [4]> cache_57_end_mask_0 = const()[name = tensor<string, []>("cache_57_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_57_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_57_cast_fp16")];
+            tensor<int32, [4]> cache_59_begin_0 = const()[name = tensor<string, []>("cache_59_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_59_end_0 = const()[name = tensor<string, []>("cache_59_end_0"), val = tensor<int32, [4]>([15, 1, 1024, 8])];
+            tensor<bool, [4]> cache_59_end_mask_0 = const()[name = tensor<string, []>("cache_59_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_59_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_59_cast_fp16")];
+            tensor<int32, [1]> input_755_axes_0 = const()[name = tensor<string, []>("input_755_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286198592)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286200704)))];
+            tensor<fp16, [1, 56, 1024]> input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286202816))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289348608))), name = tensor<string, []>("encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_14_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289348800)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_127_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289357056))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292502848))), name = tensor<string, []>("encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_14_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292503040)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_128_cast_fp16")];
+            tensor<fp16, []> var_3437_to_fp16 = const()[name = tensor<string, []>("op_3437_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3438_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3437_to_fp16)[name = tensor<string, []>("op_3438_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3438_cast_fp16)[name = tensor<string, []>("input_765_cast_fp16")];
+            tensor<int32, [1]> key_29_axes_0 = const()[name = tensor<string, []>("key_29_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292505152)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292507264)))];
+            tensor<fp16, [1, 56, 1024]> key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = tensor<string, []>("key_29_cast_fp16")];
+            tensor<bool, []> input_767_interleave_0 = const()[name = tensor<string, []>("input_767_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_767_cast_fp16 = concat(axis = var_68, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = tensor<string, []>("input_767_cast_fp16")];
+            tensor<bool, []> var_3466_interleave_0 = const()[name = tensor<string, []>("op_3466_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_3466_cast_fp16 = concat(axis = var_68, interleave = var_3466_interleave_0, values = key_29_cast_fp16)[name = tensor<string, []>("op_3466_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292509376))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293295872))), name = tensor<string, []>("encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293296064)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_129_cast_fp16")];
+            tensor<int32, [4]> var_3471 = const()[name = tensor<string, []>("op_3471"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_85_cast_fp16 = reshape(shape = var_3471, x = linear_129_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293298176))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294084672))), name = tensor<string, []>("encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294084864)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_130_cast_fp16")];
+            tensor<int32, [4]> var_3476 = const()[name = tensor<string, []>("op_3476"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_57_cast_fp16 = reshape(shape = var_3476, x = linear_130_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294086976))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294873472))), name = tensor<string, []>("encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294873664)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_131_cast_fp16")];
+            tensor<int32, [4]> var_3481 = const()[name = tensor<string, []>("op_3481"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_29_cast_fp16 = reshape(shape = var_3481, x = linear_131_cast_fp16)[name = tensor<string, []>("v_29_cast_fp16")];
+            tensor<int32, [4]> value_37_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294875776)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3494_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3494_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294877888)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3496_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3496_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_29_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_371_transpose_x_0 = const()[name = tensor<string, []>("x_371_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_371_transpose_y_0 = const()[name = tensor<string, []>("x_371_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_3498_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3498_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294880000))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295079744))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3496_cast_fp16)[name = tensor<string, []>("transpose_236")];
+            tensor<fp16, [1, 8, 56, 195]> 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_3498_to_fp16_quantized)[name = tensor<string, []>("x_371_cast_fp16")];
+            tensor<int32, [8]> x_373_pad_0 = const()[name = tensor<string, []>("x_373_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_373_mode_0 = const()[name = tensor<string, []>("x_373_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_261_to_fp16 = const()[name = tensor<string, []>("const_261_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_373_cast_fp16")];
+            tensor<int32, [4]> var_3506 = const()[name = tensor<string, []>("op_3506"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_375_cast_fp16 = reshape(shape = var_3506, x = x_373_cast_fp16)[name = tensor<string, []>("x_375_cast_fp16")];
+            tensor<int32, [4]> var_3510_begin_0 = const()[name = tensor<string, []>("op_3510_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3510_end_0 = const()[name = tensor<string, []>("op_3510_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_3510_end_mask_0 = const()[name = tensor<string, []>("op_3510_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_3510_cast_fp16 = slice_by_index(begin = var_3510_begin_0, end = var_3510_end_0, end_mask = var_3510_end_mask_0, x = x_375_cast_fp16)[name = tensor<string, []>("op_3510_cast_fp16")];
+            tensor<int32, [4]> var_3511 = const()[name = tensor<string, []>("op_3511"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_57_cast_fp16 = reshape(shape = var_3511, x = var_3510_cast_fp16)[name = tensor<string, []>("matrix_bd_57_cast_fp16")];
+            tensor<bool, []> matrix_ac_29_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_29_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_29_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_29_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_124_perm_0 = const()[name = tensor<string, []>("transpose_124_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_125_perm_0 = const()[name = tensor<string, []>("transpose_125_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57_cast_fp16)[name = tensor<string, []>("transpose_234")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_124 = transpose(perm = transpose_124_perm_0, x = var_3494_cast_fp16)[name = tensor<string, []>("transpose_235")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_29_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_59_begin_0 = const()[name = tensor<string, []>("matrix_bd_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_59_end_0 = const()[name = tensor<string, []>("matrix_bd_59_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_59_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_59_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3520_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = tensor<string, []>("op_3520_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_57_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_57_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_57_cast_fp16 = mul(x = var_3520_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_57_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_59_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_59_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3526_cast_fp16 = softmax(axis = var_59, x = scores_59_cast_fp16)[name = tensor<string, []>("op_3526_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_769_cast_fp16 = select(a = var_44_to_fp16, b = var_3526_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_769_cast_fp16")];
+            tensor<bool, []> x_377_transpose_x_0 = const()[name = tensor<string, []>("x_377_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_377_transpose_y_0 = const()[name = tensor<string, []>("x_377_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = v_29_cast_fp16)[name = tensor<string, []>("transpose_233")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_377_cast_fp16")];
+            tensor<int32, [4]> var_3530_perm_0 = const()[name = tensor<string, []>("op_3530_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3531 = const()[name = tensor<string, []>("op_3531"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_3530_cast_fp16 = transpose(perm = var_3530_perm_0, x = x_377_cast_fp16)[name = tensor<string, []>("transpose_232")];
+            tensor<fp16, [1, 56, 1024]> input_771_cast_fp16 = reshape(shape = var_3531, x = var_3530_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295080256))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295866752))), name = tensor<string, []>("encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295866944)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_133_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = tensor<string, []>("input_775_cast_fp16")];
+            tensor<int32, [1]> x_381_axes_0 = const()[name = tensor<string, []>("x_381_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295869056)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295871168)))];
+            tensor<fp16, [1, 56, 1024]> x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = tensor<string, []>("x_381_cast_fp16")];
+            tensor<int32, [3]> input_777_perm_0 = const()[name = tensor<string, []>("input_777_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_779_pad_type_0 = const()[name = tensor<string, []>("input_779_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_779_strides_0 = const()[name = tensor<string, []>("input_779_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_779_pad_0 = const()[name = tensor<string, []>("input_779_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_779_dilations_0 = const()[name = tensor<string, []>("input_779_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_779_groups_0 = const()[name = tensor<string, []>("input_779_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295873280))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297970496))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = tensor<string, []>("transpose_231")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_779_cast_fp16")];
+            tensor<int32, []> x_383_split_num_splits_0 = const()[name = tensor<string, []>("x_383_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_383_split_axis_0 = const()[name = tensor<string, []>("x_383_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_383_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_383_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = tensor<string, []>("x_383_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_383_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_781_cast_fp16 = select(a = var_44_to_fp16, b = x_383_cast_fp16, cond = var_575)[name = tensor<string, []>("input_781_cast_fp16")];
+            tensor<bool, []> new_x_59_interleave_0 = const()[name = tensor<string, []>("new_x_59_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_59_cast_fp16 = concat(axis = var_59, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = tensor<string, []>("new_x_59_cast_fp16")];
+            tensor<int32, [3]> var_3570_begin_0 = const()[name = tensor<string, []>("op_3570_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_3570_end_0 = const()[name = tensor<string, []>("op_3570_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_3570_end_mask_0 = const()[name = tensor<string, []>("op_3570_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3570_cast_fp16 = slice_by_index(begin = var_3570_begin_0, end = var_3570_end_0, end_mask = var_3570_end_mask_0, x = new_x_59_cast_fp16)[name = tensor<string, []>("op_3570_cast_fp16")];
+            tensor<string, []> x_385_pad_type_0 = const()[name = tensor<string, []>("x_385_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_385_groups_0 = const()[name = tensor<string, []>("x_385_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_385_strides_0 = const()[name = tensor<string, []>("x_385_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_385_pad_0 = const()[name = tensor<string, []>("x_385_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_385_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297974656))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297983936))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_385_cast_fp16")];
+            tensor<int32, [3]> input_783_perm_0 = const()[name = tensor<string, []>("input_783_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_387_axes_0 = const()[name = tensor<string, []>("x_387_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297986048)))];
+            tensor<fp16, [1024]> encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297988160)))];
+            tensor<fp16, [1, 56, 1024]> input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = tensor<string, []>("transpose_230")];
+            tensor<fp16, [1, 56, 1024]> x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = tensor<string, []>("x_387_cast_fp16")];
+            tensor<int32, [3]> input_785_perm_0 = const()[name = tensor<string, []>("input_785_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = tensor<string, []>("transpose_229")];
+            tensor<fp16, [1, 1024, 56]> input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = tensor<string, []>("input_787_cast_fp16")];
+            tensor<string, []> x_389_pad_type_0 = const()[name = tensor<string, []>("x_389_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_389_strides_0 = const()[name = tensor<string, []>("x_389_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_389_pad_0 = const()[name = tensor<string, []>("x_389_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_389_dilations_0 = const()[name = tensor<string, []>("x_389_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_389_groups_0 = const()[name = tensor<string, []>("x_389_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297990272))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(299038912))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_389_cast_fp16")];
+            tensor<int32, [3]> input_789_perm_0 = const()[name = tensor<string, []>("input_789_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = tensor<string, []>("transpose_228")];
+            tensor<fp16, [1, 56, 1024]> input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = tensor<string, []>("input_791_cast_fp16")];
+            tensor<int32, [1]> input_793_axes_0 = const()[name = tensor<string, []>("input_793_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(299041024)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(299043136)))];
+            tensor<fp16, [1, 56, 1024]> input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(299045248))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302191040))), name = tensor<string, []>("encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_14_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302191232)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_134_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302199488))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305345280))), name = tensor<string, []>("encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_14_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305345472)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_135_cast_fp16")];
+            tensor<fp16, []> var_3613_to_fp16 = const()[name = tensor<string, []>("op_3613_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3614_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3613_to_fp16)[name = tensor<string, []>("op_3614_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3614_cast_fp16)[name = tensor<string, []>("input_803_cast_fp16")];
+            tensor<int32, [1]> input_805_axes_0 = const()[name = tensor<string, []>("input_805_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305347584)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305349696)))];
+            tensor<fp16, [1, 56, 1024]> input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = tensor<string, []>("input_805_cast_fp16")];
+            tensor<int32, [4]> cache_61_begin_0 = const()[name = tensor<string, []>("cache_61_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_61_end_0 = const()[name = tensor<string, []>("cache_61_end_0"), val = tensor<int32, [4]>([16, 1, 42, 1024])];
+            tensor<bool, [4]> cache_61_end_mask_0 = const()[name = tensor<string, []>("cache_61_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_61_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_61_cast_fp16")];
+            tensor<int32, [4]> cache_63_begin_0 = const()[name = tensor<string, []>("cache_63_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_63_end_0 = const()[name = tensor<string, []>("cache_63_end_0"), val = tensor<int32, [4]>([16, 1, 1024, 8])];
+            tensor<bool, [4]> cache_63_end_mask_0 = const()[name = tensor<string, []>("cache_63_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_63_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_63_cast_fp16")];
+            tensor<int32, [1]> input_807_axes_0 = const()[name = tensor<string, []>("input_807_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305351808)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305353920)))];
+            tensor<fp16, [1, 56, 1024]> input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305356032))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308501824))), name = tensor<string, []>("encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_15_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308502016)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_136_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308510272))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311656064))), name = tensor<string, []>("encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_15_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311656256)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_137_cast_fp16")];
+            tensor<fp16, []> var_3650_to_fp16 = const()[name = tensor<string, []>("op_3650_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3651_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3650_to_fp16)[name = tensor<string, []>("op_3651_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3651_cast_fp16)[name = tensor<string, []>("input_817_cast_fp16")];
+            tensor<int32, [1]> key_31_axes_0 = const()[name = tensor<string, []>("key_31_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311658368)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311660480)))];
+            tensor<fp16, [1, 56, 1024]> key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = tensor<string, []>("key_31_cast_fp16")];
+            tensor<bool, []> input_819_interleave_0 = const()[name = tensor<string, []>("input_819_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_819_cast_fp16 = concat(axis = var_68, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = tensor<string, []>("input_819_cast_fp16")];
+            tensor<bool, []> var_3679_interleave_0 = const()[name = tensor<string, []>("op_3679_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_3679_cast_fp16 = concat(axis = var_68, interleave = var_3679_interleave_0, values = key_31_cast_fp16)[name = tensor<string, []>("op_3679_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311662592))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312449088))), name = tensor<string, []>("encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312449280)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_138_cast_fp16")];
+            tensor<int32, [4]> var_3684 = const()[name = tensor<string, []>("op_3684"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_91_cast_fp16 = reshape(shape = var_3684, x = linear_138_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312451392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313237888))), name = tensor<string, []>("encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313238080)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_139_cast_fp16")];
+            tensor<int32, [4]> var_3689 = const()[name = tensor<string, []>("op_3689"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_61_cast_fp16 = reshape(shape = var_3689, x = linear_139_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313240192))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314026688))), name = tensor<string, []>("encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314026880)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_140_cast_fp16")];
+            tensor<int32, [4]> var_3694 = const()[name = tensor<string, []>("op_3694"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_31_cast_fp16 = reshape(shape = var_3694, x = linear_140_cast_fp16)[name = tensor<string, []>("v_31_cast_fp16")];
+            tensor<int32, [4]> value_39_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314028992)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3707_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3707_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314031104)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3709_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3709_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_31_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_397_transpose_x_0 = const()[name = tensor<string, []>("x_397_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_397_transpose_y_0 = const()[name = tensor<string, []>("x_397_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_3711_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3711_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314033216))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314232960))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3709_cast_fp16)[name = tensor<string, []>("transpose_227")];
+            tensor<fp16, [1, 8, 56, 195]> 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_3711_to_fp16_quantized)[name = tensor<string, []>("x_397_cast_fp16")];
+            tensor<int32, [8]> x_399_pad_0 = const()[name = tensor<string, []>("x_399_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_399_mode_0 = const()[name = tensor<string, []>("x_399_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_274_to_fp16 = const()[name = tensor<string, []>("const_274_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_399_cast_fp16")];
+            tensor<int32, [4]> var_3719 = const()[name = tensor<string, []>("op_3719"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_401_cast_fp16 = reshape(shape = var_3719, x = x_399_cast_fp16)[name = tensor<string, []>("x_401_cast_fp16")];
+            tensor<int32, [4]> var_3723_begin_0 = const()[name = tensor<string, []>("op_3723_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3723_end_0 = const()[name = tensor<string, []>("op_3723_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_3723_end_mask_0 = const()[name = tensor<string, []>("op_3723_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_3723_cast_fp16 = slice_by_index(begin = var_3723_begin_0, end = var_3723_end_0, end_mask = var_3723_end_mask_0, x = x_401_cast_fp16)[name = tensor<string, []>("op_3723_cast_fp16")];
+            tensor<int32, [4]> var_3724 = const()[name = tensor<string, []>("op_3724"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_61_cast_fp16 = reshape(shape = var_3724, x = var_3723_cast_fp16)[name = tensor<string, []>("matrix_bd_61_cast_fp16")];
+            tensor<bool, []> matrix_ac_31_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_31_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_31_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_31_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_126_perm_0 = const()[name = tensor<string, []>("transpose_126_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_127_perm_0 = const()[name = tensor<string, []>("transpose_127_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61_cast_fp16)[name = tensor<string, []>("transpose_225")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_126 = transpose(perm = transpose_126_perm_0, x = var_3707_cast_fp16)[name = tensor<string, []>("transpose_226")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_31_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_63_begin_0 = const()[name = tensor<string, []>("matrix_bd_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_63_end_0 = const()[name = tensor<string, []>("matrix_bd_63_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_63_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_63_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3733_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = tensor<string, []>("op_3733_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_61_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_61_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_61_cast_fp16 = mul(x = var_3733_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_61_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_63_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_63_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3739_cast_fp16 = softmax(axis = var_59, x = scores_63_cast_fp16)[name = tensor<string, []>("op_3739_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_821_cast_fp16 = select(a = var_44_to_fp16, b = var_3739_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_821_cast_fp16")];
+            tensor<bool, []> x_403_transpose_x_0 = const()[name = tensor<string, []>("x_403_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_403_transpose_y_0 = const()[name = tensor<string, []>("x_403_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = v_31_cast_fp16)[name = tensor<string, []>("transpose_224")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_403_cast_fp16")];
+            tensor<int32, [4]> var_3743_perm_0 = const()[name = tensor<string, []>("op_3743_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3744 = const()[name = tensor<string, []>("op_3744"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_3743_cast_fp16 = transpose(perm = var_3743_perm_0, x = x_403_cast_fp16)[name = tensor<string, []>("transpose_223")];
+            tensor<fp16, [1, 56, 1024]> input_823_cast_fp16 = reshape(shape = var_3744, x = var_3743_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314233472))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315019968))), name = tensor<string, []>("encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315020160)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_142_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = tensor<string, []>("input_827_cast_fp16")];
+            tensor<int32, [1]> x_407_axes_0 = const()[name = tensor<string, []>("x_407_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315022272)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315024384)))];
+            tensor<fp16, [1, 56, 1024]> x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = tensor<string, []>("x_407_cast_fp16")];
+            tensor<int32, [3]> input_829_perm_0 = const()[name = tensor<string, []>("input_829_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_831_pad_type_0 = const()[name = tensor<string, []>("input_831_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_831_strides_0 = const()[name = tensor<string, []>("input_831_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_831_pad_0 = const()[name = tensor<string, []>("input_831_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_831_dilations_0 = const()[name = tensor<string, []>("input_831_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_831_groups_0 = const()[name = tensor<string, []>("input_831_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315026496))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317123712))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = tensor<string, []>("transpose_222")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_831_cast_fp16")];
+            tensor<int32, []> x_409_split_num_splits_0 = const()[name = tensor<string, []>("x_409_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_409_split_axis_0 = const()[name = tensor<string, []>("x_409_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_409_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_409_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = tensor<string, []>("x_409_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_409_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_833_cast_fp16 = select(a = var_44_to_fp16, b = x_409_cast_fp16, cond = var_575)[name = tensor<string, []>("input_833_cast_fp16")];
+            tensor<bool, []> new_x_63_interleave_0 = const()[name = tensor<string, []>("new_x_63_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_63_cast_fp16 = concat(axis = var_59, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = tensor<string, []>("new_x_63_cast_fp16")];
+            tensor<int32, [3]> var_3783_begin_0 = const()[name = tensor<string, []>("op_3783_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_3783_end_0 = const()[name = tensor<string, []>("op_3783_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_3783_end_mask_0 = const()[name = tensor<string, []>("op_3783_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3783_cast_fp16 = slice_by_index(begin = var_3783_begin_0, end = var_3783_end_0, end_mask = var_3783_end_mask_0, x = new_x_63_cast_fp16)[name = tensor<string, []>("op_3783_cast_fp16")];
+            tensor<string, []> x_411_pad_type_0 = const()[name = tensor<string, []>("x_411_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_411_groups_0 = const()[name = tensor<string, []>("x_411_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_411_strides_0 = const()[name = tensor<string, []>("x_411_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_411_pad_0 = const()[name = tensor<string, []>("x_411_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_411_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317127872))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317137152))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_411_cast_fp16")];
+            tensor<int32, [3]> input_835_perm_0 = const()[name = tensor<string, []>("input_835_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_413_axes_0 = const()[name = tensor<string, []>("x_413_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317139264)))];
+            tensor<fp16, [1024]> encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317141376)))];
+            tensor<fp16, [1, 56, 1024]> input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = tensor<string, []>("transpose_221")];
+            tensor<fp16, [1, 56, 1024]> x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = tensor<string, []>("x_413_cast_fp16")];
+            tensor<int32, [3]> input_837_perm_0 = const()[name = tensor<string, []>("input_837_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = tensor<string, []>("transpose_220")];
+            tensor<fp16, [1, 1024, 56]> input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = tensor<string, []>("input_839_cast_fp16")];
+            tensor<string, []> x_415_pad_type_0 = const()[name = tensor<string, []>("x_415_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_415_strides_0 = const()[name = tensor<string, []>("x_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_415_pad_0 = const()[name = tensor<string, []>("x_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_415_dilations_0 = const()[name = tensor<string, []>("x_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_415_groups_0 = const()[name = tensor<string, []>("x_415_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317143488))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318192128))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_415_cast_fp16")];
+            tensor<int32, [3]> input_841_perm_0 = const()[name = tensor<string, []>("input_841_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = tensor<string, []>("transpose_219")];
+            tensor<fp16, [1, 56, 1024]> input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = tensor<string, []>("input_843_cast_fp16")];
+            tensor<int32, [1]> input_845_axes_0 = const()[name = tensor<string, []>("input_845_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318194240)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318196352)))];
+            tensor<fp16, [1, 56, 1024]> input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318198464))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321344256))), name = tensor<string, []>("encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_15_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321344448)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_143_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321352704))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324498496))), name = tensor<string, []>("encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_15_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324498688)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_144_cast_fp16")];
+            tensor<fp16, []> var_3826_to_fp16 = const()[name = tensor<string, []>("op_3826_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3827_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3826_to_fp16)[name = tensor<string, []>("op_3827_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3827_cast_fp16)[name = tensor<string, []>("input_855_cast_fp16")];
+            tensor<int32, [1]> input_857_axes_0 = const()[name = tensor<string, []>("input_857_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324500800)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324502912)))];
+            tensor<fp16, [1, 56, 1024]> input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = tensor<string, []>("input_857_cast_fp16")];
+            tensor<int32, [4]> cache_65_begin_0 = const()[name = tensor<string, []>("cache_65_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_65_end_0 = const()[name = tensor<string, []>("cache_65_end_0"), val = tensor<int32, [4]>([17, 1, 42, 1024])];
+            tensor<bool, [4]> cache_65_end_mask_0 = const()[name = tensor<string, []>("cache_65_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_65_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_65_cast_fp16")];
+            tensor<int32, [4]> cache_67_begin_0 = const()[name = tensor<string, []>("cache_67_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_67_end_0 = const()[name = tensor<string, []>("cache_67_end_0"), val = tensor<int32, [4]>([17, 1, 1024, 8])];
+            tensor<bool, [4]> cache_67_end_mask_0 = const()[name = tensor<string, []>("cache_67_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_67_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_67_cast_fp16")];
+            tensor<int32, [1]> input_859_axes_0 = const()[name = tensor<string, []>("input_859_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324505024)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324507136)))];
+            tensor<fp16, [1, 56, 1024]> input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324509248))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(327655040))), name = tensor<string, []>("encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_16_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(327655232)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_145_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(327663488))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330809280))), name = tensor<string, []>("encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_16_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330809472)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_146_cast_fp16")];
+            tensor<fp16, []> var_3863_to_fp16 = const()[name = tensor<string, []>("op_3863_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3864_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3863_to_fp16)[name = tensor<string, []>("op_3864_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3864_cast_fp16)[name = tensor<string, []>("input_869_cast_fp16")];
+            tensor<int32, [1]> key_33_axes_0 = const()[name = tensor<string, []>("key_33_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330811584)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330813696)))];
+            tensor<fp16, [1, 56, 1024]> key_33_cast_fp16 = layer_norm(axes = key_33_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = tensor<string, []>("key_33_cast_fp16")];
+            tensor<bool, []> input_871_interleave_0 = const()[name = tensor<string, []>("input_871_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_871_cast_fp16 = concat(axis = var_68, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_33_cast_fp16))[name = tensor<string, []>("input_871_cast_fp16")];
+            tensor<bool, []> var_3892_interleave_0 = const()[name = tensor<string, []>("op_3892_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_3892_cast_fp16 = concat(axis = var_68, interleave = var_3892_interleave_0, values = key_33_cast_fp16)[name = tensor<string, []>("op_3892_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330815808))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331602304))), name = tensor<string, []>("encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331602496)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_147_cast_fp16")];
+            tensor<int32, [4]> var_3897 = const()[name = tensor<string, []>("op_3897"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_97_cast_fp16 = reshape(shape = var_3897, x = linear_147_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331604608))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332391104))), name = tensor<string, []>("encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332391296)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_148_cast_fp16")];
+            tensor<int32, [4]> var_3902 = const()[name = tensor<string, []>("op_3902"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_65_cast_fp16 = reshape(shape = var_3902, x = linear_148_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332393408))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333179904))), name = tensor<string, []>("encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333180096)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_149_cast_fp16")];
+            tensor<int32, [4]> var_3907 = const()[name = tensor<string, []>("op_3907"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_33_cast_fp16 = reshape(shape = var_3907, x = linear_149_cast_fp16)[name = tensor<string, []>("v_33_cast_fp16")];
+            tensor<int32, [4]> value_41_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333182208)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3920_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3920_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333184320)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3922_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3922_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_33_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_423_transpose_x_0 = const()[name = tensor<string, []>("x_423_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_423_transpose_y_0 = const()[name = tensor<string, []>("x_423_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_3924_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3924_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333186432))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333386176))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_33_cast_fp16 = transpose(perm = q_with_bias_v_33_perm_0, x = var_3922_cast_fp16)[name = tensor<string, []>("transpose_218")];
+            tensor<fp16, [1, 8, 56, 195]> 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_3924_to_fp16_quantized)[name = tensor<string, []>("x_423_cast_fp16")];
+            tensor<int32, [8]> x_425_pad_0 = const()[name = tensor<string, []>("x_425_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_425_mode_0 = const()[name = tensor<string, []>("x_425_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_287_to_fp16 = const()[name = tensor<string, []>("const_287_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_425_cast_fp16")];
+            tensor<int32, [4]> var_3932 = const()[name = tensor<string, []>("op_3932"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_427_cast_fp16 = reshape(shape = var_3932, x = x_425_cast_fp16)[name = tensor<string, []>("x_427_cast_fp16")];
+            tensor<int32, [4]> var_3936_begin_0 = const()[name = tensor<string, []>("op_3936_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3936_end_0 = const()[name = tensor<string, []>("op_3936_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_3936_end_mask_0 = const()[name = tensor<string, []>("op_3936_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_3936_cast_fp16 = slice_by_index(begin = var_3936_begin_0, end = var_3936_end_0, end_mask = var_3936_end_mask_0, x = x_427_cast_fp16)[name = tensor<string, []>("op_3936_cast_fp16")];
+            tensor<int32, [4]> var_3937 = const()[name = tensor<string, []>("op_3937"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_65_cast_fp16 = reshape(shape = var_3937, x = var_3936_cast_fp16)[name = tensor<string, []>("matrix_bd_65_cast_fp16")];
+            tensor<bool, []> matrix_ac_33_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_33_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_33_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_33_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_128_perm_0 = const()[name = tensor<string, []>("transpose_128_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_129_perm_0 = const()[name = tensor<string, []>("transpose_129_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65_cast_fp16)[name = tensor<string, []>("transpose_216")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_128 = transpose(perm = transpose_128_perm_0, x = var_3920_cast_fp16)[name = tensor<string, []>("transpose_217")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_33_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_67_begin_0 = const()[name = tensor<string, []>("matrix_bd_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_67_end_0 = const()[name = tensor<string, []>("matrix_bd_67_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_67_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_67_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3946_cast_fp16 = add(x = matrix_ac_33_cast_fp16, y = matrix_bd_67_cast_fp16)[name = tensor<string, []>("op_3946_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_65_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_65_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_65_cast_fp16 = mul(x = var_3946_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_65_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_67_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_67_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3952_cast_fp16 = softmax(axis = var_59, x = scores_67_cast_fp16)[name = tensor<string, []>("op_3952_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_873_cast_fp16 = select(a = var_44_to_fp16, b = var_3952_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_873_cast_fp16")];
+            tensor<bool, []> x_429_transpose_x_0 = const()[name = tensor<string, []>("x_429_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_429_transpose_y_0 = const()[name = tensor<string, []>("x_429_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = v_33_cast_fp16)[name = tensor<string, []>("transpose_215")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_429_cast_fp16")];
+            tensor<int32, [4]> var_3956_perm_0 = const()[name = tensor<string, []>("op_3956_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3957 = const()[name = tensor<string, []>("op_3957"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_3956_cast_fp16 = transpose(perm = var_3956_perm_0, x = x_429_cast_fp16)[name = tensor<string, []>("transpose_214")];
+            tensor<fp16, [1, 56, 1024]> input_875_cast_fp16 = reshape(shape = var_3957, x = var_3956_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333386688))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334173184))), name = tensor<string, []>("encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334173376)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_151_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = tensor<string, []>("input_879_cast_fp16")];
+            tensor<int32, [1]> x_433_axes_0 = const()[name = tensor<string, []>("x_433_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334175488)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334177600)))];
+            tensor<fp16, [1, 56, 1024]> x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = tensor<string, []>("x_433_cast_fp16")];
+            tensor<int32, [3]> input_881_perm_0 = const()[name = tensor<string, []>("input_881_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_883_pad_type_0 = const()[name = tensor<string, []>("input_883_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_883_strides_0 = const()[name = tensor<string, []>("input_883_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_883_pad_0 = const()[name = tensor<string, []>("input_883_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_883_dilations_0 = const()[name = tensor<string, []>("input_883_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_883_groups_0 = const()[name = tensor<string, []>("input_883_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334179712))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336276928))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = tensor<string, []>("transpose_213")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_883_cast_fp16")];
+            tensor<int32, []> x_435_split_num_splits_0 = const()[name = tensor<string, []>("x_435_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_435_split_axis_0 = const()[name = tensor<string, []>("x_435_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_435_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_435_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = tensor<string, []>("x_435_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_435_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_885_cast_fp16 = select(a = var_44_to_fp16, b = x_435_cast_fp16, cond = var_575)[name = tensor<string, []>("input_885_cast_fp16")];
+            tensor<bool, []> new_x_67_interleave_0 = const()[name = tensor<string, []>("new_x_67_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_67_cast_fp16 = concat(axis = var_59, interleave = new_x_67_interleave_0, values = (cache_67_cast_fp16, input_885_cast_fp16))[name = tensor<string, []>("new_x_67_cast_fp16")];
+            tensor<int32, [3]> var_3996_begin_0 = const()[name = tensor<string, []>("op_3996_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_3996_end_0 = const()[name = tensor<string, []>("op_3996_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_3996_end_mask_0 = const()[name = tensor<string, []>("op_3996_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3996_cast_fp16 = slice_by_index(begin = var_3996_begin_0, end = var_3996_end_0, end_mask = var_3996_end_mask_0, x = new_x_67_cast_fp16)[name = tensor<string, []>("op_3996_cast_fp16")];
+            tensor<string, []> x_437_pad_type_0 = const()[name = tensor<string, []>("x_437_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_437_groups_0 = const()[name = tensor<string, []>("x_437_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_437_strides_0 = const()[name = tensor<string, []>("x_437_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_437_pad_0 = const()[name = tensor<string, []>("x_437_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_437_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336281088))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336290368))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_437_cast_fp16")];
+            tensor<int32, [3]> input_887_perm_0 = const()[name = tensor<string, []>("input_887_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_439_axes_0 = const()[name = tensor<string, []>("x_439_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336292480)))];
+            tensor<fp16, [1024]> encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336294592)))];
+            tensor<fp16, [1, 56, 1024]> input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = tensor<string, []>("transpose_212")];
+            tensor<fp16, [1, 56, 1024]> x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = tensor<string, []>("x_439_cast_fp16")];
+            tensor<int32, [3]> input_889_perm_0 = const()[name = tensor<string, []>("input_889_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = tensor<string, []>("transpose_211")];
+            tensor<fp16, [1, 1024, 56]> input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = tensor<string, []>("input_891_cast_fp16")];
+            tensor<string, []> x_441_pad_type_0 = const()[name = tensor<string, []>("x_441_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_441_strides_0 = const()[name = tensor<string, []>("x_441_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_441_pad_0 = const()[name = tensor<string, []>("x_441_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_441_dilations_0 = const()[name = tensor<string, []>("x_441_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_441_groups_0 = const()[name = tensor<string, []>("x_441_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336296704))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337345344))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_441_cast_fp16")];
+            tensor<int32, [3]> input_893_perm_0 = const()[name = tensor<string, []>("input_893_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = tensor<string, []>("transpose_210")];
+            tensor<fp16, [1, 56, 1024]> input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = tensor<string, []>("input_895_cast_fp16")];
+            tensor<int32, [1]> input_897_axes_0 = const()[name = tensor<string, []>("input_897_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337347456)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337349568)))];
+            tensor<fp16, [1, 56, 1024]> input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337351680))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(340497472))), name = tensor<string, []>("encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_16_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(340497664)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_152_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(340505920))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343651712))), name = tensor<string, []>("encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_16_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343651904)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_153_cast_fp16")];
+            tensor<fp16, []> var_4039_to_fp16 = const()[name = tensor<string, []>("op_4039_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4040_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_4039_to_fp16)[name = tensor<string, []>("op_4040_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_907_cast_fp16 = add(x = input_895_cast_fp16, y = var_4040_cast_fp16)[name = tensor<string, []>("input_907_cast_fp16")];
+            tensor<int32, [1]> input_909_axes_0 = const()[name = tensor<string, []>("input_909_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343654016)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343656128)))];
+            tensor<fp16, [1, 56, 1024]> input_909_cast_fp16 = layer_norm(axes = input_909_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_907_cast_fp16)[name = tensor<string, []>("input_909_cast_fp16")];
+            tensor<int32, [4]> cache_69_begin_0 = const()[name = tensor<string, []>("cache_69_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> cache_69_end_0 = const()[name = tensor<string, []>("cache_69_end_0"), val = tensor<int32, [4]>([18, 1, 42, 1024])];
+            tensor<bool, [4]> cache_69_end_mask_0 = const()[name = tensor<string, []>("cache_69_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_69_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_69_cast_fp16")];
+            tensor<int32, [4]> cache_71_begin_0 = const()[name = tensor<string, []>("cache_71_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> cache_71_end_0 = const()[name = tensor<string, []>("cache_71_end_0"), val = tensor<int32, [4]>([18, 1, 1024, 8])];
+            tensor<bool, [4]> cache_71_end_mask_0 = const()[name = tensor<string, []>("cache_71_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_71_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_71_cast_fp16")];
+            tensor<int32, [1]> input_911_axes_0 = const()[name = tensor<string, []>("input_911_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343658240)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343660352)))];
+            tensor<fp16, [1, 56, 1024]> input_911_cast_fp16 = layer_norm(axes = input_911_axes_0, beta = encoder_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_feed_forward1_weight_to_fp16, x = input_909_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343662464))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346808256))), name = tensor<string, []>("encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_17_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346808448)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_154_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_915_cast_fp16 = silu(x = linear_154_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346816704))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349962496))), name = tensor<string, []>("encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_17_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349962688)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_155_cast_fp16")];
+            tensor<fp16, []> var_4076_to_fp16 = const()[name = tensor<string, []>("op_4076_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4077_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_4076_to_fp16)[name = tensor<string, []>("op_4077_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_921_cast_fp16 = add(x = input_909_cast_fp16, y = var_4077_cast_fp16)[name = tensor<string, []>("input_921_cast_fp16")];
+            tensor<int32, [1]> key_35_axes_0 = const()[name = tensor<string, []>("key_35_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349964800)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349966912)))];
+            tensor<fp16, [1, 56, 1024]> key_35_cast_fp16 = layer_norm(axes = key_35_axes_0, beta = encoder_layers_17_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_self_att_weight_to_fp16, x = input_921_cast_fp16)[name = tensor<string, []>("key_35_cast_fp16")];
+            tensor<bool, []> input_923_interleave_0 = const()[name = tensor<string, []>("input_923_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_923_cast_fp16 = concat(axis = var_68, interleave = input_923_interleave_0, values = (cache_69_cast_fp16, key_35_cast_fp16))[name = tensor<string, []>("input_923_cast_fp16")];
+            tensor<bool, []> var_4105_interleave_0 = const()[name = tensor<string, []>("op_4105_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_4105_cast_fp16 = concat(axis = var_68, interleave = var_4105_interleave_0, values = key_35_cast_fp16)[name = tensor<string, []>("op_4105_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349969024))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350755520))), name = tensor<string, []>("encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350755712)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_156_cast_fp16")];
+            tensor<int32, [4]> var_4110 = const()[name = tensor<string, []>("op_4110"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_103_cast_fp16 = reshape(shape = var_4110, x = linear_156_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350757824))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(351544320))), name = tensor<string, []>("encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(351544512)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_157_cast_fp16")];
+            tensor<int32, [4]> var_4115 = const()[name = tensor<string, []>("op_4115"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_69_cast_fp16 = reshape(shape = var_4115, x = linear_157_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(351546624))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(352333120))), name = tensor<string, []>("encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(352333312)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_158_cast_fp16")];
+            tensor<int32, [4]> var_4120 = const()[name = tensor<string, []>("op_4120"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_35_cast_fp16 = reshape(shape = var_4120, x = linear_158_cast_fp16)[name = tensor<string, []>("v_35_cast_fp16")];
+            tensor<int32, [4]> value_43_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_17_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(352335424)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4133_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_4133_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_17_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(352337536)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4135_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_4135_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_35_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_449_transpose_x_0 = const()[name = tensor<string, []>("x_449_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_449_transpose_y_0 = const()[name = tensor<string, []>("x_449_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_4137_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_4137_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(352339648))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(352539392))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_35_cast_fp16 = transpose(perm = q_with_bias_v_35_perm_0, x = var_4135_cast_fp16)[name = tensor<string, []>("transpose_209")];
+            tensor<fp16, [1, 8, 56, 195]> 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_4137_to_fp16_quantized)[name = tensor<string, []>("x_449_cast_fp16")];
+            tensor<int32, [8]> x_451_pad_0 = const()[name = tensor<string, []>("x_451_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_451_mode_0 = const()[name = tensor<string, []>("x_451_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_300_to_fp16 = const()[name = tensor<string, []>("const_300_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_451_cast_fp16")];
+            tensor<int32, [4]> var_4145 = const()[name = tensor<string, []>("op_4145"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_453_cast_fp16 = reshape(shape = var_4145, x = x_451_cast_fp16)[name = tensor<string, []>("x_453_cast_fp16")];
+            tensor<int32, [4]> var_4149_begin_0 = const()[name = tensor<string, []>("op_4149_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4149_end_0 = const()[name = tensor<string, []>("op_4149_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_4149_end_mask_0 = const()[name = tensor<string, []>("op_4149_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_4149_cast_fp16 = slice_by_index(begin = var_4149_begin_0, end = var_4149_end_0, end_mask = var_4149_end_mask_0, x = x_453_cast_fp16)[name = tensor<string, []>("op_4149_cast_fp16")];
+            tensor<int32, [4]> var_4150 = const()[name = tensor<string, []>("op_4150"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_69_cast_fp16 = reshape(shape = var_4150, x = var_4149_cast_fp16)[name = tensor<string, []>("matrix_bd_69_cast_fp16")];
+            tensor<bool, []> matrix_ac_35_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_35_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_35_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_35_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_130_perm_0 = const()[name = tensor<string, []>("transpose_130_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_131_perm_0 = const()[name = tensor<string, []>("transpose_131_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69_cast_fp16)[name = tensor<string, []>("transpose_207")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_130 = transpose(perm = transpose_130_perm_0, x = var_4133_cast_fp16)[name = tensor<string, []>("transpose_208")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_35_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_71_begin_0 = const()[name = tensor<string, []>("matrix_bd_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_71_end_0 = const()[name = tensor<string, []>("matrix_bd_71_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_71_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_71_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_4159_cast_fp16 = add(x = matrix_ac_35_cast_fp16, y = matrix_bd_71_cast_fp16)[name = tensor<string, []>("op_4159_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_69_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_69_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_69_cast_fp16 = mul(x = var_4159_cast_fp16, y = _inversed_scores_69_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_69_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_71_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_69_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_71_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_4165_cast_fp16 = softmax(axis = var_59, x = scores_71_cast_fp16)[name = tensor<string, []>("op_4165_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_925_cast_fp16 = select(a = var_44_to_fp16, b = var_4165_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_925_cast_fp16")];
+            tensor<bool, []> x_455_transpose_x_0 = const()[name = tensor<string, []>("x_455_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_455_transpose_y_0 = const()[name = tensor<string, []>("x_455_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = v_35_cast_fp16)[name = tensor<string, []>("transpose_206")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_455_cast_fp16")];
+            tensor<int32, [4]> var_4169_perm_0 = const()[name = tensor<string, []>("op_4169_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4170 = const()[name = tensor<string, []>("op_4170"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_4169_cast_fp16 = transpose(perm = var_4169_perm_0, x = x_455_cast_fp16)[name = tensor<string, []>("transpose_205")];
+            tensor<fp16, [1, 56, 1024]> input_927_cast_fp16 = reshape(shape = var_4170, x = var_4169_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(352539904))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353326400))), name = tensor<string, []>("encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353326592)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_160_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_931_cast_fp16 = add(x = input_921_cast_fp16, y = linear_160_cast_fp16)[name = tensor<string, []>("input_931_cast_fp16")];
+            tensor<int32, [1]> x_459_axes_0 = const()[name = tensor<string, []>("x_459_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353328704)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353330816)))];
+            tensor<fp16, [1, 56, 1024]> x_459_cast_fp16 = layer_norm(axes = x_459_axes_0, beta = encoder_layers_17_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_conv_weight_to_fp16, x = input_931_cast_fp16)[name = tensor<string, []>("x_459_cast_fp16")];
+            tensor<int32, [3]> input_933_perm_0 = const()[name = tensor<string, []>("input_933_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_935_pad_type_0 = const()[name = tensor<string, []>("input_935_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_935_strides_0 = const()[name = tensor<string, []>("input_935_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_935_pad_0 = const()[name = tensor<string, []>("input_935_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_935_dilations_0 = const()[name = tensor<string, []>("input_935_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_935_groups_0 = const()[name = tensor<string, []>("input_935_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353332928))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355430144))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_933_cast_fp16 = transpose(perm = input_933_perm_0, x = x_459_cast_fp16)[name = tensor<string, []>("transpose_204")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_935_cast_fp16")];
+            tensor<int32, []> x_461_split_num_splits_0 = const()[name = tensor<string, []>("x_461_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_461_split_axis_0 = const()[name = tensor<string, []>("x_461_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_461_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_461_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_461_split_1_sigmoid_cast_fp16 = sigmoid(x = x_461_split_cast_fp16_1)[name = tensor<string, []>("x_461_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_461_cast_fp16 = mul(x = x_461_split_cast_fp16_0, y = x_461_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_461_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_937_cast_fp16 = select(a = var_44_to_fp16, b = x_461_cast_fp16, cond = var_575)[name = tensor<string, []>("input_937_cast_fp16")];
+            tensor<bool, []> new_x_71_interleave_0 = const()[name = tensor<string, []>("new_x_71_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_71_cast_fp16 = concat(axis = var_59, interleave = new_x_71_interleave_0, values = (cache_71_cast_fp16, input_937_cast_fp16))[name = tensor<string, []>("new_x_71_cast_fp16")];
+            tensor<int32, [3]> var_4209_begin_0 = const()[name = tensor<string, []>("op_4209_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_4209_end_0 = const()[name = tensor<string, []>("op_4209_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_4209_end_mask_0 = const()[name = tensor<string, []>("op_4209_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4209_cast_fp16 = slice_by_index(begin = var_4209_begin_0, end = var_4209_end_0, end_mask = var_4209_end_mask_0, x = new_x_71_cast_fp16)[name = tensor<string, []>("op_4209_cast_fp16")];
+            tensor<string, []> x_463_pad_type_0 = const()[name = tensor<string, []>("x_463_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_463_groups_0 = const()[name = tensor<string, []>("x_463_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_463_strides_0 = const()[name = tensor<string, []>("x_463_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_463_pad_0 = const()[name = tensor<string, []>("x_463_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_463_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355434304))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355443584))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_463_cast_fp16")];
+            tensor<int32, [3]> input_939_perm_0 = const()[name = tensor<string, []>("input_939_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_465_axes_0 = const()[name = tensor<string, []>("x_465_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355445696)))];
+            tensor<fp16, [1024]> encoder_layers_17_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355447808)))];
+            tensor<fp16, [1, 56, 1024]> input_939_cast_fp16 = transpose(perm = input_939_perm_0, x = x_463_cast_fp16)[name = tensor<string, []>("transpose_203")];
+            tensor<fp16, [1, 56, 1024]> x_465_cast_fp16 = layer_norm(axes = x_465_axes_0, beta = encoder_layers_17_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_conv_batch_norm_weight_to_fp16, x = input_939_cast_fp16)[name = tensor<string, []>("x_465_cast_fp16")];
+            tensor<int32, [3]> input_941_perm_0 = const()[name = tensor<string, []>("input_941_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_941_cast_fp16 = transpose(perm = input_941_perm_0, x = x_465_cast_fp16)[name = tensor<string, []>("transpose_202")];
+            tensor<fp16, [1, 1024, 56]> input_943_cast_fp16 = silu(x = input_941_cast_fp16)[name = tensor<string, []>("input_943_cast_fp16")];
+            tensor<string, []> x_467_pad_type_0 = const()[name = tensor<string, []>("x_467_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_467_strides_0 = const()[name = tensor<string, []>("x_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_467_pad_0 = const()[name = tensor<string, []>("x_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_467_dilations_0 = const()[name = tensor<string, []>("x_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_467_groups_0 = const()[name = tensor<string, []>("x_467_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355449920))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356498560))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_467_cast_fp16")];
+            tensor<int32, [3]> input_945_perm_0 = const()[name = tensor<string, []>("input_945_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_945_cast_fp16 = transpose(perm = input_945_perm_0, x = x_467_cast_fp16)[name = tensor<string, []>("transpose_201")];
+            tensor<fp16, [1, 56, 1024]> input_947_cast_fp16 = add(x = input_931_cast_fp16, y = input_945_cast_fp16)[name = tensor<string, []>("input_947_cast_fp16")];
+            tensor<int32, [1]> input_949_axes_0 = const()[name = tensor<string, []>("input_949_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356500672)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356502784)))];
+            tensor<fp16, [1, 56, 1024]> input_949_cast_fp16 = layer_norm(axes = input_949_axes_0, beta = encoder_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_feed_forward2_weight_to_fp16, x = input_947_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356504896))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359650688))), name = tensor<string, []>("encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_17_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359650880)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_161_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_953_cast_fp16 = silu(x = linear_161_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359659136))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362804928))), name = tensor<string, []>("encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_17_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362805120)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_162_cast_fp16")];
+            tensor<fp16, []> var_4252_to_fp16 = const()[name = tensor<string, []>("op_4252_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4253_cast_fp16 = mul(x = linear_162_cast_fp16, y = var_4252_to_fp16)[name = tensor<string, []>("op_4253_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_959_cast_fp16 = add(x = input_947_cast_fp16, y = var_4253_cast_fp16)[name = tensor<string, []>("input_959_cast_fp16")];
+            tensor<int32, [1]> input_961_axes_0 = const()[name = tensor<string, []>("input_961_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362807232)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362809344)))];
+            tensor<fp16, [1, 56, 1024]> input_961_cast_fp16 = layer_norm(axes = input_961_axes_0, beta = encoder_layers_17_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_out_weight_to_fp16, x = input_959_cast_fp16)[name = tensor<string, []>("input_961_cast_fp16")];
+            tensor<int32, [4]> cache_73_begin_0 = const()[name = tensor<string, []>("cache_73_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> cache_73_end_0 = const()[name = tensor<string, []>("cache_73_end_0"), val = tensor<int32, [4]>([19, 1, 42, 1024])];
+            tensor<bool, [4]> cache_73_end_mask_0 = const()[name = tensor<string, []>("cache_73_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_73_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_73_cast_fp16")];
+            tensor<int32, [4]> cache_75_begin_0 = const()[name = tensor<string, []>("cache_75_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> cache_75_end_0 = const()[name = tensor<string, []>("cache_75_end_0"), val = tensor<int32, [4]>([19, 1, 1024, 8])];
+            tensor<bool, [4]> cache_75_end_mask_0 = const()[name = tensor<string, []>("cache_75_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_75_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_75_cast_fp16")];
+            tensor<int32, [1]> input_963_axes_0 = const()[name = tensor<string, []>("input_963_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362811456)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362813568)))];
+            tensor<fp16, [1, 56, 1024]> input_963_cast_fp16 = layer_norm(axes = input_963_axes_0, beta = encoder_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_feed_forward1_weight_to_fp16, x = input_961_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362815680))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(365961472))), name = tensor<string, []>("encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_18_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(365961664)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_163_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_967_cast_fp16 = silu(x = linear_163_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(365969920))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369115712))), name = tensor<string, []>("encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_18_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369115904)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_164_cast_fp16")];
+            tensor<fp16, []> var_4289_to_fp16 = const()[name = tensor<string, []>("op_4289_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4290_cast_fp16 = mul(x = linear_164_cast_fp16, y = var_4289_to_fp16)[name = tensor<string, []>("op_4290_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_973_cast_fp16 = add(x = input_961_cast_fp16, y = var_4290_cast_fp16)[name = tensor<string, []>("input_973_cast_fp16")];
+            tensor<int32, [1]> key_37_axes_0 = const()[name = tensor<string, []>("key_37_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369118016)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369120128)))];
+            tensor<fp16, [1, 56, 1024]> key_37_cast_fp16 = layer_norm(axes = key_37_axes_0, beta = encoder_layers_18_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_self_att_weight_to_fp16, x = input_973_cast_fp16)[name = tensor<string, []>("key_37_cast_fp16")];
+            tensor<bool, []> input_975_interleave_0 = const()[name = tensor<string, []>("input_975_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_975_cast_fp16 = concat(axis = var_68, interleave = input_975_interleave_0, values = (cache_73_cast_fp16, key_37_cast_fp16))[name = tensor<string, []>("input_975_cast_fp16")];
+            tensor<bool, []> var_4318_interleave_0 = const()[name = tensor<string, []>("op_4318_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_4318_cast_fp16 = concat(axis = var_68, interleave = var_4318_interleave_0, values = key_37_cast_fp16)[name = tensor<string, []>("op_4318_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369122240))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369908736))), name = tensor<string, []>("encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369908928)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_165_cast_fp16")];
+            tensor<int32, [4]> var_4323 = const()[name = tensor<string, []>("op_4323"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_109_cast_fp16 = reshape(shape = var_4323, x = linear_165_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369911040))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(370697536))), name = tensor<string, []>("encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(370697728)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_166_cast_fp16")];
+            tensor<int32, [4]> var_4328 = const()[name = tensor<string, []>("op_4328"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_73_cast_fp16 = reshape(shape = var_4328, x = linear_166_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(370699840))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371486336))), name = tensor<string, []>("encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371486528)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_167_cast_fp16")];
+            tensor<int32, [4]> var_4333 = const()[name = tensor<string, []>("op_4333"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_37_cast_fp16 = reshape(shape = var_4333, x = linear_167_cast_fp16)[name = tensor<string, []>("v_37_cast_fp16")];
+            tensor<int32, [4]> value_45_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_18_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371488640)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4346_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_4346_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_18_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371490752)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4348_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_4348_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_37_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_475_transpose_x_0 = const()[name = tensor<string, []>("x_475_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_475_transpose_y_0 = const()[name = tensor<string, []>("x_475_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_4350_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_4350_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371492864))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371692608))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_37_cast_fp16 = transpose(perm = q_with_bias_v_37_perm_0, x = var_4348_cast_fp16)[name = tensor<string, []>("transpose_200")];
+            tensor<fp16, [1, 8, 56, 195]> 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_4350_to_fp16_quantized)[name = tensor<string, []>("x_475_cast_fp16")];
+            tensor<int32, [8]> x_477_pad_0 = const()[name = tensor<string, []>("x_477_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_477_mode_0 = const()[name = tensor<string, []>("x_477_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_313_to_fp16 = const()[name = tensor<string, []>("const_313_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_477_cast_fp16")];
+            tensor<int32, [4]> var_4358 = const()[name = tensor<string, []>("op_4358"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_479_cast_fp16 = reshape(shape = var_4358, x = x_477_cast_fp16)[name = tensor<string, []>("x_479_cast_fp16")];
+            tensor<int32, [4]> var_4362_begin_0 = const()[name = tensor<string, []>("op_4362_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4362_end_0 = const()[name = tensor<string, []>("op_4362_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_4362_end_mask_0 = const()[name = tensor<string, []>("op_4362_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_4362_cast_fp16 = slice_by_index(begin = var_4362_begin_0, end = var_4362_end_0, end_mask = var_4362_end_mask_0, x = x_479_cast_fp16)[name = tensor<string, []>("op_4362_cast_fp16")];
+            tensor<int32, [4]> var_4363 = const()[name = tensor<string, []>("op_4363"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_73_cast_fp16 = reshape(shape = var_4363, x = var_4362_cast_fp16)[name = tensor<string, []>("matrix_bd_73_cast_fp16")];
+            tensor<bool, []> matrix_ac_37_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_37_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_37_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_37_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_132_perm_0 = const()[name = tensor<string, []>("transpose_132_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_133_perm_0 = const()[name = tensor<string, []>("transpose_133_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73_cast_fp16)[name = tensor<string, []>("transpose_198")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_132 = transpose(perm = transpose_132_perm_0, x = var_4346_cast_fp16)[name = tensor<string, []>("transpose_199")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_37_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_75_begin_0 = const()[name = tensor<string, []>("matrix_bd_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_75_end_0 = const()[name = tensor<string, []>("matrix_bd_75_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_75_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_75_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_4372_cast_fp16 = add(x = matrix_ac_37_cast_fp16, y = matrix_bd_75_cast_fp16)[name = tensor<string, []>("op_4372_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_73_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_73_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_73_cast_fp16 = mul(x = var_4372_cast_fp16, y = _inversed_scores_73_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_73_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_75_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_73_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_75_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_4378_cast_fp16 = softmax(axis = var_59, x = scores_75_cast_fp16)[name = tensor<string, []>("op_4378_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_977_cast_fp16 = select(a = var_44_to_fp16, b = var_4378_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_977_cast_fp16")];
+            tensor<bool, []> x_481_transpose_x_0 = const()[name = tensor<string, []>("x_481_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_481_transpose_y_0 = const()[name = tensor<string, []>("x_481_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = v_37_cast_fp16)[name = tensor<string, []>("transpose_197")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_481_cast_fp16")];
+            tensor<int32, [4]> var_4382_perm_0 = const()[name = tensor<string, []>("op_4382_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4383 = const()[name = tensor<string, []>("op_4383"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_4382_cast_fp16 = transpose(perm = var_4382_perm_0, x = x_481_cast_fp16)[name = tensor<string, []>("transpose_196")];
+            tensor<fp16, [1, 56, 1024]> input_979_cast_fp16 = reshape(shape = var_4383, x = var_4382_cast_fp16)[name = tensor<string, []>("input_979_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371693120))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372741760))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372743872)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_169_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_983_cast_fp16 = add(x = input_973_cast_fp16, y = linear_169_cast_fp16)[name = tensor<string, []>("input_983_cast_fp16")];
+            tensor<int32, [1]> x_485_axes_0 = const()[name = tensor<string, []>("x_485_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372745984)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372748096)))];
+            tensor<fp16, [1, 56, 1024]> x_485_cast_fp16 = layer_norm(axes = x_485_axes_0, beta = encoder_layers_18_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_conv_weight_to_fp16, x = input_983_cast_fp16)[name = tensor<string, []>("x_485_cast_fp16")];
+            tensor<int32, [3]> input_985_perm_0 = const()[name = tensor<string, []>("input_985_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_987_pad_type_0 = const()[name = tensor<string, []>("input_987_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_987_strides_0 = const()[name = tensor<string, []>("input_987_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_987_pad_0 = const()[name = tensor<string, []>("input_987_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_987_dilations_0 = const()[name = tensor<string, []>("input_987_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_987_groups_0 = const()[name = tensor<string, []>("input_987_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372750208))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(374847424))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_985_cast_fp16 = transpose(perm = input_985_perm_0, x = x_485_cast_fp16)[name = tensor<string, []>("transpose_195")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_987_cast_fp16")];
+            tensor<int32, []> x_487_split_num_splits_0 = const()[name = tensor<string, []>("x_487_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_487_split_axis_0 = const()[name = tensor<string, []>("x_487_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_487_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_487_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_487_split_1_sigmoid_cast_fp16 = sigmoid(x = x_487_split_cast_fp16_1)[name = tensor<string, []>("x_487_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_487_cast_fp16 = mul(x = x_487_split_cast_fp16_0, y = x_487_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_487_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_989_cast_fp16 = select(a = var_44_to_fp16, b = x_487_cast_fp16, cond = var_575)[name = tensor<string, []>("input_989_cast_fp16")];
+            tensor<bool, []> new_x_75_interleave_0 = const()[name = tensor<string, []>("new_x_75_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_75_cast_fp16 = concat(axis = var_59, interleave = new_x_75_interleave_0, values = (cache_75_cast_fp16, input_989_cast_fp16))[name = tensor<string, []>("new_x_75_cast_fp16")];
+            tensor<int32, [3]> var_4422_begin_0 = const()[name = tensor<string, []>("op_4422_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_4422_end_0 = const()[name = tensor<string, []>("op_4422_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_4422_end_mask_0 = const()[name = tensor<string, []>("op_4422_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4422_cast_fp16 = slice_by_index(begin = var_4422_begin_0, end = var_4422_end_0, end_mask = var_4422_end_mask_0, x = new_x_75_cast_fp16)[name = tensor<string, []>("op_4422_cast_fp16")];
+            tensor<string, []> x_489_pad_type_0 = const()[name = tensor<string, []>("x_489_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_489_groups_0 = const()[name = tensor<string, []>("x_489_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_489_strides_0 = const()[name = tensor<string, []>("x_489_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_489_pad_0 = const()[name = tensor<string, []>("x_489_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_489_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(374851584))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(374860864))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_489_cast_fp16")];
+            tensor<int32, [3]> input_991_perm_0 = const()[name = tensor<string, []>("input_991_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_491_axes_0 = const()[name = tensor<string, []>("x_491_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(374862976)))];
+            tensor<fp16, [1024]> encoder_layers_18_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(374865088)))];
+            tensor<fp16, [1, 56, 1024]> input_991_cast_fp16 = transpose(perm = input_991_perm_0, x = x_489_cast_fp16)[name = tensor<string, []>("transpose_194")];
+            tensor<fp16, [1, 56, 1024]> x_491_cast_fp16 = layer_norm(axes = x_491_axes_0, beta = encoder_layers_18_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_conv_batch_norm_weight_to_fp16, x = input_991_cast_fp16)[name = tensor<string, []>("x_491_cast_fp16")];
+            tensor<int32, [3]> input_993_perm_0 = const()[name = tensor<string, []>("input_993_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_993_cast_fp16 = transpose(perm = input_993_perm_0, x = x_491_cast_fp16)[name = tensor<string, []>("transpose_193")];
+            tensor<fp16, [1, 1024, 56]> input_995_cast_fp16 = silu(x = input_993_cast_fp16)[name = tensor<string, []>("input_995_cast_fp16")];
+            tensor<string, []> x_493_pad_type_0 = const()[name = tensor<string, []>("x_493_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_493_strides_0 = const()[name = tensor<string, []>("x_493_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_493_pad_0 = const()[name = tensor<string, []>("x_493_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_493_dilations_0 = const()[name = tensor<string, []>("x_493_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_493_groups_0 = const()[name = tensor<string, []>("x_493_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(374867200))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(375915840))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_493_cast_fp16")];
+            tensor<int32, [3]> input_997_perm_0 = const()[name = tensor<string, []>("input_997_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_997_cast_fp16 = transpose(perm = input_997_perm_0, x = x_493_cast_fp16)[name = tensor<string, []>("transpose_192")];
+            tensor<fp16, [1, 56, 1024]> input_999_cast_fp16 = add(x = input_983_cast_fp16, y = input_997_cast_fp16)[name = tensor<string, []>("input_999_cast_fp16")];
+            tensor<int32, [1]> input_1001_axes_0 = const()[name = tensor<string, []>("input_1001_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(375917952)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(375920064)))];
+            tensor<fp16, [1, 56, 1024]> input_1001_cast_fp16 = layer_norm(axes = input_1001_axes_0, beta = encoder_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_feed_forward2_weight_to_fp16, x = input_999_cast_fp16)[name = tensor<string, []>("input_1001_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(375922176))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(380116544))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805184)))];
+            tensor<fp16, [4096]> encoder_layers_18_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(380124800)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_170_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1005_cast_fp16 = silu(x = linear_170_cast_fp16)[name = tensor<string, []>("input_1005_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(380133056))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384327424))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_18_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384329536)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_171_cast_fp16")];
+            tensor<fp16, []> var_4465_to_fp16 = const()[name = tensor<string, []>("op_4465_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4466_cast_fp16 = mul(x = linear_171_cast_fp16, y = var_4465_to_fp16)[name = tensor<string, []>("op_4466_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1011_cast_fp16 = add(x = input_999_cast_fp16, y = var_4466_cast_fp16)[name = tensor<string, []>("input_1011_cast_fp16")];
+            tensor<int32, [1]> input_1013_axes_0 = const()[name = tensor<string, []>("input_1013_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384331648)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384333760)))];
+            tensor<fp16, [1, 56, 1024]> input_1013_cast_fp16 = layer_norm(axes = input_1013_axes_0, beta = encoder_layers_18_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_out_weight_to_fp16, x = input_1011_cast_fp16)[name = tensor<string, []>("input_1013_cast_fp16")];
+            tensor<int32, [4]> cache_77_begin_0 = const()[name = tensor<string, []>("cache_77_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> cache_77_end_0 = const()[name = tensor<string, []>("cache_77_end_0"), val = tensor<int32, [4]>([20, 1, 42, 1024])];
+            tensor<bool, [4]> cache_77_end_mask_0 = const()[name = tensor<string, []>("cache_77_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_77_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_77_cast_fp16")];
+            tensor<int32, [4]> cache_79_begin_0 = const()[name = tensor<string, []>("cache_79_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> cache_79_end_0 = const()[name = tensor<string, []>("cache_79_end_0"), val = tensor<int32, [4]>([20, 1, 1024, 8])];
+            tensor<bool, [4]> cache_79_end_mask_0 = const()[name = tensor<string, []>("cache_79_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_79_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_79_cast_fp16")];
+            tensor<int32, [1]> input_1015_axes_0 = const()[name = tensor<string, []>("input_1015_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384335872)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384337984)))];
+            tensor<fp16, [1, 56, 1024]> input_1015_cast_fp16 = layer_norm(axes = input_1015_axes_0, beta = encoder_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_feed_forward1_weight_to_fp16, x = input_1013_cast_fp16)[name = tensor<string, []>("input_1015_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384340096))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(388534464))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805184)))];
+            tensor<fp16, [4096]> encoder_layers_19_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(388542720)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_172_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1019_cast_fp16 = silu(x = linear_172_cast_fp16)[name = tensor<string, []>("input_1019_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(388550976))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(392745344))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_19_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(392747456)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_173_cast_fp16")];
+            tensor<fp16, []> var_4502_to_fp16 = const()[name = tensor<string, []>("op_4502_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4503_cast_fp16 = mul(x = linear_173_cast_fp16, y = var_4502_to_fp16)[name = tensor<string, []>("op_4503_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1025_cast_fp16 = add(x = input_1013_cast_fp16, y = var_4503_cast_fp16)[name = tensor<string, []>("input_1025_cast_fp16")];
+            tensor<int32, [1]> key_39_axes_0 = const()[name = tensor<string, []>("key_39_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(392749568)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(392751680)))];
+            tensor<fp16, [1, 56, 1024]> key_39_cast_fp16 = layer_norm(axes = key_39_axes_0, beta = encoder_layers_19_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_self_att_weight_to_fp16, x = input_1025_cast_fp16)[name = tensor<string, []>("key_39_cast_fp16")];
+            tensor<bool, []> input_1027_interleave_0 = const()[name = tensor<string, []>("input_1027_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_1027_cast_fp16 = concat(axis = var_68, interleave = input_1027_interleave_0, values = (cache_77_cast_fp16, key_39_cast_fp16))[name = tensor<string, []>("input_1027_cast_fp16")];
+            tensor<bool, []> var_4531_interleave_0 = const()[name = tensor<string, []>("op_4531_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_4531_cast_fp16 = concat(axis = var_68, interleave = var_4531_interleave_0, values = key_39_cast_fp16)[name = tensor<string, []>("op_4531_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(392753792))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393802432))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393804544)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_174_cast_fp16")];
+            tensor<int32, [4]> var_4536 = const()[name = tensor<string, []>("op_4536"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_115_cast_fp16 = reshape(shape = var_4536, x = linear_174_cast_fp16)[name = tensor<string, []>("q_115_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393806656))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(394855296))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(394857408)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_175_cast_fp16")];
+            tensor<int32, [4]> var_4541 = const()[name = tensor<string, []>("op_4541"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_77_cast_fp16 = reshape(shape = var_4541, x = linear_175_cast_fp16)[name = tensor<string, []>("k_77_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(394859520))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395908160))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395910272)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_176_cast_fp16")];
+            tensor<int32, [4]> var_4546 = const()[name = tensor<string, []>("op_4546"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_39_cast_fp16 = reshape(shape = var_4546, x = linear_176_cast_fp16)[name = tensor<string, []>("v_39_cast_fp16")];
+            tensor<int32, [4]> value_47_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_19_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395912384)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4559_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_4559_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_19_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395914496)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4561_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_4561_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_39_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_501_transpose_x_0 = const()[name = tensor<string, []>("x_501_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_501_transpose_y_0 = const()[name = tensor<string, []>("x_501_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_4563_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_4563_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395916608))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(396116352))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_39_cast_fp16 = transpose(perm = q_with_bias_v_39_perm_0, x = var_4561_cast_fp16)[name = tensor<string, []>("transpose_191")];
+            tensor<fp16, [1, 8, 56, 195]> 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_4563_to_fp16_quantized)[name = tensor<string, []>("x_501_cast_fp16")];
+            tensor<int32, [8]> x_503_pad_0 = const()[name = tensor<string, []>("x_503_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_503_mode_0 = const()[name = tensor<string, []>("x_503_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_326_to_fp16 = const()[name = tensor<string, []>("const_326_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_503_cast_fp16")];
+            tensor<int32, [4]> var_4571 = const()[name = tensor<string, []>("op_4571"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_505_cast_fp16 = reshape(shape = var_4571, x = x_503_cast_fp16)[name = tensor<string, []>("x_505_cast_fp16")];
+            tensor<int32, [4]> var_4575_begin_0 = const()[name = tensor<string, []>("op_4575_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4575_end_0 = const()[name = tensor<string, []>("op_4575_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_4575_end_mask_0 = const()[name = tensor<string, []>("op_4575_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_4575_cast_fp16 = slice_by_index(begin = var_4575_begin_0, end = var_4575_end_0, end_mask = var_4575_end_mask_0, x = x_505_cast_fp16)[name = tensor<string, []>("op_4575_cast_fp16")];
+            tensor<int32, [4]> var_4576 = const()[name = tensor<string, []>("op_4576"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_77_cast_fp16 = reshape(shape = var_4576, x = var_4575_cast_fp16)[name = tensor<string, []>("matrix_bd_77_cast_fp16")];
+            tensor<bool, []> matrix_ac_39_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_39_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_39_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_39_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_134_perm_0 = const()[name = tensor<string, []>("transpose_134_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_135_perm_0 = const()[name = tensor<string, []>("transpose_135_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77_cast_fp16)[name = tensor<string, []>("transpose_189")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_134 = transpose(perm = transpose_134_perm_0, x = var_4559_cast_fp16)[name = tensor<string, []>("transpose_190")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_39_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_79_begin_0 = const()[name = tensor<string, []>("matrix_bd_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_79_end_0 = const()[name = tensor<string, []>("matrix_bd_79_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_79_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_79_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_4585_cast_fp16 = add(x = matrix_ac_39_cast_fp16, y = matrix_bd_79_cast_fp16)[name = tensor<string, []>("op_4585_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_77_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_77_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_77_cast_fp16 = mul(x = var_4585_cast_fp16, y = _inversed_scores_77_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_77_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_79_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_77_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_79_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_4591_cast_fp16 = softmax(axis = var_59, x = scores_79_cast_fp16)[name = tensor<string, []>("op_4591_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_1029_cast_fp16 = select(a = var_44_to_fp16, b = var_4591_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_1029_cast_fp16")];
+            tensor<bool, []> x_507_transpose_x_0 = const()[name = tensor<string, []>("x_507_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_507_transpose_y_0 = const()[name = tensor<string, []>("x_507_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = v_39_cast_fp16)[name = tensor<string, []>("transpose_188")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_507_cast_fp16")];
+            tensor<int32, [4]> var_4595_perm_0 = const()[name = tensor<string, []>("op_4595_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4596 = const()[name = tensor<string, []>("op_4596"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_4595_cast_fp16 = transpose(perm = var_4595_perm_0, x = x_507_cast_fp16)[name = tensor<string, []>("transpose_187")];
+            tensor<fp16, [1, 56, 1024]> input_1031_cast_fp16 = reshape(shape = var_4596, x = var_4595_cast_fp16)[name = tensor<string, []>("input_1031_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(396116864))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(397165504))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(397167616)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_178_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1035_cast_fp16 = add(x = input_1025_cast_fp16, y = linear_178_cast_fp16)[name = tensor<string, []>("input_1035_cast_fp16")];
+            tensor<int32, [1]> x_511_axes_0 = const()[name = tensor<string, []>("x_511_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(397169728)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(397171840)))];
+            tensor<fp16, [1, 56, 1024]> x_511_cast_fp16 = layer_norm(axes = x_511_axes_0, beta = encoder_layers_19_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_conv_weight_to_fp16, x = input_1035_cast_fp16)[name = tensor<string, []>("x_511_cast_fp16")];
+            tensor<int32, [3]> input_1037_perm_0 = const()[name = tensor<string, []>("input_1037_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_1039_pad_type_0 = const()[name = tensor<string, []>("input_1039_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_1039_strides_0 = const()[name = tensor<string, []>("input_1039_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1039_pad_0 = const()[name = tensor<string, []>("input_1039_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1039_dilations_0 = const()[name = tensor<string, []>("input_1039_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_1039_groups_0 = const()[name = tensor<string, []>("input_1039_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(397173952))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(399271168))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_1037_cast_fp16 = transpose(perm = input_1037_perm_0, x = x_511_cast_fp16)[name = tensor<string, []>("transpose_186")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_1039_cast_fp16")];
+            tensor<int32, []> x_513_split_num_splits_0 = const()[name = tensor<string, []>("x_513_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_513_split_axis_0 = const()[name = tensor<string, []>("x_513_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_513_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_513_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_513_split_1_sigmoid_cast_fp16 = sigmoid(x = x_513_split_cast_fp16_1)[name = tensor<string, []>("x_513_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_513_cast_fp16 = mul(x = x_513_split_cast_fp16_0, y = x_513_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_513_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_1041_cast_fp16 = select(a = var_44_to_fp16, b = x_513_cast_fp16, cond = var_575)[name = tensor<string, []>("input_1041_cast_fp16")];
+            tensor<bool, []> new_x_79_interleave_0 = const()[name = tensor<string, []>("new_x_79_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_79_cast_fp16 = concat(axis = var_59, interleave = new_x_79_interleave_0, values = (cache_79_cast_fp16, input_1041_cast_fp16))[name = tensor<string, []>("new_x_79_cast_fp16")];
+            tensor<int32, [3]> var_4635_begin_0 = const()[name = tensor<string, []>("op_4635_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_4635_end_0 = const()[name = tensor<string, []>("op_4635_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_4635_end_mask_0 = const()[name = tensor<string, []>("op_4635_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4635_cast_fp16 = slice_by_index(begin = var_4635_begin_0, end = var_4635_end_0, end_mask = var_4635_end_mask_0, x = new_x_79_cast_fp16)[name = tensor<string, []>("op_4635_cast_fp16")];
+            tensor<string, []> x_515_pad_type_0 = const()[name = tensor<string, []>("x_515_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_515_groups_0 = const()[name = tensor<string, []>("x_515_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_515_strides_0 = const()[name = tensor<string, []>("x_515_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_515_pad_0 = const()[name = tensor<string, []>("x_515_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_515_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(399275328))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(399284608))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_515_cast_fp16")];
+            tensor<int32, [3]> input_1043_perm_0 = const()[name = tensor<string, []>("input_1043_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_517_axes_0 = const()[name = tensor<string, []>("x_517_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(399286720)))];
+            tensor<fp16, [1024]> encoder_layers_19_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(399288832)))];
+            tensor<fp16, [1, 56, 1024]> input_1043_cast_fp16 = transpose(perm = input_1043_perm_0, x = x_515_cast_fp16)[name = tensor<string, []>("transpose_185")];
+            tensor<fp16, [1, 56, 1024]> x_517_cast_fp16 = layer_norm(axes = x_517_axes_0, beta = encoder_layers_19_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_conv_batch_norm_weight_to_fp16, x = input_1043_cast_fp16)[name = tensor<string, []>("x_517_cast_fp16")];
+            tensor<int32, [3]> input_1045_perm_0 = const()[name = tensor<string, []>("input_1045_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_1045_cast_fp16 = transpose(perm = input_1045_perm_0, x = x_517_cast_fp16)[name = tensor<string, []>("transpose_184")];
+            tensor<fp16, [1, 1024, 56]> input_1047_cast_fp16 = silu(x = input_1045_cast_fp16)[name = tensor<string, []>("input_1047_cast_fp16")];
+            tensor<string, []> x_519_pad_type_0 = const()[name = tensor<string, []>("x_519_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_519_strides_0 = const()[name = tensor<string, []>("x_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_519_pad_0 = const()[name = tensor<string, []>("x_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_519_dilations_0 = const()[name = tensor<string, []>("x_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_519_groups_0 = const()[name = tensor<string, []>("x_519_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(399290944))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(400339584))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_519_cast_fp16")];
+            tensor<int32, [3]> input_1049_perm_0 = const()[name = tensor<string, []>("input_1049_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_1049_cast_fp16 = transpose(perm = input_1049_perm_0, x = x_519_cast_fp16)[name = tensor<string, []>("transpose_183")];
+            tensor<fp16, [1, 56, 1024]> input_1051_cast_fp16 = add(x = input_1035_cast_fp16, y = input_1049_cast_fp16)[name = tensor<string, []>("input_1051_cast_fp16")];
+            tensor<int32, [1]> input_1053_axes_0 = const()[name = tensor<string, []>("input_1053_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(400341696)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(400343808)))];
+            tensor<fp16, [1, 56, 1024]> input_1053_cast_fp16 = layer_norm(axes = input_1053_axes_0, beta = encoder_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_feed_forward2_weight_to_fp16, x = input_1051_cast_fp16)[name = tensor<string, []>("input_1053_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(400345920))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(404540288))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805184)))];
+            tensor<fp16, [4096]> encoder_layers_19_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(404548544)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_179_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1057_cast_fp16 = silu(x = linear_179_cast_fp16)[name = tensor<string, []>("input_1057_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(404556800))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408751168))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_19_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408753280)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_180_cast_fp16")];
+            tensor<fp16, []> var_4678_to_fp16 = const()[name = tensor<string, []>("op_4678_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4679_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_4678_to_fp16)[name = tensor<string, []>("op_4679_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1063_cast_fp16 = add(x = input_1051_cast_fp16, y = var_4679_cast_fp16)[name = tensor<string, []>("input_1063_cast_fp16")];
+            tensor<int32, [1]> input_1065_axes_0 = const()[name = tensor<string, []>("input_1065_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408755392)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408757504)))];
+            tensor<fp16, [1, 56, 1024]> input_1065_cast_fp16 = layer_norm(axes = input_1065_axes_0, beta = encoder_layers_19_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_out_weight_to_fp16, x = input_1063_cast_fp16)[name = tensor<string, []>("input_1065_cast_fp16")];
+            tensor<int32, [4]> cache_81_begin_0 = const()[name = tensor<string, []>("cache_81_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> cache_81_end_0 = const()[name = tensor<string, []>("cache_81_end_0"), val = tensor<int32, [4]>([21, 1, 42, 1024])];
+            tensor<bool, [4]> cache_81_end_mask_0 = const()[name = tensor<string, []>("cache_81_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_81_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_81_cast_fp16")];
+            tensor<int32, [4]> cache_83_begin_0 = const()[name = tensor<string, []>("cache_83_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> cache_83_end_0 = const()[name = tensor<string, []>("cache_83_end_0"), val = tensor<int32, [4]>([21, 1, 1024, 8])];
+            tensor<bool, [4]> cache_83_end_mask_0 = const()[name = tensor<string, []>("cache_83_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_83_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_83_cast_fp16")];
+            tensor<int32, [1]> input_1067_axes_0 = const()[name = tensor<string, []>("input_1067_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408759616)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408761728)))];
+            tensor<fp16, [1, 56, 1024]> input_1067_cast_fp16 = layer_norm(axes = input_1067_axes_0, beta = encoder_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_feed_forward1_weight_to_fp16, x = input_1065_cast_fp16)[name = tensor<string, []>("input_1067_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408763840))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(412958208))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805184)))];
+            tensor<fp16, [4096]> encoder_layers_20_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(412966464)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_181_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1071_cast_fp16 = silu(x = linear_181_cast_fp16)[name = tensor<string, []>("input_1071_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(412974720))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417169088))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_20_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417171200)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_182_cast_fp16")];
+            tensor<fp16, []> var_4715_to_fp16 = const()[name = tensor<string, []>("op_4715_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4716_cast_fp16 = mul(x = linear_182_cast_fp16, y = var_4715_to_fp16)[name = tensor<string, []>("op_4716_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1077_cast_fp16 = add(x = input_1065_cast_fp16, y = var_4716_cast_fp16)[name = tensor<string, []>("input_1077_cast_fp16")];
+            tensor<int32, [1]> key_41_axes_0 = const()[name = tensor<string, []>("key_41_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417173312)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417175424)))];
+            tensor<fp16, [1, 56, 1024]> key_41_cast_fp16 = layer_norm(axes = key_41_axes_0, beta = encoder_layers_20_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_self_att_weight_to_fp16, x = input_1077_cast_fp16)[name = tensor<string, []>("key_41_cast_fp16")];
+            tensor<bool, []> input_1079_interleave_0 = const()[name = tensor<string, []>("input_1079_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_1079_cast_fp16 = concat(axis = var_68, interleave = input_1079_interleave_0, values = (cache_81_cast_fp16, key_41_cast_fp16))[name = tensor<string, []>("input_1079_cast_fp16")];
+            tensor<bool, []> var_4744_interleave_0 = const()[name = tensor<string, []>("op_4744_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_4744_cast_fp16 = concat(axis = var_68, interleave = var_4744_interleave_0, values = key_41_cast_fp16)[name = tensor<string, []>("op_4744_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417177536))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(418226176))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(418228288)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_183_cast_fp16")];
+            tensor<int32, [4]> var_4749 = const()[name = tensor<string, []>("op_4749"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_121_cast_fp16 = reshape(shape = var_4749, x = linear_183_cast_fp16)[name = tensor<string, []>("q_121_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(418230400))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419279040))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419281152)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_184_cast_fp16")];
+            tensor<int32, [4]> var_4754 = const()[name = tensor<string, []>("op_4754"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_81_cast_fp16 = reshape(shape = var_4754, x = linear_184_cast_fp16)[name = tensor<string, []>("k_81_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419283264))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420331904))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420334016)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_185_cast_fp16")];
+            tensor<int32, [4]> var_4759 = const()[name = tensor<string, []>("op_4759"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_41_cast_fp16 = reshape(shape = var_4759, x = linear_185_cast_fp16)[name = tensor<string, []>("v_41_cast_fp16")];
+            tensor<int32, [4]> value_49_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_20_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420336128)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4772_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_4772_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_20_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420338240)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4774_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_4774_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_41_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_527_transpose_x_0 = const()[name = tensor<string, []>("x_527_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_527_transpose_y_0 = const()[name = tensor<string, []>("x_527_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_4776_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_4776_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420340352))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420540096))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_41_cast_fp16 = transpose(perm = q_with_bias_v_41_perm_0, x = var_4774_cast_fp16)[name = tensor<string, []>("transpose_182")];
+            tensor<fp16, [1, 8, 56, 195]> 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_4776_to_fp16_quantized)[name = tensor<string, []>("x_527_cast_fp16")];
+            tensor<int32, [8]> x_529_pad_0 = const()[name = tensor<string, []>("x_529_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_529_mode_0 = const()[name = tensor<string, []>("x_529_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_339_to_fp16 = const()[name = tensor<string, []>("const_339_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_529_cast_fp16")];
+            tensor<int32, [4]> var_4784 = const()[name = tensor<string, []>("op_4784"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_531_cast_fp16 = reshape(shape = var_4784, x = x_529_cast_fp16)[name = tensor<string, []>("x_531_cast_fp16")];
+            tensor<int32, [4]> var_4788_begin_0 = const()[name = tensor<string, []>("op_4788_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4788_end_0 = const()[name = tensor<string, []>("op_4788_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_4788_end_mask_0 = const()[name = tensor<string, []>("op_4788_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_4788_cast_fp16 = slice_by_index(begin = var_4788_begin_0, end = var_4788_end_0, end_mask = var_4788_end_mask_0, x = x_531_cast_fp16)[name = tensor<string, []>("op_4788_cast_fp16")];
+            tensor<int32, [4]> var_4789 = const()[name = tensor<string, []>("op_4789"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_81_cast_fp16 = reshape(shape = var_4789, x = var_4788_cast_fp16)[name = tensor<string, []>("matrix_bd_81_cast_fp16")];
+            tensor<bool, []> matrix_ac_41_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_41_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_41_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_41_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_136_perm_0 = const()[name = tensor<string, []>("transpose_136_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_137_perm_0 = const()[name = tensor<string, []>("transpose_137_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81_cast_fp16)[name = tensor<string, []>("transpose_180")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_136 = transpose(perm = transpose_136_perm_0, x = var_4772_cast_fp16)[name = tensor<string, []>("transpose_181")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_41_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_83_begin_0 = const()[name = tensor<string, []>("matrix_bd_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_83_end_0 = const()[name = tensor<string, []>("matrix_bd_83_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_83_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_83_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_4798_cast_fp16 = add(x = matrix_ac_41_cast_fp16, y = matrix_bd_83_cast_fp16)[name = tensor<string, []>("op_4798_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_81_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_81_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_81_cast_fp16 = mul(x = var_4798_cast_fp16, y = _inversed_scores_81_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_81_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_83_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_81_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_83_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_4804_cast_fp16 = softmax(axis = var_59, x = scores_83_cast_fp16)[name = tensor<string, []>("op_4804_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_1081_cast_fp16 = select(a = var_44_to_fp16, b = var_4804_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_1081_cast_fp16")];
+            tensor<bool, []> x_533_transpose_x_0 = const()[name = tensor<string, []>("x_533_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_533_transpose_y_0 = const()[name = tensor<string, []>("x_533_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = v_41_cast_fp16)[name = tensor<string, []>("transpose_179")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_533_cast_fp16")];
+            tensor<int32, [4]> var_4808_perm_0 = const()[name = tensor<string, []>("op_4808_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4809 = const()[name = tensor<string, []>("op_4809"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_4808_cast_fp16 = transpose(perm = var_4808_perm_0, x = x_533_cast_fp16)[name = tensor<string, []>("transpose_178")];
+            tensor<fp16, [1, 56, 1024]> input_1083_cast_fp16 = reshape(shape = var_4809, x = var_4808_cast_fp16)[name = tensor<string, []>("input_1083_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420540608))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421589248))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421591360)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_187_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1087_cast_fp16 = add(x = input_1077_cast_fp16, y = linear_187_cast_fp16)[name = tensor<string, []>("input_1087_cast_fp16")];
+            tensor<int32, [1]> x_537_axes_0 = const()[name = tensor<string, []>("x_537_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421593472)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421595584)))];
+            tensor<fp16, [1, 56, 1024]> x_537_cast_fp16 = layer_norm(axes = x_537_axes_0, beta = encoder_layers_20_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_conv_weight_to_fp16, x = input_1087_cast_fp16)[name = tensor<string, []>("x_537_cast_fp16")];
+            tensor<int32, [3]> input_1089_perm_0 = const()[name = tensor<string, []>("input_1089_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_1091_pad_type_0 = const()[name = tensor<string, []>("input_1091_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_1091_strides_0 = const()[name = tensor<string, []>("input_1091_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1091_pad_0 = const()[name = tensor<string, []>("input_1091_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1091_dilations_0 = const()[name = tensor<string, []>("input_1091_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_1091_groups_0 = const()[name = tensor<string, []>("input_1091_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421597696))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(423694912))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_1089_cast_fp16 = transpose(perm = input_1089_perm_0, x = x_537_cast_fp16)[name = tensor<string, []>("transpose_177")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_1091_cast_fp16")];
+            tensor<int32, []> x_539_split_num_splits_0 = const()[name = tensor<string, []>("x_539_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_539_split_axis_0 = const()[name = tensor<string, []>("x_539_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_539_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_539_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_539_split_1_sigmoid_cast_fp16 = sigmoid(x = x_539_split_cast_fp16_1)[name = tensor<string, []>("x_539_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_539_cast_fp16 = mul(x = x_539_split_cast_fp16_0, y = x_539_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_539_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_1093_cast_fp16 = select(a = var_44_to_fp16, b = x_539_cast_fp16, cond = var_575)[name = tensor<string, []>("input_1093_cast_fp16")];
+            tensor<bool, []> new_x_83_interleave_0 = const()[name = tensor<string, []>("new_x_83_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_83_cast_fp16 = concat(axis = var_59, interleave = new_x_83_interleave_0, values = (cache_83_cast_fp16, input_1093_cast_fp16))[name = tensor<string, []>("new_x_83_cast_fp16")];
+            tensor<int32, [3]> var_4848_begin_0 = const()[name = tensor<string, []>("op_4848_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_4848_end_0 = const()[name = tensor<string, []>("op_4848_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_4848_end_mask_0 = const()[name = tensor<string, []>("op_4848_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4848_cast_fp16 = slice_by_index(begin = var_4848_begin_0, end = var_4848_end_0, end_mask = var_4848_end_mask_0, x = new_x_83_cast_fp16)[name = tensor<string, []>("op_4848_cast_fp16")];
+            tensor<string, []> x_541_pad_type_0 = const()[name = tensor<string, []>("x_541_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_541_groups_0 = const()[name = tensor<string, []>("x_541_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_541_strides_0 = const()[name = tensor<string, []>("x_541_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_541_pad_0 = const()[name = tensor<string, []>("x_541_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_541_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(423699072))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(423708352))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_541_cast_fp16")];
+            tensor<int32, [3]> input_1095_perm_0 = const()[name = tensor<string, []>("input_1095_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_543_axes_0 = const()[name = tensor<string, []>("x_543_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(423710464)))];
+            tensor<fp16, [1024]> encoder_layers_20_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(423712576)))];
+            tensor<fp16, [1, 56, 1024]> input_1095_cast_fp16 = transpose(perm = input_1095_perm_0, x = x_541_cast_fp16)[name = tensor<string, []>("transpose_176")];
+            tensor<fp16, [1, 56, 1024]> x_543_cast_fp16 = layer_norm(axes = x_543_axes_0, beta = encoder_layers_20_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_conv_batch_norm_weight_to_fp16, x = input_1095_cast_fp16)[name = tensor<string, []>("x_543_cast_fp16")];
+            tensor<int32, [3]> input_1097_perm_0 = const()[name = tensor<string, []>("input_1097_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_1097_cast_fp16 = transpose(perm = input_1097_perm_0, x = x_543_cast_fp16)[name = tensor<string, []>("transpose_175")];
+            tensor<fp16, [1, 1024, 56]> input_1099_cast_fp16 = silu(x = input_1097_cast_fp16)[name = tensor<string, []>("input_1099_cast_fp16")];
+            tensor<string, []> x_545_pad_type_0 = const()[name = tensor<string, []>("x_545_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_545_strides_0 = const()[name = tensor<string, []>("x_545_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_545_pad_0 = const()[name = tensor<string, []>("x_545_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_545_dilations_0 = const()[name = tensor<string, []>("x_545_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_545_groups_0 = const()[name = tensor<string, []>("x_545_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(423714688))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(424763328))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_545_cast_fp16")];
+            tensor<int32, [3]> input_1101_perm_0 = const()[name = tensor<string, []>("input_1101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_1101_cast_fp16 = transpose(perm = input_1101_perm_0, x = x_545_cast_fp16)[name = tensor<string, []>("transpose_174")];
+            tensor<fp16, [1, 56, 1024]> input_1103_cast_fp16 = add(x = input_1087_cast_fp16, y = input_1101_cast_fp16)[name = tensor<string, []>("input_1103_cast_fp16")];
+            tensor<int32, [1]> input_1105_axes_0 = const()[name = tensor<string, []>("input_1105_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(424765440)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(424767552)))];
+            tensor<fp16, [1, 56, 1024]> input_1105_cast_fp16 = layer_norm(axes = input_1105_axes_0, beta = encoder_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_feed_forward2_weight_to_fp16, x = input_1103_cast_fp16)[name = tensor<string, []>("input_1105_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(424769664))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(428964032))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805184)))];
+            tensor<fp16, [4096]> encoder_layers_20_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(428972288)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_188_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1109_cast_fp16 = silu(x = linear_188_cast_fp16)[name = tensor<string, []>("input_1109_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(428980544))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(433174912))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_20_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(433177024)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_189_cast_fp16")];
+            tensor<fp16, []> var_4891_to_fp16 = const()[name = tensor<string, []>("op_4891_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4892_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_4891_to_fp16)[name = tensor<string, []>("op_4892_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1115_cast_fp16 = add(x = input_1103_cast_fp16, y = var_4892_cast_fp16)[name = tensor<string, []>("input_1115_cast_fp16")];
+            tensor<int32, [1]> input_1117_axes_0 = const()[name = tensor<string, []>("input_1117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(433179136)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(433181248)))];
+            tensor<fp16, [1, 56, 1024]> input_1117_cast_fp16 = layer_norm(axes = input_1117_axes_0, beta = encoder_layers_20_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_out_weight_to_fp16, x = input_1115_cast_fp16)[name = tensor<string, []>("input_1117_cast_fp16")];
+            tensor<int32, [4]> cache_85_begin_0 = const()[name = tensor<string, []>("cache_85_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> cache_85_end_0 = const()[name = tensor<string, []>("cache_85_end_0"), val = tensor<int32, [4]>([22, 1, 42, 1024])];
+            tensor<bool, [4]> cache_85_end_mask_0 = const()[name = tensor<string, []>("cache_85_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_85_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_85_cast_fp16")];
+            tensor<int32, [4]> cache_87_begin_0 = const()[name = tensor<string, []>("cache_87_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> cache_87_end_0 = const()[name = tensor<string, []>("cache_87_end_0"), val = tensor<int32, [4]>([22, 1, 1024, 8])];
+            tensor<bool, [4]> cache_87_end_mask_0 = const()[name = tensor<string, []>("cache_87_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_87_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_87_cast_fp16")];
+            tensor<int32, [1]> input_1119_axes_0 = const()[name = tensor<string, []>("input_1119_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(433183360)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(433185472)))];
+            tensor<fp16, [1, 56, 1024]> input_1119_cast_fp16 = layer_norm(axes = input_1119_axes_0, beta = encoder_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_feed_forward1_weight_to_fp16, x = input_1117_cast_fp16)[name = tensor<string, []>("input_1119_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(433187584))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(437381952))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805184)))];
+            tensor<fp16, [4096]> encoder_layers_21_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(437390208)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_190_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1123_cast_fp16 = silu(x = linear_190_cast_fp16)[name = tensor<string, []>("input_1123_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(437398464))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(441592832))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_21_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(441594944)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_191_cast_fp16")];
+            tensor<fp16, []> var_4928_to_fp16 = const()[name = tensor<string, []>("op_4928_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4929_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_4928_to_fp16)[name = tensor<string, []>("op_4929_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1129_cast_fp16 = add(x = input_1117_cast_fp16, y = var_4929_cast_fp16)[name = tensor<string, []>("input_1129_cast_fp16")];
+            tensor<int32, [1]> key_43_axes_0 = const()[name = tensor<string, []>("key_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(441597056)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(441599168)))];
+            tensor<fp16, [1, 56, 1024]> key_43_cast_fp16 = layer_norm(axes = key_43_axes_0, beta = encoder_layers_21_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_self_att_weight_to_fp16, x = input_1129_cast_fp16)[name = tensor<string, []>("key_43_cast_fp16")];
+            tensor<bool, []> input_1131_interleave_0 = const()[name = tensor<string, []>("input_1131_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_1131_cast_fp16 = concat(axis = var_68, interleave = input_1131_interleave_0, values = (cache_85_cast_fp16, key_43_cast_fp16))[name = tensor<string, []>("input_1131_cast_fp16")];
+            tensor<bool, []> var_4957_interleave_0 = const()[name = tensor<string, []>("op_4957_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_4957_cast_fp16 = concat(axis = var_68, interleave = var_4957_interleave_0, values = key_43_cast_fp16)[name = tensor<string, []>("op_4957_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(441601280))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442649920))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442652032)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_192_cast_fp16")];
+            tensor<int32, [4]> var_4962 = const()[name = tensor<string, []>("op_4962"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_127_cast_fp16 = reshape(shape = var_4962, x = linear_192_cast_fp16)[name = tensor<string, []>("q_127_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442654144))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(443702784))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(443704896)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_193_cast_fp16")];
+            tensor<int32, [4]> var_4967 = const()[name = tensor<string, []>("op_4967"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_85_cast_fp16 = reshape(shape = var_4967, x = linear_193_cast_fp16)[name = tensor<string, []>("k_85_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(443707008))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444755648))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444757760)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_194_cast_fp16")];
+            tensor<int32, [4]> var_4972 = const()[name = tensor<string, []>("op_4972"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_43_cast_fp16 = reshape(shape = var_4972, x = linear_194_cast_fp16)[name = tensor<string, []>("v_43_cast_fp16")];
+            tensor<int32, [4]> value_51_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_21_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444759872)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4985_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_4985_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_21_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444761984)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4987_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_4987_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_43_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_553_transpose_x_0 = const()[name = tensor<string, []>("x_553_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_553_transpose_y_0 = const()[name = tensor<string, []>("x_553_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_4989_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_4989_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444764096))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444963840))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_43_cast_fp16 = transpose(perm = q_with_bias_v_43_perm_0, x = var_4987_cast_fp16)[name = tensor<string, []>("transpose_173")];
+            tensor<fp16, [1, 8, 56, 195]> 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_4989_to_fp16_quantized)[name = tensor<string, []>("x_553_cast_fp16")];
+            tensor<int32, [8]> x_555_pad_0 = const()[name = tensor<string, []>("x_555_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_555_mode_0 = const()[name = tensor<string, []>("x_555_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_352_to_fp16 = const()[name = tensor<string, []>("const_352_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_555_cast_fp16")];
+            tensor<int32, [4]> var_4997 = const()[name = tensor<string, []>("op_4997"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_557_cast_fp16 = reshape(shape = var_4997, x = x_555_cast_fp16)[name = tensor<string, []>("x_557_cast_fp16")];
+            tensor<int32, [4]> var_5001_begin_0 = const()[name = tensor<string, []>("op_5001_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5001_end_0 = const()[name = tensor<string, []>("op_5001_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_5001_end_mask_0 = const()[name = tensor<string, []>("op_5001_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_5001_cast_fp16 = slice_by_index(begin = var_5001_begin_0, end = var_5001_end_0, end_mask = var_5001_end_mask_0, x = x_557_cast_fp16)[name = tensor<string, []>("op_5001_cast_fp16")];
+            tensor<int32, [4]> var_5002 = const()[name = tensor<string, []>("op_5002"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_85_cast_fp16 = reshape(shape = var_5002, x = var_5001_cast_fp16)[name = tensor<string, []>("matrix_bd_85_cast_fp16")];
+            tensor<bool, []> matrix_ac_43_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_43_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_43_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_43_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_138_perm_0 = const()[name = tensor<string, []>("transpose_138_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_139_perm_0 = const()[name = tensor<string, []>("transpose_139_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85_cast_fp16)[name = tensor<string, []>("transpose_171")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_138 = transpose(perm = transpose_138_perm_0, x = var_4985_cast_fp16)[name = tensor<string, []>("transpose_172")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_43_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_87_begin_0 = const()[name = tensor<string, []>("matrix_bd_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_87_end_0 = const()[name = tensor<string, []>("matrix_bd_87_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_87_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_87_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_5011_cast_fp16 = add(x = matrix_ac_43_cast_fp16, y = matrix_bd_87_cast_fp16)[name = tensor<string, []>("op_5011_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_85_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_85_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_85_cast_fp16 = mul(x = var_5011_cast_fp16, y = _inversed_scores_85_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_85_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_87_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_85_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_87_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_5017_cast_fp16 = softmax(axis = var_59, x = scores_87_cast_fp16)[name = tensor<string, []>("op_5017_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_1133_cast_fp16 = select(a = var_44_to_fp16, b = var_5017_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_1133_cast_fp16")];
+            tensor<bool, []> x_559_transpose_x_0 = const()[name = tensor<string, []>("x_559_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_559_transpose_y_0 = const()[name = tensor<string, []>("x_559_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = v_43_cast_fp16)[name = tensor<string, []>("transpose_170")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_559_cast_fp16")];
+            tensor<int32, [4]> var_5021_perm_0 = const()[name = tensor<string, []>("op_5021_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5022 = const()[name = tensor<string, []>("op_5022"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_5021_cast_fp16 = transpose(perm = var_5021_perm_0, x = x_559_cast_fp16)[name = tensor<string, []>("transpose_169")];
+            tensor<fp16, [1, 56, 1024]> input_1135_cast_fp16 = reshape(shape = var_5022, x = var_5021_cast_fp16)[name = tensor<string, []>("input_1135_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444964352)))];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(447061568)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_196_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1139_cast_fp16 = add(x = input_1129_cast_fp16, y = linear_196_cast_fp16)[name = tensor<string, []>("input_1139_cast_fp16")];
+            tensor<int32, [1]> x_563_axes_0 = const()[name = tensor<string, []>("x_563_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(447063680)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(447065792)))];
+            tensor<fp16, [1, 56, 1024]> x_563_cast_fp16 = layer_norm(axes = x_563_axes_0, beta = encoder_layers_21_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_conv_weight_to_fp16, x = input_1139_cast_fp16)[name = tensor<string, []>("x_563_cast_fp16")];
+            tensor<int32, [3]> input_1141_perm_0 = const()[name = tensor<string, []>("input_1141_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_1143_pad_type_0 = const()[name = tensor<string, []>("input_1143_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_1143_strides_0 = const()[name = tensor<string, []>("input_1143_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1143_pad_0 = const()[name = tensor<string, []>("input_1143_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1143_dilations_0 = const()[name = tensor<string, []>("input_1143_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_1143_groups_0 = const()[name = tensor<string, []>("input_1143_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(447067904))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(449165120))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_1141_cast_fp16 = transpose(perm = input_1141_perm_0, x = x_563_cast_fp16)[name = tensor<string, []>("transpose_168")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_1143_cast_fp16")];
+            tensor<int32, []> x_565_split_num_splits_0 = const()[name = tensor<string, []>("x_565_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_565_split_axis_0 = const()[name = tensor<string, []>("x_565_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_565_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_565_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_565_split_1_sigmoid_cast_fp16 = sigmoid(x = x_565_split_cast_fp16_1)[name = tensor<string, []>("x_565_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_565_cast_fp16 = mul(x = x_565_split_cast_fp16_0, y = x_565_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_565_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_1145_cast_fp16 = select(a = var_44_to_fp16, b = x_565_cast_fp16, cond = var_575)[name = tensor<string, []>("input_1145_cast_fp16")];
+            tensor<bool, []> new_x_87_interleave_0 = const()[name = tensor<string, []>("new_x_87_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_87_cast_fp16 = concat(axis = var_59, interleave = new_x_87_interleave_0, values = (cache_87_cast_fp16, input_1145_cast_fp16))[name = tensor<string, []>("new_x_87_cast_fp16")];
+            tensor<int32, [3]> var_5061_begin_0 = const()[name = tensor<string, []>("op_5061_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_5061_end_0 = const()[name = tensor<string, []>("op_5061_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_5061_end_mask_0 = const()[name = tensor<string, []>("op_5061_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_5061_cast_fp16 = slice_by_index(begin = var_5061_begin_0, end = var_5061_end_0, end_mask = var_5061_end_mask_0, x = new_x_87_cast_fp16)[name = tensor<string, []>("op_5061_cast_fp16")];
+            tensor<string, []> x_567_pad_type_0 = const()[name = tensor<string, []>("x_567_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_567_groups_0 = const()[name = tensor<string, []>("x_567_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_567_strides_0 = const()[name = tensor<string, []>("x_567_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_567_pad_0 = const()[name = tensor<string, []>("x_567_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_567_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(449169280))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(449178560))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_567_cast_fp16")];
+            tensor<int32, [3]> input_1147_perm_0 = const()[name = tensor<string, []>("input_1147_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_569_axes_0 = const()[name = tensor<string, []>("x_569_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(449180672)))];
+            tensor<fp16, [1024]> encoder_layers_21_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(449182784)))];
+            tensor<fp16, [1, 56, 1024]> input_1147_cast_fp16 = transpose(perm = input_1147_perm_0, x = x_567_cast_fp16)[name = tensor<string, []>("transpose_167")];
+            tensor<fp16, [1, 56, 1024]> x_569_cast_fp16 = layer_norm(axes = x_569_axes_0, beta = encoder_layers_21_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_conv_batch_norm_weight_to_fp16, x = input_1147_cast_fp16)[name = tensor<string, []>("x_569_cast_fp16")];
+            tensor<int32, [3]> input_1149_perm_0 = const()[name = tensor<string, []>("input_1149_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_1149_cast_fp16 = transpose(perm = input_1149_perm_0, x = x_569_cast_fp16)[name = tensor<string, []>("transpose_166")];
+            tensor<fp16, [1, 1024, 56]> input_1151_cast_fp16 = silu(x = input_1149_cast_fp16)[name = tensor<string, []>("input_1151_cast_fp16")];
+            tensor<string, []> x_571_pad_type_0 = const()[name = tensor<string, []>("x_571_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_571_strides_0 = const()[name = tensor<string, []>("x_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_571_pad_0 = const()[name = tensor<string, []>("x_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_571_dilations_0 = const()[name = tensor<string, []>("x_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_571_groups_0 = const()[name = tensor<string, []>("x_571_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(449184896))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(450233536))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_571_cast_fp16")];
+            tensor<int32, [3]> input_1153_perm_0 = const()[name = tensor<string, []>("input_1153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_1153_cast_fp16 = transpose(perm = input_1153_perm_0, x = x_571_cast_fp16)[name = tensor<string, []>("transpose_165")];
+            tensor<fp16, [1, 56, 1024]> input_1155_cast_fp16 = add(x = input_1139_cast_fp16, y = input_1153_cast_fp16)[name = tensor<string, []>("input_1155_cast_fp16")];
+            tensor<int32, [1]> input_1157_axes_0 = const()[name = tensor<string, []>("input_1157_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(450235648)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(450237760)))];
+            tensor<fp16, [1, 56, 1024]> input_1157_cast_fp16 = layer_norm(axes = input_1157_axes_0, beta = encoder_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_feed_forward2_weight_to_fp16, x = input_1155_cast_fp16)[name = tensor<string, []>("input_1157_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_21_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(450239872)))];
+            tensor<fp16, [4096]> encoder_layers_21_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(458628544)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_197_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1161_cast_fp16 = silu(x = linear_197_cast_fp16)[name = tensor<string, []>("input_1161_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_21_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(458636800)))];
+            tensor<fp16, [1024]> encoder_layers_21_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(467025472)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_198_cast_fp16")];
+            tensor<fp16, []> var_5104_to_fp16 = const()[name = tensor<string, []>("op_5104_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_5105_cast_fp16 = mul(x = linear_198_cast_fp16, y = var_5104_to_fp16)[name = tensor<string, []>("op_5105_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1167_cast_fp16 = add(x = input_1155_cast_fp16, y = var_5105_cast_fp16)[name = tensor<string, []>("input_1167_cast_fp16")];
+            tensor<int32, [1]> input_1169_axes_0 = const()[name = tensor<string, []>("input_1169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(467027584)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(467029696)))];
+            tensor<fp16, [1, 56, 1024]> input_1169_cast_fp16 = layer_norm(axes = input_1169_axes_0, beta = encoder_layers_21_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_out_weight_to_fp16, x = input_1167_cast_fp16)[name = tensor<string, []>("input_1169_cast_fp16")];
+            tensor<int32, [4]> cache_89_begin_0 = const()[name = tensor<string, []>("cache_89_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> cache_89_end_0 = const()[name = tensor<string, []>("cache_89_end_0"), val = tensor<int32, [4]>([23, 1, 42, 1024])];
+            tensor<bool, [4]> cache_89_end_mask_0 = const()[name = tensor<string, []>("cache_89_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_89_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_89_cast_fp16")];
+            tensor<int32, [4]> cache_91_begin_0 = const()[name = tensor<string, []>("cache_91_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> cache_91_end_0 = const()[name = tensor<string, []>("cache_91_end_0"), val = tensor<int32, [4]>([23, 1, 1024, 8])];
+            tensor<bool, [4]> cache_91_end_mask_0 = const()[name = tensor<string, []>("cache_91_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_91_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_91_cast_fp16")];
+            tensor<int32, [1]> input_1171_axes_0 = const()[name = tensor<string, []>("input_1171_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(467031808)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(467033920)))];
+            tensor<fp16, [1, 56, 1024]> input_1171_cast_fp16 = layer_norm(axes = input_1171_axes_0, beta = encoder_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_feed_forward1_weight_to_fp16, x = input_1169_cast_fp16)[name = tensor<string, []>("input_1171_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_22_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(467036032)))];
+            tensor<fp16, [4096]> encoder_layers_22_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(475424704)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_199_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1175_cast_fp16 = silu(x = linear_199_cast_fp16)[name = tensor<string, []>("input_1175_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_22_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(475432960)))];
+            tensor<fp16, [1024]> encoder_layers_22_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483821632)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_200_cast_fp16")];
+            tensor<fp16, []> var_5141_to_fp16 = const()[name = tensor<string, []>("op_5141_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_5142_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_5141_to_fp16)[name = tensor<string, []>("op_5142_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1181_cast_fp16 = add(x = input_1169_cast_fp16, y = var_5142_cast_fp16)[name = tensor<string, []>("input_1181_cast_fp16")];
+            tensor<int32, [1]> key_45_axes_0 = const()[name = tensor<string, []>("key_45_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483823744)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483825856)))];
+            tensor<fp16, [1, 56, 1024]> key_45_cast_fp16 = layer_norm(axes = key_45_axes_0, beta = encoder_layers_22_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_self_att_weight_to_fp16, x = input_1181_cast_fp16)[name = tensor<string, []>("key_45_cast_fp16")];
+            tensor<bool, []> input_1183_interleave_0 = const()[name = tensor<string, []>("input_1183_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_1183_cast_fp16 = concat(axis = var_68, interleave = input_1183_interleave_0, values = (cache_89_cast_fp16, key_45_cast_fp16))[name = tensor<string, []>("input_1183_cast_fp16")];
+            tensor<bool, []> var_5170_interleave_0 = const()[name = tensor<string, []>("op_5170_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> var_5170_cast_fp16 = concat(axis = var_68, interleave = var_5170_interleave_0, values = key_45_cast_fp16)[name = tensor<string, []>("op_5170_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483827968)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(485925184)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_201_cast_fp16")];
+            tensor<int32, [4]> var_5175 = const()[name = tensor<string, []>("op_5175"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_133_cast_fp16 = reshape(shape = var_5175, x = linear_201_cast_fp16)[name = tensor<string, []>("q_133_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(485927296)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(488024512)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_202_cast_fp16")];
+            tensor<int32, [4]> var_5180 = const()[name = tensor<string, []>("op_5180"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_89_cast_fp16 = reshape(shape = var_5180, x = linear_202_cast_fp16)[name = tensor<string, []>("k_89_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(488026624)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490123840)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_203_cast_fp16")];
+            tensor<int32, [4]> var_5185 = const()[name = tensor<string, []>("op_5185"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_45_cast_fp16 = reshape(shape = var_5185, x = linear_203_cast_fp16)[name = tensor<string, []>("v_45_cast_fp16")];
+            tensor<int32, [4]> value_53_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_22_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490125952)))];
+            tensor<fp16, [1, 56, 8, 128]> var_5198_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_5198_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_22_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490128064)))];
+            tensor<fp16, [1, 56, 8, 128]> var_5200_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_5200_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_45_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_579_transpose_x_0 = const()[name = tensor<string, []>("x_579_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_579_transpose_y_0 = const()[name = tensor<string, []>("x_579_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_5202_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_5202_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490130176))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490329920))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_45_cast_fp16 = transpose(perm = q_with_bias_v_45_perm_0, x = var_5200_cast_fp16)[name = tensor<string, []>("transpose_164")];
+            tensor<fp16, [1, 8, 56, 195]> 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_5202_to_fp16_quantized)[name = tensor<string, []>("x_579_cast_fp16")];
+            tensor<int32, [8]> x_581_pad_0 = const()[name = tensor<string, []>("x_581_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_581_mode_0 = const()[name = tensor<string, []>("x_581_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_365_to_fp16 = const()[name = tensor<string, []>("const_365_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_581_cast_fp16")];
+            tensor<int32, [4]> var_5210 = const()[name = tensor<string, []>("op_5210"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_583_cast_fp16 = reshape(shape = var_5210, x = x_581_cast_fp16)[name = tensor<string, []>("x_583_cast_fp16")];
+            tensor<int32, [4]> var_5214_begin_0 = const()[name = tensor<string, []>("op_5214_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5214_end_0 = const()[name = tensor<string, []>("op_5214_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_5214_end_mask_0 = const()[name = tensor<string, []>("op_5214_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_5214_cast_fp16 = slice_by_index(begin = var_5214_begin_0, end = var_5214_end_0, end_mask = var_5214_end_mask_0, x = x_583_cast_fp16)[name = tensor<string, []>("op_5214_cast_fp16")];
+            tensor<int32, [4]> var_5215 = const()[name = tensor<string, []>("op_5215"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_89_cast_fp16 = reshape(shape = var_5215, x = var_5214_cast_fp16)[name = tensor<string, []>("matrix_bd_89_cast_fp16")];
+            tensor<bool, []> matrix_ac_45_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_45_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_45_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_45_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_140_perm_0 = const()[name = tensor<string, []>("transpose_140_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_141_perm_0 = const()[name = tensor<string, []>("transpose_141_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89_cast_fp16)[name = tensor<string, []>("transpose_162")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_140 = transpose(perm = transpose_140_perm_0, x = var_5198_cast_fp16)[name = tensor<string, []>("transpose_163")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_45_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_91_begin_0 = const()[name = tensor<string, []>("matrix_bd_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_91_end_0 = const()[name = tensor<string, []>("matrix_bd_91_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_91_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_91_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_5224_cast_fp16 = add(x = matrix_ac_45_cast_fp16, y = matrix_bd_91_cast_fp16)[name = tensor<string, []>("op_5224_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_89_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_89_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_89_cast_fp16 = mul(x = var_5224_cast_fp16, y = _inversed_scores_89_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_89_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_91_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_89_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_91_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_5230_cast_fp16 = softmax(axis = var_59, x = scores_91_cast_fp16)[name = tensor<string, []>("op_5230_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_1185_cast_fp16 = select(a = var_44_to_fp16, b = var_5230_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_1185_cast_fp16")];
+            tensor<bool, []> x_585_transpose_x_0 = const()[name = tensor<string, []>("x_585_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_585_transpose_y_0 = const()[name = tensor<string, []>("x_585_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = v_45_cast_fp16)[name = tensor<string, []>("transpose_161")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_585_cast_fp16")];
+            tensor<int32, [4]> var_5234_perm_0 = const()[name = tensor<string, []>("op_5234_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5235 = const()[name = tensor<string, []>("op_5235"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_5234_cast_fp16 = transpose(perm = var_5234_perm_0, x = x_585_cast_fp16)[name = tensor<string, []>("transpose_160")];
+            tensor<fp16, [1, 56, 1024]> input_1187_cast_fp16 = reshape(shape = var_5235, x = var_5234_cast_fp16)[name = tensor<string, []>("input_1187_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490330432)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492427648)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_205_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1191_cast_fp16 = add(x = input_1181_cast_fp16, y = linear_205_cast_fp16)[name = tensor<string, []>("input_1191_cast_fp16")];
+            tensor<int32, [1]> x_589_axes_0 = const()[name = tensor<string, []>("x_589_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492429760)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492431872)))];
+            tensor<fp16, [1, 56, 1024]> x_589_cast_fp16 = layer_norm(axes = x_589_axes_0, beta = encoder_layers_22_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_conv_weight_to_fp16, x = input_1191_cast_fp16)[name = tensor<string, []>("x_589_cast_fp16")];
+            tensor<int32, [3]> input_1193_perm_0 = const()[name = tensor<string, []>("input_1193_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_1195_pad_type_0 = const()[name = tensor<string, []>("input_1195_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_1195_strides_0 = const()[name = tensor<string, []>("input_1195_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1195_pad_0 = const()[name = tensor<string, []>("input_1195_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1195_dilations_0 = const()[name = tensor<string, []>("input_1195_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_1195_groups_0 = const()[name = tensor<string, []>("input_1195_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492433984))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(494531200))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_1193_cast_fp16 = transpose(perm = input_1193_perm_0, x = x_589_cast_fp16)[name = tensor<string, []>("transpose_159")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_1195_cast_fp16")];
+            tensor<int32, []> x_591_split_num_splits_0 = const()[name = tensor<string, []>("x_591_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_591_split_axis_0 = const()[name = tensor<string, []>("x_591_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_591_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_591_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_591_split_1_sigmoid_cast_fp16 = sigmoid(x = x_591_split_cast_fp16_1)[name = tensor<string, []>("x_591_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_591_cast_fp16 = mul(x = x_591_split_cast_fp16_0, y = x_591_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_591_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_1197_cast_fp16 = select(a = var_44_to_fp16, b = x_591_cast_fp16, cond = var_575)[name = tensor<string, []>("input_1197_cast_fp16")];
+            tensor<bool, []> new_x_91_interleave_0 = const()[name = tensor<string, []>("new_x_91_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_91_cast_fp16 = concat(axis = var_59, interleave = new_x_91_interleave_0, values = (cache_91_cast_fp16, input_1197_cast_fp16))[name = tensor<string, []>("new_x_91_cast_fp16")];
+            tensor<int32, [3]> var_5274_begin_0 = const()[name = tensor<string, []>("op_5274_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_5274_end_0 = const()[name = tensor<string, []>("op_5274_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_5274_end_mask_0 = const()[name = tensor<string, []>("op_5274_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_5274_cast_fp16 = slice_by_index(begin = var_5274_begin_0, end = var_5274_end_0, end_mask = var_5274_end_mask_0, x = new_x_91_cast_fp16)[name = tensor<string, []>("op_5274_cast_fp16")];
+            tensor<string, []> x_593_pad_type_0 = const()[name = tensor<string, []>("x_593_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_593_groups_0 = const()[name = tensor<string, []>("x_593_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_593_strides_0 = const()[name = tensor<string, []>("x_593_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_593_pad_0 = const()[name = tensor<string, []>("x_593_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_593_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(494535360))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(494544640))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_593_cast_fp16")];
+            tensor<int32, [3]> input_1199_perm_0 = const()[name = tensor<string, []>("input_1199_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_595_axes_0 = const()[name = tensor<string, []>("x_595_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(494546752)))];
+            tensor<fp16, [1024]> encoder_layers_22_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(494548864)))];
+            tensor<fp16, [1, 56, 1024]> input_1199_cast_fp16 = transpose(perm = input_1199_perm_0, x = x_593_cast_fp16)[name = tensor<string, []>("transpose_158")];
+            tensor<fp16, [1, 56, 1024]> x_595_cast_fp16 = layer_norm(axes = x_595_axes_0, beta = encoder_layers_22_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_conv_batch_norm_weight_to_fp16, x = input_1199_cast_fp16)[name = tensor<string, []>("x_595_cast_fp16")];
+            tensor<int32, [3]> input_1201_perm_0 = const()[name = tensor<string, []>("input_1201_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_1201_cast_fp16 = transpose(perm = input_1201_perm_0, x = x_595_cast_fp16)[name = tensor<string, []>("transpose_157")];
+            tensor<fp16, [1, 1024, 56]> input_1203_cast_fp16 = silu(x = input_1201_cast_fp16)[name = tensor<string, []>("input_1203_cast_fp16")];
+            tensor<string, []> x_597_pad_type_0 = const()[name = tensor<string, []>("x_597_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_597_strides_0 = const()[name = tensor<string, []>("x_597_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_597_pad_0 = const()[name = tensor<string, []>("x_597_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_597_dilations_0 = const()[name = tensor<string, []>("x_597_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_597_groups_0 = const()[name = tensor<string, []>("x_597_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(494550976))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(495599616))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_597_cast_fp16")];
+            tensor<int32, [3]> input_1205_perm_0 = const()[name = tensor<string, []>("input_1205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_1205_cast_fp16 = transpose(perm = input_1205_perm_0, x = x_597_cast_fp16)[name = tensor<string, []>("transpose_156")];
+            tensor<fp16, [1, 56, 1024]> input_1207_cast_fp16 = add(x = input_1191_cast_fp16, y = input_1205_cast_fp16)[name = tensor<string, []>("input_1207_cast_fp16")];
+            tensor<int32, [1]> input_1209_axes_0 = const()[name = tensor<string, []>("input_1209_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(495601728)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(495603840)))];
+            tensor<fp16, [1, 56, 1024]> input_1209_cast_fp16 = layer_norm(axes = input_1209_axes_0, beta = encoder_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_feed_forward2_weight_to_fp16, x = input_1207_cast_fp16)[name = tensor<string, []>("input_1209_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_22_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(495605952)))];
+            tensor<fp16, [4096]> encoder_layers_22_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(503994624)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_206_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1213_cast_fp16 = silu(x = linear_206_cast_fp16)[name = tensor<string, []>("input_1213_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_22_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504002880)))];
+            tensor<fp16, [1024]> encoder_layers_22_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(512391552)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_207_cast_fp16")];
+            tensor<fp16, []> var_5317_to_fp16 = const()[name = tensor<string, []>("op_5317_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_5318_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_5317_to_fp16)[name = tensor<string, []>("op_5318_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1219_cast_fp16 = add(x = input_1207_cast_fp16, y = var_5318_cast_fp16)[name = tensor<string, []>("input_1219_cast_fp16")];
+            tensor<int32, [1]> input_1221_axes_0 = const()[name = tensor<string, []>("input_1221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(512393664)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(512395776)))];
+            tensor<fp16, [1, 56, 1024]> input_1221_cast_fp16 = layer_norm(axes = input_1221_axes_0, beta = encoder_layers_22_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_out_weight_to_fp16, x = input_1219_cast_fp16)[name = tensor<string, []>("input_1221_cast_fp16")];
+            tensor<int32, [4]> cache_93_begin_0 = const()[name = tensor<string, []>("cache_93_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> cache_93_end_0 = const()[name = tensor<string, []>("cache_93_end_0"), val = tensor<int32, [4]>([24, 1, 42, 1024])];
+            tensor<bool, [4]> cache_93_end_mask_0 = const()[name = tensor<string, []>("cache_93_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_93_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_93_cast_fp16")];
+            tensor<int32, [4]> cache_begin_0 = const()[name = tensor<string, []>("cache_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> cache_end_0 = const()[name = tensor<string, []>("cache_end_0"), val = tensor<int32, [4]>([24, 1, 1024, 8])];
+            tensor<bool, [4]> cache_end_mask_0 = const()[name = tensor<string, []>("cache_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_cast_fp16")];
+            tensor<int32, [1]> input_1223_axes_0 = const()[name = tensor<string, []>("input_1223_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(512397888)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(512400000)))];
+            tensor<fp16, [1, 56, 1024]> input_1223_cast_fp16 = layer_norm(axes = input_1223_axes_0, beta = encoder_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_feed_forward1_weight_to_fp16, x = input_1221_cast_fp16)[name = tensor<string, []>("input_1223_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_23_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(512402112)))];
+            tensor<fp16, [4096]> encoder_layers_23_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(520790784)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_208_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1227_cast_fp16 = silu(x = linear_208_cast_fp16)[name = tensor<string, []>("input_1227_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_23_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(520799040)))];
+            tensor<fp16, [1024]> encoder_layers_23_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529187712)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_209_cast_fp16")];
+            tensor<fp16, []> var_5354_to_fp16 = const()[name = tensor<string, []>("op_5354_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_5355_cast_fp16 = mul(x = linear_209_cast_fp16, y = var_5354_to_fp16)[name = tensor<string, []>("op_5355_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1233_cast_fp16 = add(x = input_1221_cast_fp16, y = var_5355_cast_fp16)[name = tensor<string, []>("input_1233_cast_fp16")];
+            tensor<int32, [1]> key_axes_0 = const()[name = tensor<string, []>("key_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529189824)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529191936)))];
+            tensor<fp16, [1, 56, 1024]> key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_23_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_self_att_weight_to_fp16, x = input_1233_cast_fp16)[name = tensor<string, []>("key_cast_fp16")];
+            tensor<bool, []> input_1235_interleave_0 = const()[name = tensor<string, []>("input_1235_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 98, 1024]> input_1235_cast_fp16 = concat(axis = var_68, interleave = input_1235_interleave_0, values = (cache_93_cast_fp16, key_cast_fp16))[name = tensor<string, []>("input_1235_cast_fp16")];
+            tensor<bool, []> cache_last_channel_cur_interleave_0 = const()[name = tensor<string, []>("cache_last_channel_cur_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1024]> cache_last_channel_cur_cast_fp16 = concat(axis = var_68, interleave = cache_last_channel_cur_interleave_0, values = key_cast_fp16)[name = tensor<string, []>("cache_last_channel_cur_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529194048)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(531291264)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_210_cast_fp16")];
+            tensor<int32, [4]> var_5388 = const()[name = tensor<string, []>("op_5388"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_139_cast_fp16 = reshape(shape = var_5388, x = linear_210_cast_fp16)[name = tensor<string, []>("q_139_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(531293376)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533390592)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_211_cast_fp16")];
+            tensor<int32, [4]> var_5393 = const()[name = tensor<string, []>("op_5393"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_93_cast_fp16 = reshape(shape = var_5393, x = linear_211_cast_fp16)[name = tensor<string, []>("k_93_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533392704)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(535489920)))];
+            tensor<fp16, [1, 98, 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 = tensor<string, []>("linear_212_cast_fp16")];
+            tensor<int32, [4]> var_5398 = const()[name = tensor<string, []>("op_5398"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_cast_fp16 = reshape(shape = var_5398, x = linear_212_cast_fp16)[name = tensor<string, []>("v_cast_fp16")];
+            tensor<int32, [4]> value_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_23_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(535492032)))];
+            tensor<fp16, [1, 56, 8, 128]> var_5411_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_5411_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_23_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(535494144)))];
+            tensor<fp16, [1, 56, 8, 128]> var_5413_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_5413_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_605_transpose_x_0 = const()[name = tensor<string, []>("x_605_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_605_transpose_y_0 = const()[name = tensor<string, []>("x_605_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_5415_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_5415_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(535496256))), scale = tensor<fp16, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(535696000))), zero_point = tensor<int8, [195]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16391232)))];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_5413_cast_fp16)[name = tensor<string, []>("transpose_155")];
+            tensor<fp16, [1, 8, 56, 195]> 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_5415_to_fp16_quantized)[name = tensor<string, []>("x_605_cast_fp16")];
+            tensor<int32, [8]> x_607_pad_0 = const()[name = tensor<string, []>("x_607_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_607_mode_0 = const()[name = tensor<string, []>("x_607_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_378_to_fp16 = const()[name = tensor<string, []>("const_378_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> 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 = tensor<string, []>("x_607_cast_fp16")];
+            tensor<int32, [4]> var_5423 = const()[name = tensor<string, []>("op_5423"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_609_cast_fp16 = reshape(shape = var_5423, x = x_607_cast_fp16)[name = tensor<string, []>("x_609_cast_fp16")];
+            tensor<int32, [4]> var_5427_begin_0 = const()[name = tensor<string, []>("op_5427_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5427_end_0 = const()[name = tensor<string, []>("op_5427_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_5427_end_mask_0 = const()[name = tensor<string, []>("op_5427_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_5427_cast_fp16 = slice_by_index(begin = var_5427_begin_0, end = var_5427_end_0, end_mask = var_5427_end_mask_0, x = x_609_cast_fp16)[name = tensor<string, []>("op_5427_cast_fp16")];
+            tensor<int32, [4]> var_5428 = const()[name = tensor<string, []>("op_5428"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_93_cast_fp16 = reshape(shape = var_5428, x = var_5427_cast_fp16)[name = tensor<string, []>("matrix_bd_93_cast_fp16")];
+            tensor<bool, []> matrix_ac_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_142_perm_0 = const()[name = tensor<string, []>("transpose_142_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_143_perm_0 = const()[name = tensor<string, []>("transpose_143_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93_cast_fp16)[name = tensor<string, []>("transpose_153")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_142 = transpose(perm = transpose_142_perm_0, x = var_5411_cast_fp16)[name = tensor<string, []>("transpose_154")];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_ac_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_begin_0 = const()[name = tensor<string, []>("matrix_bd_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_end_0 = const()[name = tensor<string, []>("matrix_bd_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> 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 = tensor<string, []>("matrix_bd_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_5437_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = tensor<string, []>("op_5437_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_93_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_93_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_93_cast_fp16 = mul(x = var_5437_cast_fp16, y = _inversed_scores_93_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_93_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_93_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_5443_cast_fp16 = softmax(axis = var_59, x = scores_cast_fp16)[name = tensor<string, []>("op_5443_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_1237_cast_fp16 = select(a = var_44_to_fp16, b = var_5443_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_1237_cast_fp16")];
+            tensor<bool, []> x_611_transpose_x_0 = const()[name = tensor<string, []>("x_611_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_611_transpose_y_0 = const()[name = tensor<string, []>("x_611_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = tensor<string, []>("transpose_152")];
+            tensor<fp16, [1, 8, 56, 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 = tensor<string, []>("x_611_cast_fp16")];
+            tensor<int32, [4]> var_5447_perm_0 = const()[name = tensor<string, []>("op_5447_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5448 = const()[name = tensor<string, []>("op_5448"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_5447_cast_fp16 = transpose(perm = var_5447_perm_0, x = x_611_cast_fp16)[name = tensor<string, []>("transpose_151")];
+            tensor<fp16, [1, 56, 1024]> input_1239_cast_fp16 = reshape(shape = var_5448, x = var_5447_cast_fp16)[name = tensor<string, []>("input_1239_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(535696512)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(537793728)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_214_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1243_cast_fp16 = add(x = input_1233_cast_fp16, y = linear_214_cast_fp16)[name = tensor<string, []>("input_1243_cast_fp16")];
+            tensor<int32, [1]> x_615_axes_0 = const()[name = tensor<string, []>("x_615_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(537795840)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(537797952)))];
+            tensor<fp16, [1, 56, 1024]> x_615_cast_fp16 = layer_norm(axes = x_615_axes_0, beta = encoder_layers_23_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_conv_weight_to_fp16, x = input_1243_cast_fp16)[name = tensor<string, []>("x_615_cast_fp16")];
+            tensor<int32, [3]> input_1245_perm_0 = const()[name = tensor<string, []>("input_1245_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_1247_pad_type_0 = const()[name = tensor<string, []>("input_1247_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_1247_strides_0 = const()[name = tensor<string, []>("input_1247_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1247_pad_0 = const()[name = tensor<string, []>("input_1247_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1247_dilations_0 = const()[name = tensor<string, []>("input_1247_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_1247_groups_0 = const()[name = tensor<string, []>("input_1247_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(537800064))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(539897280))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19546368)))];
+            tensor<fp16, [1, 1024, 56]> input_1245_cast_fp16 = transpose(perm = input_1245_perm_0, x = x_615_cast_fp16)[name = tensor<string, []>("transpose_150")];
+            tensor<fp16, [1, 2048, 56]> 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 = tensor<string, []>("input_1247_cast_fp16")];
+            tensor<int32, []> x_617_split_num_splits_0 = const()[name = tensor<string, []>("x_617_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_617_split_axis_0 = const()[name = tensor<string, []>("x_617_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 56]> x_617_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_617_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_617_split_1_sigmoid_cast_fp16 = sigmoid(x = x_617_split_cast_fp16_1)[name = tensor<string, []>("x_617_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_617_cast_fp16 = mul(x = x_617_split_cast_fp16_0, y = x_617_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_617_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_1249_cast_fp16 = select(a = var_44_to_fp16, b = x_617_cast_fp16, cond = var_575)[name = tensor<string, []>("input_1249_cast_fp16")];
+            tensor<bool, []> new_x_interleave_0 = const()[name = tensor<string, []>("new_x_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_cast_fp16 = concat(axis = var_59, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_1249_cast_fp16))[name = tensor<string, []>("new_x_cast_fp16")];
+            tensor<int32, [3]> cache_last_time_cur_begin_0 = const()[name = tensor<string, []>("cache_last_time_cur_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> cache_last_time_cur_end_0 = const()[name = tensor<string, []>("cache_last_time_cur_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> cache_last_time_cur_end_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_last_time_cur_cast_fp16")];
+            tensor<string, []> x_619_pad_type_0 = const()[name = tensor<string, []>("x_619_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_619_groups_0 = const()[name = tensor<string, []>("x_619_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_619_strides_0 = const()[name = tensor<string, []>("x_619_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_619_pad_0 = const()[name = tensor<string, []>("x_619_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_619_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(539901440))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(539910720))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_619_cast_fp16")];
+            tensor<int32, [3]> input_1251_perm_0 = const()[name = tensor<string, []>("input_1251_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_621_axes_0 = const()[name = tensor<string, []>("x_621_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(539912832)))];
+            tensor<fp16, [1024]> encoder_layers_23_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(539914944)))];
+            tensor<fp16, [1, 56, 1024]> input_1251_cast_fp16 = transpose(perm = input_1251_perm_0, x = x_619_cast_fp16)[name = tensor<string, []>("transpose_149")];
+            tensor<fp16, [1, 56, 1024]> x_621_cast_fp16 = layer_norm(axes = x_621_axes_0, beta = encoder_layers_23_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_conv_batch_norm_weight_to_fp16, x = input_1251_cast_fp16)[name = tensor<string, []>("x_621_cast_fp16")];
+            tensor<int32, [3]> input_1253_perm_0 = const()[name = tensor<string, []>("input_1253_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_1253_cast_fp16 = transpose(perm = input_1253_perm_0, x = x_621_cast_fp16)[name = tensor<string, []>("transpose_148")];
+            tensor<fp16, [1, 1024, 56]> input_1255_cast_fp16 = silu(x = input_1253_cast_fp16)[name = tensor<string, []>("input_1255_cast_fp16")];
+            tensor<string, []> x_623_pad_type_0 = const()[name = tensor<string, []>("x_623_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_623_strides_0 = const()[name = tensor<string, []>("x_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_623_pad_0 = const()[name = tensor<string, []>("x_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_623_dilations_0 = const()[name = tensor<string, []>("x_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_623_groups_0 = const()[name = tensor<string, []>("x_623_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(539917056))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(540965696))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 56]> 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 = tensor<string, []>("x_623_cast_fp16")];
+            tensor<int32, [3]> input_1257_perm_0 = const()[name = tensor<string, []>("input_1257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_1257_cast_fp16 = transpose(perm = input_1257_perm_0, x = x_623_cast_fp16)[name = tensor<string, []>("transpose_147")];
+            tensor<fp16, [1, 56, 1024]> input_1259_cast_fp16 = add(x = input_1243_cast_fp16, y = input_1257_cast_fp16)[name = tensor<string, []>("input_1259_cast_fp16")];
+            tensor<int32, [1]> input_1261_axes_0 = const()[name = tensor<string, []>("input_1261_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(540967808)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(540969920)))];
+            tensor<fp16, [1, 56, 1024]> input_1261_cast_fp16 = layer_norm(axes = input_1261_axes_0, beta = encoder_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_feed_forward2_weight_to_fp16, x = input_1259_cast_fp16)[name = tensor<string, []>("input_1261_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_23_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(540972032)))];
+            tensor<fp16, [4096]> encoder_layers_23_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(549360704)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_215_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1265_cast_fp16 = silu(x = linear_215_cast_fp16)[name = tensor<string, []>("input_1265_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_23_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(549368960)))];
+            tensor<fp16, [1024]> encoder_layers_23_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(557757632)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_216_cast_fp16")];
+            tensor<fp16, []> var_5530_to_fp16 = const()[name = tensor<string, []>("op_5530_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_5531_cast_fp16 = mul(x = linear_216_cast_fp16, y = var_5530_to_fp16)[name = tensor<string, []>("op_5531_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1271_cast_fp16 = add(x = input_1259_cast_fp16, y = var_5531_cast_fp16)[name = tensor<string, []>("input_1271_cast_fp16")];
+            tensor<int32, [1]> audio_signal_axes_0 = const()[name = tensor<string, []>("audio_signal_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(557759744)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(557761856)))];
+            tensor<fp16, [1, 56, 1024]> audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_23_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_out_weight_to_fp16, x = input_1271_cast_fp16)[name = tensor<string, []>("audio_signal_cast_fp16")];
+            tensor<int32, []> obj_5_axis_0 = const()[name = tensor<string, []>("obj_5_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [24, 1, 56, 1024]> obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_484_cast_fp16, var_697_cast_fp16, var_910_cast_fp16, var_1123_cast_fp16, var_1336_cast_fp16, var_1549_cast_fp16, var_1762_cast_fp16, var_1975_cast_fp16, var_2188_cast_fp16, var_2401_cast_fp16, var_2614_cast_fp16, var_2827_cast_fp16, var_3040_cast_fp16, var_3253_cast_fp16, var_3466_cast_fp16, var_3679_cast_fp16, var_3892_cast_fp16, var_4105_cast_fp16, var_4318_cast_fp16, var_4531_cast_fp16, var_4744_cast_fp16, var_4957_cast_fp16, var_5170_cast_fp16, cache_last_channel_cur_cast_fp16))[name = tensor<string, []>("obj_5_cast_fp16")];
+            tensor<int32, []> obj_7_axis_0 = const()[name = tensor<string, []>("obj_7_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [24, 1, 1024, 8]> obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_588_cast_fp16, var_801_cast_fp16, var_1014_cast_fp16, var_1227_cast_fp16, var_1440_cast_fp16, var_1653_cast_fp16, var_1866_cast_fp16, var_2079_cast_fp16, var_2292_cast_fp16, var_2505_cast_fp16, var_2718_cast_fp16, var_2931_cast_fp16, var_3144_cast_fp16, var_3357_cast_fp16, var_3570_cast_fp16, var_3783_cast_fp16, var_3996_cast_fp16, var_4209_cast_fp16, var_4422_cast_fp16, var_4635_cast_fp16, var_4848_cast_fp16, var_5061_cast_fp16, var_5274_cast_fp16, cache_last_time_cur_cast_fp16))[name = tensor<string, []>("obj_7_cast_fp16")];
+            tensor<int32, [1]> var_5547 = add(x = cache_len, y = max_audio_length_1)[name = tensor<string, []>("op_5547")];
+            tensor<string, []> var_5547_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_5547_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> const_384_to_fp16 = const()[name = tensor<string, []>("const_384_to_fp16"), val = tensor<fp16, []>(-inf)];
+            tensor<fp16, []> var_49_promoted_to_fp16 = const()[name = tensor<string, []>("op_49_promoted_to_fp16"), val = tensor<fp16, []>(0x1.5p+5)];
+            tensor<fp16, [1]> var_5547_to_fp16 = cast(dtype = var_5547_promoted_to_fp16_dtype_0, x = var_5547)[name = tensor<string, []>("cast_6")];
+            tensor<fp16, [1]> clip_1_cast_fp16 = clip(alpha = const_384_to_fp16, beta = var_49_promoted_to_fp16, x = var_5547_to_fp16)[name = tensor<string, []>("clip_1_cast_fp16")];
+            tensor<int32, [4]> var_5569_begin_0 = const()[name = tensor<string, []>("op_5569_begin_0"), val = tensor<int32, [4]>([0, 0, 14, 0])];
+            tensor<int32, [4]> var_5569_end_0 = const()[name = tensor<string, []>("op_5569_end_0"), val = tensor<int32, [4]>([24, 1, 56, 1024])];
+            tensor<bool, [4]> var_5569_end_mask_0 = const()[name = tensor<string, []>("op_5569_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [24, 1, 42, 1024]> var_5569_cast_fp16 = slice_by_index(begin = var_5569_begin_0, end = var_5569_end_0, end_mask = var_5569_end_mask_0, x = obj_5_cast_fp16)[name = tensor<string, []>("op_5569_cast_fp16")];
+            tensor<int32, []> var_5584 = const()[name = tensor<string, []>("op_5584"), val = tensor<int32, []>(0)];
+            tensor<int32, []> one_hot_1_batch_dims_0 = const()[name = tensor<string, []>("one_hot_1_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> one_hot_1_validate_indices_0 = const()[name = tensor<string, []>("one_hot_1_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [128, 128]> to_onehot_identity_table_to_fp16 = const()[name = tensor<string, []>("to_onehot_identity_table_to_fp16"), val = tensor<fp16, [128, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(557763968)))];
+            tensor<string, []> prompt_id_to_int16_dtype_0 = const()[name = tensor<string, []>("prompt_id_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1]> prompt_id_to_int16 = cast(dtype = prompt_id_to_int16_dtype_0, x = prompt_id)[name = tensor<string, []>("cast_5")];
+            tensor<fp16, [1, 128]> one_hot_1_cast_fp16_cast_uint16 = gather(axis = var_5584, batch_dims = one_hot_1_batch_dims_0, indices = prompt_id_to_int16, validate_indices = one_hot_1_validate_indices_0, x = to_onehot_identity_table_to_fp16)[name = tensor<string, []>("one_hot_1_cast_fp16_cast_uint16")];
+            tensor<int32, [1]> var_5593_axes_0 = const()[name = tensor<string, []>("op_5593_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 128]> var_5593_cast_fp16 = expand_dims(axes = var_5593_axes_0, x = one_hot_1_cast_fp16_cast_uint16)[name = tensor<string, []>("op_5593_cast_fp16")];
+            tensor<int32, [3]> one_hot_reps_0 = const()[name = tensor<string, []>("one_hot_reps_0"), val = tensor<int32, [3]>([1, 56, 1])];
+            tensor<fp16, [1, 56, 128]> one_hot_cast_fp16 = tile(reps = one_hot_reps_0, x = var_5593_cast_fp16)[name = tensor<string, []>("one_hot_cast_fp16")];
+            tensor<int32, []> var_5602 = const()[name = tensor<string, []>("op_5602"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> input_1273_interleave_0 = const()[name = tensor<string, []>("input_1273_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 56, 1152]> input_1273_cast_fp16 = concat(axis = var_5602, interleave = input_1273_interleave_0, values = (audio_signal_cast_fp16, one_hot_cast_fp16))[name = tensor<string, []>("input_1273_cast_fp16")];
+            tensor<fp16, [2048, 1152]> prompt_kernel_0_weight_to_fp16 = const()[name = tensor<string, []>("prompt_kernel_0_weight_to_fp16"), val = tensor<fp16, [2048, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(557796800)))];
+            tensor<fp16, [2048]> prompt_kernel_0_bias_to_fp16 = const()[name = tensor<string, []>("prompt_kernel_0_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(562515456)))];
+            tensor<fp16, [1, 56, 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 = tensor<string, []>("linear_217_cast_fp16")];
+            tensor<fp16, [1, 56, 2048]> input_cast_fp16 = relu(x = linear_217_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<fp16, [1024, 2048]> prompt_kernel_2_weight_to_fp16 = const()[name = tensor<string, []>("prompt_kernel_2_weight_to_fp16"), val = tensor<fp16, [1024, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(562519616)))];
+            tensor<fp16, [1024]> prompt_kernel_2_bias_to_fp16 = const()[name = tensor<string, []>("prompt_kernel_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(566713984)))];
+            tensor<fp16, [1, 56, 1024]> linear_218_cast_fp16 = linear(bias = prompt_kernel_2_bias_to_fp16, weight = prompt_kernel_2_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("linear_218_cast_fp16")];
+            tensor<int32, [3]> var_5615_perm_0 = const()[name = tensor<string, []>("op_5615_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> var_5615_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_5615_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<string, []> cast_246_dtype_0 = const()[name = tensor<string, []>("cast_246_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [4]> var_5623_perm_0 = const()[name = tensor<string, []>("op_5623_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<string, []> var_5623_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_5623_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [4]> var_5626_perm_0 = const()[name = tensor<string, []>("op_5626_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<string, []> var_5626_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_5626_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<string, []> cast_247_dtype_0 = const()[name = tensor<string, []>("cast_247_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1]> cache_len_out = cast(dtype = cast_247_dtype_0, x = clip_1_cast_fp16)[name = tensor<string, []>("cast_0")];
+            tensor<fp16, [1, 24, 1024, 8]> var_5626_cast_fp16 = transpose(perm = var_5626_perm_0, x = obj_7_cast_fp16)[name = tensor<string, []>("transpose_144")];
+            tensor<fp32, [1, 24, 1024, 8]> cache_time_out = cast(dtype = var_5626_cast_fp16_to_fp32_dtype_0, x = var_5626_cast_fp16)[name = tensor<string, []>("cast_1")];
+            tensor<fp16, [1, 24, 42, 1024]> var_5623_cast_fp16 = transpose(perm = var_5623_perm_0, x = var_5569_cast_fp16)[name = tensor<string, []>("transpose_145")];
+            tensor<fp32, [1, 24, 42, 1024]> cache_channel_out = cast(dtype = var_5623_cast_fp16_to_fp32_dtype_0, x = var_5623_cast_fp16)[name = tensor<string, []>("cast_2")];
+            tensor<int32, [1]> encoded_length = cast(dtype = cast_246_dtype_0, x = clip_0_cast_fp16)[name = tensor<string, []>("cast_3")];
+            tensor<fp16, [1, 1024, 56]> var_5615_cast_fp16 = transpose(perm = var_5615_perm_0, x = linear_218_cast_fp16)[name = tensor<string, []>("transpose_146")];
+            tensor<fp32, [1, 1024, 56]> encoded = cast(dtype = var_5615_cast_fp16_to_fp32_dtype_0, x = var_5615_cast_fp16)[name = tensor<string, []>("cast_4")];
+        } -> (encoded, encoded_length, cache_channel_out, cache_time_out, cache_len_out);
+}
\ No newline at end of file
diff --git a/latin/4480ms/encoder.mlmodelc/weights/weight.bin b/latin/4480ms/encoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..e2f89e3e4cfffde62ac5a610aa502cbeeeaefe67
--- /dev/null
+++ b/latin/4480ms/encoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:faf7625ad245102745aadcb7aa80793ff74130eea5185bf991bda11dcd4f2e64
+size 566716096
diff --git a/latin/4480ms/joint.mlmodelc/coremldata.bin b/latin/4480ms/joint.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..2eb3414f924bc2cb1344e4dfadcddafe5615f414
--- /dev/null
+++ b/latin/4480ms/joint.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed9ab0bdf1d741ecdbeafc300017a00cabb683e401736c9e207df20bfd820db5
+size 341
diff --git a/latin/4480ms/joint.mlmodelc/model.mil b/latin/4480ms/joint.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..58ce97d6dc855c13232cf61f81845dbca8e3759f
--- /dev/null
+++ b/latin/4480ms/joint.mlmodelc/model.mil
@@ -0,0 +1,31 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<fp32, [1, 640, 1]> decoder, tensor<fp32, [1, 1024, 1]> encoder) {
+            tensor<int32, [3]> input_1_perm_0 = const()[name = tensor<string, []>("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> encoder_to_fp16_dtype_0 = const()[name = tensor<string, []>("encoder_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> decoder_to_fp16_dtype_0 = const()[name = tensor<string, []>("decoder_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [640, 1024]> module_enc_weight_to_fp16 = const()[name = tensor<string, []>("module_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [640]> module_enc_bias_to_fp16 = const()[name = tensor<string, []>("module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1310848)))];
+            tensor<fp16, [1, 1024, 1]> encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = tensor<string, []>("cast_2")];
+            tensor<fp16, [1, 1, 1024]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_to_fp16)[name = tensor<string, []>("transpose_1")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = module_enc_bias_to_fp16, weight = module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> module_pred_weight_to_fp16 = const()[name = tensor<string, []>("module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1312192)))];
+            tensor<fp16, [640]> module_pred_bias_to_fp16 = const()[name = tensor<string, []>("module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2131456)))];
+            tensor<fp16, [1, 640, 1]> decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = tensor<string, []>("cast_1")];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_to_fp16)[name = tensor<string, []>("transpose_0")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = module_pred_bias_to_fp16, weight = module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_23_axes_0 = const()[name = tensor<string, []>("op_23_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_23_cast_fp16")];
+            tensor<int32, [1]> var_25_axes_0 = const()[name = tensor<string, []>("op_25_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_25_cast_fp16 = expand_dims(axes = var_25_axes_0, x = linear_1_cast_fp16)[name = tensor<string, []>("op_25_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_25_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<fp16, [2829, 640]> module_joint_net_2_weight_to_fp16 = const()[name = tensor<string, []>("module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [2829, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2132800)))];
+            tensor<fp16, [2829]> module_joint_net_2_bias_to_fp16 = const()[name = tensor<string, []>("module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [2829]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5753984)))];
+            tensor<fp16, [1, 1, 1, 2829]> linear_2_cast_fp16 = linear(bias = module_joint_net_2_bias_to_fp16, weight = module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<string, []> linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("linear_2_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 1, 1, 2829]> logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = tensor<string, []>("cast_0")];
+        } -> (logits);
+}
\ No newline at end of file
diff --git a/latin/4480ms/joint.mlmodelc/weights/weight.bin b/latin/4480ms/joint.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..baa3ed0a821090eb9e42550135aba83fb8e89bbe
--- /dev/null
+++ b/latin/4480ms/joint.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66b9d7405fad48fdd8413a7275c04d44933cb9beb80d318c158f80e52ff43fba
+size 5759706
diff --git a/latin/4480ms/preprocessor.mlmodelc/analytics/coremldata.bin b/latin/4480ms/preprocessor.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..09555a22fd71f6e9eab8b151f4ec43086af5e5de
--- /dev/null
+++ b/latin/4480ms/preprocessor.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:394f3db1360792923d29ee45e42733a51ebce50c73cd36459ac2cfe3fed2ed5d
+size 243
diff --git a/latin/4480ms/preprocessor.mlmodelc/coremldata.bin b/latin/4480ms/preprocessor.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..3388d8db6989bf0a6377d6c80f7d5d8693619a4d
--- /dev/null
+++ b/latin/4480ms/preprocessor.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f6f8787b47a3d6b7b7dafd655b9fa9e562b99c28d2abca2ea2f8ac597b29549
+size 371
diff --git a/latin/4480ms/preprocessor.mlmodelc/model.mil b/latin/4480ms/preprocessor.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..3dd53065828833a3eb30fdc575ba97353ee57d11
--- /dev/null
+++ b/latin/4480ms/preprocessor.mlmodelc/model.mil
@@ -0,0 +1,122 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<fp32, [1, ?]> audio, tensor<int32, [1]> audio_length) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"audio", [1, 1]}}), ("RangeDims", {{"audio", [[1, 1], [1, 1280000]]}})))] {
+            tensor<int32, []> var_9 = const()[name = tensor<string, []>("op_9"), val = tensor<int32, []>(1)];
+            tensor<int32, []> var_10 = const()[name = tensor<string, []>("op_10"), val = tensor<int32, []>(160)];
+            tensor<int32, []> var_12 = const()[name = tensor<string, []>("op_12"), val = tensor<int32, []>(0)];
+            tensor<int32, []> var_33 = const()[name = tensor<string, []>("op_33"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> var_34 = add(x = audio_length, y = var_33)[name = tensor<string, []>("op_34")];
+            tensor<int32, []> var_35 = const()[name = tensor<string, []>("op_35"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> var_36 = sub(x = var_34, y = var_35)[name = tensor<string, []>("op_36")];
+            tensor<int32, [1]> floor_div_0 = floor_div(x = var_36, y = var_10)[name = tensor<string, []>("floor_div_0")];
+            tensor<bool, [1]> var_39 = equal(x = audio_length, y = var_12)[name = tensor<string, []>("op_39")];
+            tensor<int32, [1]> var_40 = const()[name = tensor<string, []>("op_40"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> mel_length = select(a = var_40, b = floor_div_0, cond = var_39)[name = tensor<string, []>("seq_len")];
+            tensor<string, []> audio_to_fp16_dtype_0 = const()[name = tensor<string, []>("audio_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, ?]> audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = tensor<string, []>("cast_14")];
+            tensor<int32, [2]> var_42_shape_cast_fp16 = shape(x = audio_to_fp16)[name = tensor<string, []>("op_42_shape_cast_fp16")];
+            tensor<int32, []> gather_0_axis_0 = const()[name = tensor<string, []>("gather_0_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_0_batch_dims_0 = const()[name = tensor<string, []>("gather_0_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_0_validate_indices_0 = const()[name = tensor<string, []>("gather_0_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> var_42_shape_cast_fp16_to_int16_dtype_0 = const()[name = tensor<string, []>("op_42_shape_cast_fp16_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<uint16, []> select_0_to_uint16 = const()[name = tensor<string, []>("select_0_to_uint16"), val = tensor<uint16, []>(1)];
+            tensor<int16, [2]> var_42_shape_cast_fp16_to_int16 = cast(dtype = var_42_shape_cast_fp16_to_int16_dtype_0, x = var_42_shape_cast_fp16)[name = tensor<string, []>("cast_13")];
+            tensor<int16, []> gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_42_shape_cast_fp16_to_int16)[name = tensor<string, []>("gather_0_cast_uint16")];
+            tensor<string, []> gather_0_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_0_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, []> const_0 = const()[name = tensor<string, []>("const_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> const_1 = const()[name = tensor<string, []>("const_1"), val = tensor<int32, []>(1)];
+            tensor<int32, []> gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = tensor<string, []>("cast_12")];
+            tensor<int32, [?]> var_43 = range_1d(end = gather_0_cast_uint16_to_int32, start = const_0, step = const_1)[name = tensor<string, []>("op_43")];
+            tensor<int32, [1]> var_44_axes_0 = const()[name = tensor<string, []>("op_44_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = tensor<string, []>("op_44")];
+            tensor<int32, [1]> var_45_axes_0 = const()[name = tensor<string, []>("op_45_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_45 = expand_dims(axes = var_45_axes_0, x = audio_length)[name = tensor<string, []>("op_45")];
+            tensor<bool, [1, ?]> timemask = less(x = var_44, y = var_45)[name = tensor<string, []>("timemask")];
+            tensor<int32, [2]> var_48_begin_0 = const()[name = tensor<string, []>("op_48_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_48_end_0 = const()[name = tensor<string, []>("op_48_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> var_48_end_mask_0 = const()[name = tensor<string, []>("op_48_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<bool, [2]> var_48_squeeze_mask_0 = const()[name = tensor<string, []>("op_48_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
+            tensor<fp16, [1]> var_48_cast_fp16 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, squeeze_mask = var_48_squeeze_mask_0, x = audio_to_fp16)[name = tensor<string, []>("op_48_cast_fp16")];
+            tensor<int32, [1]> var_49_axes_0 = const()[name = tensor<string, []>("op_49_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1]> var_49_cast_fp16 = expand_dims(axes = var_49_axes_0, x = var_48_cast_fp16)[name = tensor<string, []>("op_49_cast_fp16")];
+            tensor<int32, [2]> var_51_begin_0 = const()[name = tensor<string, []>("op_51_begin_0"), val = tensor<int32, [2]>([0, 1])];
+            tensor<int32, [2]> var_51_end_0 = const()[name = tensor<string, []>("op_51_end_0"), val = tensor<int32, [2]>([1, 0])];
+            tensor<bool, [2]> var_51_end_mask_0 = const()[name = tensor<string, []>("op_51_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<fp16, [1, ?]> var_51_cast_fp16 = slice_by_index(begin = var_51_begin_0, end = var_51_end_0, end_mask = var_51_end_mask_0, x = audio_to_fp16)[name = tensor<string, []>("op_51_cast_fp16")];
+            tensor<int32, [2]> var_53_begin_0 = const()[name = tensor<string, []>("op_53_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_53_end_0 = const()[name = tensor<string, []>("op_53_end_0"), val = tensor<int32, [2]>([1, -1])];
+            tensor<bool, [2]> var_53_end_mask_0 = const()[name = tensor<string, []>("op_53_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp16, [1, ?]> var_53_cast_fp16 = slice_by_index(begin = var_53_begin_0, end = var_53_end_0, end_mask = var_53_end_mask_0, x = audio_to_fp16)[name = tensor<string, []>("op_53_cast_fp16")];
+            tensor<fp16, []> var_54_to_fp16 = const()[name = tensor<string, []>("op_54_to_fp16"), val = tensor<fp16, []>(0x1.f0cp-1)];
+            tensor<fp16, [1, ?]> var_55_cast_fp16 = mul(x = var_53_cast_fp16, y = var_54_to_fp16)[name = tensor<string, []>("op_55_cast_fp16")];
+            tensor<fp16, [1, ?]> var_56_cast_fp16 = sub(x = var_51_cast_fp16, y = var_55_cast_fp16)[name = tensor<string, []>("op_56_cast_fp16")];
+            tensor<bool, []> x_3_interleave_0 = const()[name = tensor<string, []>("x_3_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, ?]> x_3_cast_fp16 = concat(axis = var_9, interleave = x_3_interleave_0, values = (var_49_cast_fp16, var_56_cast_fp16))[name = tensor<string, []>("x_3_cast_fp16")];
+            tensor<bool, [1, ?]> var_59 = logical_not(x = timemask)[name = tensor<string, []>("op_59")];
+            tensor<fp16, []> var_16_to_fp16 = const()[name = tensor<string, []>("op_16_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, ?]> input_1_cast_fp16 = select(a = var_16_to_fp16, b = x_3_cast_fp16, cond = var_59)[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<int32, [3]> concat_1x = const()[name = tensor<string, []>("concat_1x"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, ?]> input_3_cast_fp16 = reshape(shape = concat_1x, x = input_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<int32, [6]> input_5_pad_0 = const()[name = tensor<string, []>("input_5_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 256, 256])];
+            tensor<string, []> input_5_mode_0 = const()[name = tensor<string, []>("input_5_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_3_to_fp16 = const()[name = tensor<string, []>("const_3_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 1, ?]> input_5_cast_fp16 = pad(constant_val = const_3_to_fp16, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<int32, [2]> concat_2x = const()[name = tensor<string, []>("concat_2x"), val = tensor<int32, [2]>([1, -1])];
+            tensor<fp16, [1, ?]> input_cast_fp16 = reshape(shape = concat_2x, x = input_5_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [1]> expand_dims_3 = const()[name = tensor<string, []>("expand_dims_3"), val = tensor<int32, [1]>([160])];
+            tensor<int32, [1]> expand_dims_4_axes_0 = const()[name = tensor<string, []>("expand_dims_4_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, ?]> expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = input_cast_fp16)[name = tensor<string, []>("expand_dims_4_cast_fp16")];
+            tensor<string, []> conv_0_pad_type_0 = const()[name = tensor<string, []>("conv_0_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> conv_0_pad_0 = const()[name = tensor<string, []>("conv_0_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_0_dilations_0 = const()[name = tensor<string, []>("conv_0_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> conv_0_groups_0 = const()[name = tensor<string, []>("conv_0_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [257, 1, 512]> expand_dims_1_to_fp16 = const()[name = tensor<string, []>("expand_dims_1_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [1, 257, ?]> conv_0_cast_fp16 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1_to_fp16, x = expand_dims_4_cast_fp16)[name = tensor<string, []>("conv_0_cast_fp16")];
+            tensor<string, []> conv_1_pad_type_0 = const()[name = tensor<string, []>("conv_1_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> conv_1_pad_0 = const()[name = tensor<string, []>("conv_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_1_dilations_0 = const()[name = tensor<string, []>("conv_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> conv_1_groups_0 = const()[name = tensor<string, []>("conv_1_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [257, 1, 512]> expand_dims_2_to_fp16 = const()[name = tensor<string, []>("expand_dims_2_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263296)))];
+            tensor<fp16, [1, 257, ?]> conv_1_cast_fp16 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2_to_fp16, x = expand_dims_4_cast_fp16)[name = tensor<string, []>("conv_1_cast_fp16")];
+            tensor<int32, []> stack_0_axis_0 = const()[name = tensor<string, []>("stack_0_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 257, ?, 2]> stack_0_cast_fp16 = stack(axis = stack_0_axis_0, values = (conv_0_cast_fp16, conv_1_cast_fp16))[name = tensor<string, []>("stack_0_cast_fp16")];
+            tensor<fp16, []> var_19_promoted_to_fp16 = const()[name = tensor<string, []>("op_19_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 257, ?, 2]> var_74_cast_fp16 = pow(x = stack_0_cast_fp16, y = var_19_promoted_to_fp16)[name = tensor<string, []>("op_74_cast_fp16")];
+            tensor<int32, [1]> var_76_axes_0 = const()[name = tensor<string, []>("op_76_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> var_76_keep_dims_0 = const()[name = tensor<string, []>("op_76_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 257, ?]> var_76_cast_fp16 = reduce_sum(axes = var_76_axes_0, keep_dims = var_76_keep_dims_0, x = var_74_cast_fp16)[name = tensor<string, []>("op_76_cast_fp16")];
+            tensor<fp16, [1, 257, ?]> x_11_cast_fp16 = identity(x = var_76_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
+            tensor<bool, []> x_13_transpose_x_0 = const()[name = tensor<string, []>("x_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_13_transpose_y_0 = const()[name = tensor<string, []>("x_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 128, 257]> const_4_to_fp16 = const()[name = tensor<string, []>("const_4_to_fp16"), val = tensor<fp16, [1, 128, 257]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526528)))];
+            tensor<fp16, [1, 128, ?]> x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = const_4_to_fp16, y = x_11_cast_fp16)[name = tensor<string, []>("x_13_cast_fp16")];
+            tensor<fp16, []> var_83_to_fp16 = const()[name = tensor<string, []>("op_83_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [1, 128, ?]> var_84_cast_fp16 = add(x = x_13_cast_fp16, y = var_83_to_fp16)[name = tensor<string, []>("op_84_cast_fp16")];
+            tensor<fp32, []> x_epsilon_0 = const()[name = tensor<string, []>("x_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
+            tensor<fp16, [1, 128, ?]> x_cast_fp16 = log(epsilon = x_epsilon_0, x = var_84_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
+            tensor<int32, [3]> var_86_shape_cast_fp16 = shape(x = x_cast_fp16)[name = tensor<string, []>("op_86_shape_cast_fp16")];
+            tensor<int32, []> gather_5_axis_0 = const()[name = tensor<string, []>("gather_5_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_5_batch_dims_0 = const()[name = tensor<string, []>("gather_5_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_5_validate_indices_0 = const()[name = tensor<string, []>("gather_5_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> var_86_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_86_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
+            tensor<uint16, []> select_5_to_uint16 = const()[name = tensor<string, []>("select_5_to_uint16"), val = tensor<uint16, []>(2)];
+            tensor<uint16, [3]> var_86_shape_cast_fp16_to_uint16 = cast(dtype = var_86_shape_cast_fp16_to_uint16_dtype_0, x = var_86_shape_cast_fp16)[name = tensor<string, []>("cast_11")];
+            tensor<uint16, []> gather_5_cast_uint16 = gather(axis = gather_5_axis_0, batch_dims = gather_5_batch_dims_0, indices = select_5_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_86_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_5_cast_uint16")];
+            tensor<string, []> gather_5_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_5_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, []> const_5 = const()[name = tensor<string, []>("const_5"), val = tensor<int32, []>(0)];
+            tensor<int32, []> const_6 = const()[name = tensor<string, []>("const_6"), val = tensor<int32, []>(1)];
+            tensor<int32, []> gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16)[name = tensor<string, []>("cast_10")];
+            tensor<int32, [?]> mask_1 = range_1d(end = gather_5_cast_uint16_to_int32, start = const_5, step = const_6)[name = tensor<string, []>("mask_1")];
+            tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = tensor<string, []>("expand_dims_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = tensor<string, []>("expand_dims_0")];
+            tensor<int32, [1]> var_91_axes_0 = const()[name = tensor<string, []>("op_91_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_91 = expand_dims(axes = var_91_axes_0, x = mel_length)[name = tensor<string, []>("op_91")];
+            tensor<bool, [1, ?]> mask = greater_equal(x = expand_dims_0, y = var_91)[name = tensor<string, []>("mask")];
+            tensor<int32, [1]> var_93_axes_0 = const()[name = tensor<string, []>("op_93_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, ?]> var_93 = expand_dims(axes = var_93_axes_0, x = mask)[name = tensor<string, []>("op_93")];
+            tensor<fp16, [1, 128, ?]> processed_signal_cast_fp16 = select(a = var_16_to_fp16, b = x_cast_fp16, cond = var_93)[name = tensor<string, []>("processed_signal_cast_fp16")];
+            tensor<string, []> processed_signal_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("processed_signal_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 128, ?]> mel = cast(dtype = processed_signal_cast_fp16_to_fp32_dtype_0, x = processed_signal_cast_fp16)[name = tensor<string, []>("cast_9")];
+        } -> (mel, mel_length);
+}
\ No newline at end of file
diff --git a/latin/4480ms/preprocessor.mlmodelc/weights/weight.bin b/latin/4480ms/preprocessor.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411
--- /dev/null
+++ b/latin/4480ms/preprocessor.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3
+size 592384
diff --git a/latin/560ms/decoder.mlmodelc/analytics/coremldata.bin b/latin/560ms/decoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..41488c8c608114babc56299e9f3532f96cb44b81
--- /dev/null
+++ b/latin/560ms/decoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55908585dbf2b684a9b8b51947317fb3f097664df47030c71f5751985e99c8a7
+size 243
diff --git a/latin/560ms/decoder.mlmodelc/coremldata.bin b/latin/560ms/decoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..81fe951046888805a7a4d02bbc0633cfe2f122ea
--- /dev/null
+++ b/latin/560ms/decoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:76a3ae4c7f2dcd6cd4bf3a1ccbe40ad0c0aecd693e1cdebd8cb85fa30f44a8b9
+size 433
diff --git a/latin/560ms/decoder.mlmodelc/model.mil b/latin/560ms/decoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..7394ffc00c3770474de6dc940583a31934b72c03
--- /dev/null
+++ b/latin/560ms/decoder.mlmodelc/model.mil
@@ -0,0 +1,64 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            tensor<int32, []> y_axis_0 = const()[name = tensor<string, []>("y_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> y_batch_dims_0 = const()[name = tensor<string, []>("y_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> y_validate_indices_0 = const()[name = tensor<string, []>("y_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [2829, 640]> module_prediction_embed_weight_to_fp16 = const()[name = tensor<string, []>("module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [2829, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<string, []> token_to_int16_dtype_0 = const()[name = tensor<string, []>("token_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = tensor<string, []>("cast_8")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = tensor<string, []>("y_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, []> split_0_num_splits_0 = const()[name = tensor<string, []>("split_0_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_0_axis_0 = const()[name = tensor<string, []>("split_0_axis_0"), val = tensor<int32, []>(0)];
+            tensor<string, []> h_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("h_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = tensor<string, []>("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = tensor<string, []>("split_0_cast_fp16")];
+            tensor<int32, []> split_1_num_splits_0 = const()[name = tensor<string, []>("split_1_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_1_axis_0 = const()[name = tensor<string, []>("split_1_axis_0"), val = tensor<int32, []>(0)];
+            tensor<string, []> c_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("c_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = tensor<string, []>("cast_6")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = tensor<string, []>("split_1_cast_fp16")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = tensor<string, []>("input_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = tensor<string, []>("input_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> input_lstm_layer_0_direction_0 = const()[name = tensor<string, []>("input_lstm_layer_0_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_lstm_layer_0_output_sequence_0 = const()[name = tensor<string, []>("input_lstm_layer_0_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_lstm_layer_0_recurrent_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_lstm_layer_0_cell_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_lstm_layer_0_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = tensor<string, []>("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3621248)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = tensor<string, []>("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6898112)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = tensor<string, []>("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10174976)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_2")];
+            tensor<fp16, [1, 1, 640]> input_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_lstm_layer_0_cast_fp16_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = tensor<string, []>("input_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = tensor<string, []>("input_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> input_direction_0 = const()[name = tensor<string, []>("input_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_output_sequence_0 = const()[name = tensor<string, []>("input_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_recurrent_activation_0 = const()[name = tensor<string, []>("input_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_cell_activation_0 = const()[name = tensor<string, []>("input_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_activation_0 = const()[name = tensor<string, []>("input_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = tensor<string, []>("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10180160)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = tensor<string, []>("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13457024)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = tensor<string, []>("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16733888)))];
+            tensor<fp16, [1, 1, 640]> input_cast_fp16_0, tensor<fp16, [1, 640]> input_cast_fp16_1, tensor<fp16, [1, 640]> input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_3_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_lstm_layer_0_cast_fp16_0)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, []> obj_3_axis_0 = const()[name = tensor<string, []>("obj_3_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_cast_fp16_1, input_cast_fp16_1))[name = tensor<string, []>("obj_3_cast_fp16")];
+            tensor<string, []> obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_3_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, []> obj_axis_0 = const()[name = tensor<string, []>("obj_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_lstm_layer_0_cast_fp16_2, input_cast_fp16_2))[name = tensor<string, []>("obj_cast_fp16")];
+            tensor<string, []> obj_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = tensor<string, []>("transpose_0_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
+            tensor<string, []> transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("transpose_0_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp16, [1, 640, 1]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = tensor<string, []>("transpose_1")];
+            tensor<fp32, [1, 640, 1]> decoder_out = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = tensor<string, []>("cast_3")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = tensor<string, []>("cast_4")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = tensor<string, []>("cast_5")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = tensor<string, []>("token_length_tmp")];
+        } -> (decoder_out, h_out, c_out);
+}
\ No newline at end of file
diff --git a/latin/560ms/decoder.mlmodelc/weights/weight.bin b/latin/560ms/decoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..b25cf03aaabac63a0f946036e3e3ce9c5f8abffb
--- /dev/null
+++ b/latin/560ms/decoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:923df635dad8f9688d46afdf1689c0913cc977f9944db30e35a78e5b41799903
+size 16739072
diff --git a/latin/560ms/decoder_joint.mlmodelc/analytics/coremldata.bin b/latin/560ms/decoder_joint.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..01a4d8e659ce38221f7d62422cc52c6dd728461e
--- /dev/null
+++ b/latin/560ms/decoder_joint.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0908aa04510db0fde1188519ae0a067789a7826613e8f96c47832ded7d0de131
+size 243
diff --git a/latin/560ms/decoder_joint.mlmodelc/coremldata.bin b/latin/560ms/decoder_joint.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..38c637f3c757ef47e33e4e5d673c089d066fe91b
--- /dev/null
+++ b/latin/560ms/decoder_joint.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e57d6034c1ff58fadcc4e89ab8ae708a153f4707a836cf7cea45b9f5daaeb77f
+size 454
diff --git a/latin/560ms/decoder_joint.mlmodelc/model.mil b/latin/560ms/decoder_joint.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..a74099fa94fca54b433d1c19fba076c3ebec7948
--- /dev/null
+++ b/latin/560ms/decoder_joint.mlmodelc/model.mil
@@ -0,0 +1,83 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [1, 1024, 1]> encoder, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            tensor<int32, []> y_axis_0 = const()[name = tensor<string, []>("y_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> y_batch_dims_0 = const()[name = tensor<string, []>("y_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> y_validate_indices_0 = const()[name = tensor<string, []>("y_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [2829, 640]> decoder_module_prediction_embed_weight_to_fp16 = const()[name = tensor<string, []>("decoder_module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [2829, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<string, []> token_to_int16_dtype_0 = const()[name = tensor<string, []>("token_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = tensor<string, []>("cast_9")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = tensor<string, []>("y_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, []> split_0_num_splits_0 = const()[name = tensor<string, []>("split_0_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_0_axis_0 = const()[name = tensor<string, []>("split_0_axis_0"), val = tensor<int32, []>(0)];
+            tensor<string, []> h_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("h_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = tensor<string, []>("cast_8")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = tensor<string, []>("split_0_cast_fp16")];
+            tensor<int32, []> split_1_num_splits_0 = const()[name = tensor<string, []>("split_1_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_1_axis_0 = const()[name = tensor<string, []>("split_1_axis_0"), val = tensor<int32, []>(0)];
+            tensor<string, []> c_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("c_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = tensor<string, []>("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = tensor<string, []>("split_1_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = tensor<string, []>("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = tensor<string, []>("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> input_5_lstm_layer_0_direction_0 = const()[name = tensor<string, []>("input_5_lstm_layer_0_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_5_lstm_layer_0_output_sequence_0 = const()[name = tensor<string, []>("input_5_lstm_layer_0_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_5_lstm_layer_0_recurrent_activation_0 = const()[name = tensor<string, []>("input_5_lstm_layer_0_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_5_lstm_layer_0_cell_activation_0 = const()[name = tensor<string, []>("input_5_lstm_layer_0_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_5_lstm_layer_0_activation_0 = const()[name = tensor<string, []>("input_5_lstm_layer_0_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = tensor<string, []>("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3621248)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = tensor<string, []>("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6898112)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = tensor<string, []>("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10174976)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = tensor<string, []>("transpose_4")];
+            tensor<fp16, [1, 1, 640]> input_5_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_5_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = tensor<string, []>("input_5_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_5_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = tensor<string, []>("input_5_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> input_5_direction_0 = const()[name = tensor<string, []>("input_5_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_5_output_sequence_0 = const()[name = tensor<string, []>("input_5_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_5_recurrent_activation_0 = const()[name = tensor<string, []>("input_5_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_5_cell_activation_0 = const()[name = tensor<string, []>("input_5_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_5_activation_0 = const()[name = tensor<string, []>("input_5_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = tensor<string, []>("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10180160)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = tensor<string, []>("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13457024)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = tensor<string, []>("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16733888)))];
+            tensor<fp16, [1, 1, 640]> input_5_cast_fp16_0, tensor<fp16, [1, 640]> input_5_cast_fp16_1, tensor<fp16, [1, 640]> input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<int32, []> obj_3_axis_0 = const()[name = tensor<string, []>("obj_3_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = tensor<string, []>("obj_3_cast_fp16")];
+            tensor<string, []> obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_3_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, []> obj_axis_0 = const()[name = tensor<string, []>("obj_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = tensor<string, []>("obj_cast_fp16")];
+            tensor<string, []> obj_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [3]> transpose_1_perm_0 = const()[name = tensor<string, []>("transpose_1_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [3]> input_7_perm_0 = const()[name = tensor<string, []>("input_7_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> encoder_to_fp16_dtype_0 = const()[name = tensor<string, []>("encoder_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [640, 1024]> joint_module_enc_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16739072)))];
+            tensor<fp16, [640]> joint_module_enc_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18049856)))];
+            tensor<fp16, [1, 1024, 1]> encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = tensor<string, []>("cast_4")];
+            tensor<fp16, [1, 1, 1024]> input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = encoder_to_fp16)[name = tensor<string, []>("transpose_2")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18051200)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18870464)))];
+            tensor<fp16, [1, 1, 640]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = tensor<string, []>("transpose_3")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_79_axes_0 = const()[name = tensor<string, []>("op_79_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_79_cast_fp16 = expand_dims(axes = var_79_axes_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_79_cast_fp16")];
+            tensor<int32, [1]> var_80_axes_0 = const()[name = tensor<string, []>("op_80_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_80_cast_fp16 = expand_dims(axes = var_80_axes_0, x = linear_1_cast_fp16)[name = tensor<string, []>("op_80_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_11_cast_fp16 = add(x = var_79_cast_fp16, y = var_80_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_13_cast_fp16 = relu(x = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
+            tensor<fp16, [2829, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [2829, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18871808)))];
+            tensor<fp16, [2829]> joint_module_joint_net_2_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [2829]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22492992)))];
+            tensor<fp16, [1, 1, 1, 2829]> linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<string, []> linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("linear_2_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 1, 1, 2829]> logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = tensor<string, []>("cast_3")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = tensor<string, []>("cast_5")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = tensor<string, []>("cast_6")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = tensor<string, []>("token_length_tmp")];
+        } -> (logits, h_out, c_out);
+}
\ No newline at end of file
diff --git a/latin/560ms/decoder_joint.mlmodelc/weights/weight.bin b/latin/560ms/decoder_joint.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..68d2b0afd28b8baf6308f64f419042ce35a5579d
--- /dev/null
+++ b/latin/560ms/decoder_joint.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5792bda2b88daae2ccb66689069da286e801f7bd168a95aa430590efa93e8e04
+size 22498714
diff --git a/latin/560ms/encoder.mlmodelc/analytics/coremldata.bin b/latin/560ms/encoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..57fa5b34d956c93cc83c850b20d1509ed7d2c77c
--- /dev/null
+++ b/latin/560ms/encoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3409887ed9eadc8689bd4c64754d95e4119152dc0b9ee1650664d2a6a689c783
+size 243
diff --git a/latin/560ms/encoder.mlmodelc/coremldata.bin b/latin/560ms/encoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..4b95c7866e6193933c2458ee77168c565a780e59
--- /dev/null
+++ b/latin/560ms/encoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea0651d7d9662d984039c554a2094ed07c281ffb127fba77a6f99847c64b0ca2
+size 572
diff --git a/latin/560ms/encoder.mlmodelc/model.mil b/latin/560ms/encoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..7fcd32956ef866e27d1041b33fb9b334a70c44c0
--- /dev/null
+++ b/latin/560ms/encoder.mlmodelc/model.mil
@@ -0,0 +1,4413 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})]
+{
+    func main<ios17>(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, 65]> mel, tensor<int32, [1]> mel_length, tensor<int32, [1]> prompt_id) {
+            tensor<int32, [4]> value_3_perm_0 = const()[name = tensor<string, []>("value_3_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<string, []> cache_channel_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_channel_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<int32, [4]> value_5_perm_0 = const()[name = tensor<string, []>("value_5_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<string, []> cache_time_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_time_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<int32, []> var_59 = const()[name = tensor<string, []>("op_59"), val = tensor<int32, []>(-1)];
+            tensor<int32, []> var_68 = const()[name = tensor<string, []>("op_68"), val = tensor<int32, []>(1)];
+            tensor<int32, [3]> x_1_perm_0 = const()[name = tensor<string, []>("x_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> mel_to_fp16_dtype_0 = const()[name = tensor<string, []>("mel_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<int32, [1]> tensor_1_axes_0 = const()[name = tensor<string, []>("tensor_1_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 128, 65]> mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = tensor<string, []>("cast_18")];
+            tensor<fp16, [1, 65, 128]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = mel_to_fp16)[name = tensor<string, []>("transpose_367")];
+            tensor<fp16, [1, 1, 65, 128]> tensor_1_cast_fp16 = expand_dims(axes = tensor_1_axes_0, x = x_1_cast_fp16)[name = tensor<string, []>("tensor_1_cast_fp16")];
+            tensor<int32, [1, 65]> expand_dims_0 = const()[name = tensor<string, []>("expand_dims_0"), val = tensor<int32, [1, 65]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64]])];
+            tensor<int32, [1]> var_137_axes_0 = const()[name = tensor<string, []>("op_137_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_137 = expand_dims(axes = var_137_axes_0, x = mel_length)[name = tensor<string, []>("op_137")];
+            tensor<bool, [1, 65]> time_mask_1 = less(x = expand_dims_0, y = var_137)[name = tensor<string, []>("time_mask_1")];
+            tensor<int32, [1]> var_139_axes_0 = const()[name = tensor<string, []>("op_139_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 65, 1]> var_139 = expand_dims(axes = var_139_axes_0, x = time_mask_1)[name = tensor<string, []>("op_139")];
+            tensor<int32, [3]> var_141_reps_0 = const()[name = tensor<string, []>("op_141_reps_0"), val = tensor<int32, [3]>([1, 1, 128])];
+            tensor<bool, [1, 65, 128]> var_141 = tile(reps = var_141_reps_0, x = var_139)[name = tensor<string, []>("op_141")];
+            tensor<int32, [1]> var_147_axes_0 = const()[name = tensor<string, []>("op_147_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<string, []> cast_4_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_4_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 65, 128]> var_141_to_fp16 = cast(dtype = cast_4_to_fp16_dtype_0, x = var_141)[name = tensor<string, []>("cast_17")];
+            tensor<fp16, [1, 1, 65, 128]> var_147_cast_fp16 = expand_dims(axes = var_147_axes_0, x = var_141_to_fp16)[name = tensor<string, []>("op_147_cast_fp16")];
+            tensor<fp16, [1, 1, 65, 128]> input_1_cast_fp16 = mul(x = tensor_1_cast_fp16, y = var_147_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<int32, [8]> input_3_pad_0 = const()[name = tensor<string, []>("input_3_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_3_mode_0 = const()[name = tensor<string, []>("input_3_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_9_to_fp16 = const()[name = tensor<string, []>("const_9_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 1, 68, 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 = tensor<string, []>("input_3_cast_fp16")];
+            tensor<string, []> tensor_3_pad_type_0 = const()[name = tensor<string, []>("tensor_3_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_3_strides_0 = const()[name = tensor<string, []>("tensor_3_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, [4]> tensor_3_pad_0 = const()[name = tensor<string, []>("tensor_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_3_dilations_0 = const()[name = tensor<string, []>("tensor_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> tensor_3_groups_0 = const()[name = tensor<string, []>("tensor_3_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_0_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_pre_encode_conv_0_weight_to_fp16_quantized"), quantized_data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64))), scale = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2752))), zero_point = tensor<int8, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2432)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3328)))];
+            tensor<fp16, [1, 256, 33, 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 = tensor<string, []>("tensor_3_cast_fp16")];
+            tensor<string, []> cast_2_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> var_160_promoted_to_fp16 = const()[name = tensor<string, []>("op_160_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> mel_length_to_fp16 = cast(dtype = cast_2_to_fp16_dtype_0, x = mel_length)[name = tensor<string, []>("cast_16")];
+            tensor<fp16, [1]> var_161_cast_fp16 = add(x = mel_length_to_fp16, y = var_160_promoted_to_fp16)[name = tensor<string, []>("op_161_cast_fp16")];
+            tensor<fp16, []> var_162_promoted_to_fp16 = const()[name = tensor<string, []>("op_162_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> var_163_cast_fp16 = add(x = var_161_cast_fp16, y = var_162_promoted_to_fp16)[name = tensor<string, []>("op_163_cast_fp16")];
+            tensor<fp16, []> var_164_promoted_to_fp16 = const()[name = tensor<string, []>("op_164_promoted_to_fp16"), val = tensor<fp16, []>(0x1.8p+1)];
+            tensor<fp16, [1]> var_165_cast_fp16 = sub(x = var_163_cast_fp16, y = var_164_promoted_to_fp16)[name = tensor<string, []>("op_165_cast_fp16")];
+            tensor<fp16, []> var_56_promoted_to_fp16 = const()[name = tensor<string, []>("op_56_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> floor_div_0_cast_fp16 = floor_div(x = var_165_cast_fp16, y = var_56_promoted_to_fp16)[name = tensor<string, []>("floor_div_0_cast_fp16")];
+            tensor<fp16, []> var_167_promoted_to_fp16 = const()[name = tensor<string, []>("op_167_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_3_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_167_promoted_to_fp16)[name = tensor<string, []>("current_lengths_3_cast_fp16")];
+            tensor<string, []> cast_5_dtype_0 = const()[name = tensor<string, []>("cast_5_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 33]> expand_dims_1 = const()[name = tensor<string, []>("expand_dims_1"), val = tensor<int32, [1, 33]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32]])];
+            tensor<int32, [1]> var_176_axes_0 = const()[name = tensor<string, []>("op_176_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_3_cast_fp16_to_int32 = cast(dtype = cast_5_dtype_0, x = current_lengths_3_cast_fp16)[name = tensor<string, []>("cast_15")];
+            tensor<int32, [1, 1]> var_176 = expand_dims(axes = var_176_axes_0, x = current_lengths_3_cast_fp16_to_int32)[name = tensor<string, []>("op_176")];
+            tensor<bool, [1, 33]> time_mask_3 = less(x = expand_dims_1, y = var_176)[name = tensor<string, []>("time_mask_3")];
+            tensor<int32, [1]> var_178_axes_0 = const()[name = tensor<string, []>("op_178_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 33, 1]> var_178 = expand_dims(axes = var_178_axes_0, x = time_mask_3)[name = tensor<string, []>("op_178")];
+            tensor<int32, [3]> var_180_reps_0 = const()[name = tensor<string, []>("op_180_reps_0"), val = tensor<int32, [3]>([1, 1, 65])];
+            tensor<bool, [1, 33, 65]> var_180 = tile(reps = var_180_reps_0, x = var_178)[name = tensor<string, []>("op_180")];
+            tensor<int32, [1]> var_186_axes_0 = const()[name = tensor<string, []>("op_186_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<string, []> cast_6_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_6_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 33, 65]> var_180_to_fp16 = cast(dtype = cast_6_to_fp16_dtype_0, x = var_180)[name = tensor<string, []>("cast_14")];
+            tensor<fp16, [1, 1, 33, 65]> var_186_cast_fp16 = expand_dims(axes = var_186_axes_0, x = var_180_to_fp16)[name = tensor<string, []>("op_186_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_3_reps_0 = const()[name = tensor<string, []>("expanded_mask_3_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 33, 65]> expanded_mask_3_cast_fp16 = tile(reps = expanded_mask_3_reps_0, x = var_186_cast_fp16)[name = tensor<string, []>("expanded_mask_3_cast_fp16")];
+            tensor<fp16, [1, 256, 33, 65]> input_5_cast_fp16 = mul(x = tensor_3_cast_fp16, y = expanded_mask_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [1, 256, 33, 65]> tensor_5_cast_fp16 = relu(x = input_5_cast_fp16)[name = tensor<string, []>("tensor_5_cast_fp16")];
+            tensor<fp16, [1, 256, 33, 65]> input_7_cast_fp16 = mul(x = tensor_5_cast_fp16, y = expanded_mask_3_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<int32, [8]> input_9_pad_0 = const()[name = tensor<string, []>("input_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_9_mode_0 = const()[name = tensor<string, []>("input_9_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_23_to_fp16 = const()[name = tensor<string, []>("const_23_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 256, 36, 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 = tensor<string, []>("input_9_cast_fp16")];
+            tensor<string, []> tensor_7_pad_type_0 = const()[name = tensor<string, []>("tensor_7_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_7_strides_0 = const()[name = tensor<string, []>("tensor_7_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, []> tensor_7_groups_0 = const()[name = tensor<string, []>("tensor_7_groups_0"), val = tensor<int32, []>(256)];
+            tensor<int32, [4]> tensor_7_pad_0 = const()[name = tensor<string, []>("tensor_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_7_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_pre_encode_conv_2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3904))), scale = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6272))), zero_point = tensor<int8, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2432)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6848)))];
+            tensor<fp16, [1, 256, 17, 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 = tensor<string, []>("tensor_7_cast_fp16")];
+            tensor<fp16, []> var_208_promoted_to_fp16 = const()[name = tensor<string, []>("op_208_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> var_209_cast_fp16 = add(x = current_lengths_3_cast_fp16, y = var_208_promoted_to_fp16)[name = tensor<string, []>("op_209_cast_fp16")];
+            tensor<fp16, []> var_210_promoted_to_fp16 = const()[name = tensor<string, []>("op_210_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> var_211_cast_fp16 = add(x = var_209_cast_fp16, y = var_210_promoted_to_fp16)[name = tensor<string, []>("op_211_cast_fp16")];
+            tensor<fp16, []> var_212_promoted_to_fp16 = const()[name = tensor<string, []>("op_212_promoted_to_fp16"), val = tensor<fp16, []>(0x1.8p+1)];
+            tensor<fp16, [1]> var_213_cast_fp16 = sub(x = var_211_cast_fp16, y = var_212_promoted_to_fp16)[name = tensor<string, []>("op_213_cast_fp16")];
+            tensor<fp16, []> var_56_promoted_1_to_fp16 = const()[name = tensor<string, []>("op_56_promoted_1_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> floor_div_1_cast_fp16 = floor_div(x = var_213_cast_fp16, y = var_56_promoted_1_to_fp16)[name = tensor<string, []>("floor_div_1_cast_fp16")];
+            tensor<fp16, []> var_215_promoted_to_fp16 = const()[name = tensor<string, []>("op_215_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_5_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_215_promoted_to_fp16)[name = tensor<string, []>("current_lengths_5_cast_fp16")];
+            tensor<string, []> cast_7_dtype_0 = const()[name = tensor<string, []>("cast_7_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 17]> expand_dims_2 = const()[name = tensor<string, []>("expand_dims_2"), val = tensor<int32, [1, 17]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]])];
+            tensor<int32, [1]> var_224_axes_0 = const()[name = tensor<string, []>("op_224_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_5_cast_fp16_to_int32 = cast(dtype = cast_7_dtype_0, x = current_lengths_5_cast_fp16)[name = tensor<string, []>("cast_13")];
+            tensor<int32, [1, 1]> var_224 = expand_dims(axes = var_224_axes_0, x = current_lengths_5_cast_fp16_to_int32)[name = tensor<string, []>("op_224")];
+            tensor<bool, [1, 17]> time_mask_5 = less(x = expand_dims_2, y = var_224)[name = tensor<string, []>("time_mask_5")];
+            tensor<int32, [1]> var_226_axes_0 = const()[name = tensor<string, []>("op_226_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 17, 1]> var_226 = expand_dims(axes = var_226_axes_0, x = time_mask_5)[name = tensor<string, []>("op_226")];
+            tensor<int32, [3]> var_228_reps_0 = const()[name = tensor<string, []>("op_228_reps_0"), val = tensor<int32, [3]>([1, 1, 33])];
+            tensor<bool, [1, 17, 33]> var_228 = tile(reps = var_228_reps_0, x = var_226)[name = tensor<string, []>("op_228")];
+            tensor<int32, [1]> var_234_axes_0 = const()[name = tensor<string, []>("op_234_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<string, []> cast_8_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_8_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 17, 33]> var_228_to_fp16 = cast(dtype = cast_8_to_fp16_dtype_0, x = var_228)[name = tensor<string, []>("cast_12")];
+            tensor<fp16, [1, 1, 17, 33]> var_234_cast_fp16 = expand_dims(axes = var_234_axes_0, x = var_228_to_fp16)[name = tensor<string, []>("op_234_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_7_reps_0 = const()[name = tensor<string, []>("expanded_mask_7_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 17, 33]> expanded_mask_7_cast_fp16 = tile(reps = expanded_mask_7_reps_0, x = var_234_cast_fp16)[name = tensor<string, []>("expanded_mask_7_cast_fp16")];
+            tensor<fp16, [1, 256, 17, 33]> input_11_cast_fp16 = mul(x = tensor_7_cast_fp16, y = expanded_mask_7_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<string, []> tensor_9_pad_type_0 = const()[name = tensor<string, []>("tensor_9_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_9_strides_0 = const()[name = tensor<string, []>("tensor_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_9_pad_0 = const()[name = tensor<string, []>("tensor_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_9_dilations_0 = const()[name = tensor<string, []>("tensor_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> tensor_9_groups_0 = const()[name = tensor<string, []>("tensor_9_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_3_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_pre_encode_conv_3_weight_to_fp16_quantized"), quantized_data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7424))), scale = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73024))), zero_point = tensor<int8, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2432)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73600)))];
+            tensor<fp16, [1, 256, 17, 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 = tensor<string, []>("tensor_9_cast_fp16")];
+            tensor<fp16, [1, 256, 17, 33]> input_13_cast_fp16 = mul(x = tensor_9_cast_fp16, y = expanded_mask_7_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
+            tensor<fp16, [1, 256, 17, 33]> tensor_11_cast_fp16 = relu(x = input_13_cast_fp16)[name = tensor<string, []>("tensor_11_cast_fp16")];
+            tensor<fp16, [1, 256, 17, 33]> input_15_cast_fp16 = mul(x = tensor_11_cast_fp16, y = expanded_mask_7_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
+            tensor<int32, [8]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_17_mode_0 = const()[name = tensor<string, []>("input_17_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_41_to_fp16 = const()[name = tensor<string, []>("const_41_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 256, 20, 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 = tensor<string, []>("input_17_cast_fp16")];
+            tensor<string, []> tensor_13_pad_type_0 = const()[name = tensor<string, []>("tensor_13_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_13_strides_0 = const()[name = tensor<string, []>("tensor_13_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, []> tensor_13_groups_0 = const()[name = tensor<string, []>("tensor_13_groups_0"), val = tensor<int32, []>(256)];
+            tensor<int32, [4]> tensor_13_pad_0 = const()[name = tensor<string, []>("tensor_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_13_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_pre_encode_conv_5_weight_to_fp16_quantized"), quantized_data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74176))), scale = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76544))), zero_point = tensor<int8, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2432)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77120)))];
+            tensor<fp16, [1, 256, 9, 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 = tensor<string, []>("tensor_13_cast_fp16")];
+            tensor<fp16, []> var_271_promoted_to_fp16 = const()[name = tensor<string, []>("op_271_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> var_272_cast_fp16 = add(x = current_lengths_5_cast_fp16, y = var_271_promoted_to_fp16)[name = tensor<string, []>("op_272_cast_fp16")];
+            tensor<fp16, []> var_273_promoted_to_fp16 = const()[name = tensor<string, []>("op_273_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> var_274_cast_fp16 = add(x = var_272_cast_fp16, y = var_273_promoted_to_fp16)[name = tensor<string, []>("op_274_cast_fp16")];
+            tensor<fp16, []> var_275_promoted_to_fp16 = const()[name = tensor<string, []>("op_275_promoted_to_fp16"), val = tensor<fp16, []>(0x1.8p+1)];
+            tensor<fp16, [1]> var_276_cast_fp16 = sub(x = var_274_cast_fp16, y = var_275_promoted_to_fp16)[name = tensor<string, []>("op_276_cast_fp16")];
+            tensor<fp16, []> var_56_promoted_2_to_fp16 = const()[name = tensor<string, []>("op_56_promoted_2_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1]> floor_div_2_cast_fp16 = floor_div(x = var_276_cast_fp16, y = var_56_promoted_2_to_fp16)[name = tensor<string, []>("floor_div_2_cast_fp16")];
+            tensor<fp16, []> var_278_promoted_to_fp16 = const()[name = tensor<string, []>("op_278_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_278_promoted_to_fp16)[name = tensor<string, []>("current_lengths_cast_fp16")];
+            tensor<string, []> cast_9_dtype_0 = const()[name = tensor<string, []>("cast_9_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 9]> expand_dims_3 = const()[name = tensor<string, []>("expand_dims_3"), val = tensor<int32, [1, 9]>([[0, 1, 2, 3, 4, 5, 6, 7, 8]])];
+            tensor<int32, [1]> var_287_axes_0 = const()[name = tensor<string, []>("op_287_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_cast_fp16_to_int32 = cast(dtype = cast_9_dtype_0, x = current_lengths_cast_fp16)[name = tensor<string, []>("cast_11")];
+            tensor<int32, [1, 1]> var_287 = expand_dims(axes = var_287_axes_0, x = current_lengths_cast_fp16_to_int32)[name = tensor<string, []>("op_287")];
+            tensor<bool, [1, 9]> time_mask = less(x = expand_dims_3, y = var_287)[name = tensor<string, []>("time_mask")];
+            tensor<int32, [1]> var_289_axes_0 = const()[name = tensor<string, []>("op_289_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 9, 1]> var_289 = expand_dims(axes = var_289_axes_0, x = time_mask)[name = tensor<string, []>("op_289")];
+            tensor<int32, [3]> var_291_reps_0 = const()[name = tensor<string, []>("op_291_reps_0"), val = tensor<int32, [3]>([1, 1, 17])];
+            tensor<bool, [1, 9, 17]> var_291 = tile(reps = var_291_reps_0, x = var_289)[name = tensor<string, []>("op_291")];
+            tensor<int32, [1]> var_297_axes_0 = const()[name = tensor<string, []>("op_297_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<string, []> cast_10_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_10_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 9, 17]> var_291_to_fp16 = cast(dtype = cast_10_to_fp16_dtype_0, x = var_291)[name = tensor<string, []>("cast_10")];
+            tensor<fp16, [1, 1, 9, 17]> var_297_cast_fp16 = expand_dims(axes = var_297_axes_0, x = var_291_to_fp16)[name = tensor<string, []>("op_297_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_13_reps_0 = const()[name = tensor<string, []>("expanded_mask_13_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 9, 17]> expanded_mask_13_cast_fp16 = tile(reps = expanded_mask_13_reps_0, x = var_297_cast_fp16)[name = tensor<string, []>("expanded_mask_13_cast_fp16")];
+            tensor<fp16, [1, 256, 9, 17]> input_19_cast_fp16 = mul(x = tensor_13_cast_fp16, y = expanded_mask_13_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
+            tensor<string, []> tensor_15_pad_type_0 = const()[name = tensor<string, []>("tensor_15_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_15_strides_0 = const()[name = tensor<string, []>("tensor_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_15_pad_0 = const()[name = tensor<string, []>("tensor_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_15_dilations_0 = const()[name = tensor<string, []>("tensor_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> tensor_15_groups_0 = const()[name = tensor<string, []>("tensor_15_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_6_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_pre_encode_conv_6_weight_to_fp16_quantized"), quantized_data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77696))), scale = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143296))), zero_point = tensor<int8, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2432)))];
+            tensor<fp16, [256]> encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143872)))];
+            tensor<fp16, [1, 256, 9, 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 = tensor<string, []>("tensor_15_cast_fp16")];
+            tensor<fp16, [1, 256, 9, 17]> input_21_cast_fp16 = mul(x = tensor_15_cast_fp16, y = expanded_mask_13_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
+            tensor<fp16, [1, 256, 9, 17]> tensor_cast_fp16 = relu(x = input_21_cast_fp16)[name = tensor<string, []>("tensor_cast_fp16")];
+            tensor<fp16, [1, 256, 9, 17]> x_3_cast_fp16 = mul(x = tensor_cast_fp16, y = expanded_mask_13_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
+            tensor<int32, [4]> var_331_perm_0 = const()[name = tensor<string, []>("op_331_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_332 = const()[name = tensor<string, []>("op_332"), val = tensor<int32, [3]>([1, 9, -1])];
+            tensor<fp16, [1, 9, 256, 17]> var_331_cast_fp16 = transpose(perm = var_331_perm_0, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_366")];
+            tensor<fp16, [1, 9, 4352]> input_23_cast_fp16 = reshape(shape = var_332, x = var_331_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
+            tensor<fp16, [1024, 4352]> encoder_pre_encode_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_pre_encode_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4352]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144448))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4602048))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_pre_encode_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_pre_encode_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4604160)))];
+            tensor<fp16, [1, 9, 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 = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<int32, [3]> var_342_begin_0 = const()[name = tensor<string, []>("op_342_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_342_end_0 = const()[name = tensor<string, []>("op_342_end_0"), val = tensor<int32, [3]>([1, 9, 1024])];
+            tensor<bool, [3]> var_342_end_mask_0 = const()[name = tensor<string, []>("op_342_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 7, 1024]> var_342_cast_fp16 = slice_by_index(begin = var_342_begin_0, end = var_342_end_0, end_mask = var_342_end_mask_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_342_cast_fp16")];
+            tensor<int32, []> var_344 = const()[name = tensor<string, []>("op_344"), val = tensor<int32, []>(2)];
+            tensor<int32, [1]> var_345 = sub(x = current_lengths_cast_fp16_to_int32, y = var_344)[name = tensor<string, []>("op_345")];
+            tensor<string, []> var_345_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_345_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> var_62_promoted_to_fp16 = const()[name = tensor<string, []>("op_62_promoted_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, []> const_61_to_fp16 = const()[name = tensor<string, []>("const_61_to_fp16"), val = tensor<fp16, []>(inf)];
+            tensor<fp16, [1]> var_345_to_fp16 = cast(dtype = var_345_promoted_to_fp16_dtype_0, x = var_345)[name = tensor<string, []>("cast_9")];
+            tensor<fp16, [1]> clip_0_cast_fp16 = clip(alpha = var_62_promoted_to_fp16, beta = const_61_to_fp16, x = var_345_to_fp16)[name = tensor<string, []>("clip_0_cast_fp16")];
+            tensor<int32, [1]> max_audio_length_1 = const()[name = tensor<string, []>("max_audio_length_1"), val = tensor<int32, [1]>([7])];
+            tensor<fp16, []> var_361_promoted_to_fp16 = const()[name = tensor<string, []>("op_361_promoted_to_fp16"), val = tensor<fp16, []>(0x1.5p+5)];
+            tensor<fp16, [1]> padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_361_promoted_to_fp16)[name = tensor<string, []>("padding_length_cast_fp16")];
+            tensor<int32, []> const_63 = const()[name = tensor<string, []>("const_63"), val = tensor<int32, []>(-1)];
+            tensor<int32, [1]> var_363 = mul(x = cache_len, y = const_63)[name = tensor<string, []>("op_363")];
+            tensor<int32, []> var_364 = const()[name = tensor<string, []>("op_364"), val = tensor<int32, []>(42)];
+            tensor<int32, [1]> offset = add(x = var_363, y = var_364)[name = tensor<string, []>("offset")];
+            tensor<int32, [1]> var_404_axes_0 = const()[name = tensor<string, []>("op_404_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1]> var_404_cast_fp16 = expand_dims(axes = var_404_axes_0, x = padding_length_cast_fp16)[name = tensor<string, []>("op_404_cast_fp16")];
+            tensor<fp16, [1, 49]> var_403_promoted_to_fp16 = const()[name = tensor<string, []>("op_403_promoted_to_fp16"), val = tensor<fp16, [1, 49]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4606272)))];
+            tensor<bool, [1, 49]> pad_mask_1_cast_fp16 = less(x = var_403_promoted_to_fp16, y = var_404_cast_fp16)[name = tensor<string, []>("pad_mask_1_cast_fp16")];
+            tensor<int32, [1, 49]> expand_dims_5 = const()[name = tensor<string, []>("expand_dims_5"), val = tensor<int32, [1, 49]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48]])];
+            tensor<int32, [1]> var_410_axes_0 = const()[name = tensor<string, []>("op_410_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 1]> var_410 = expand_dims(axes = var_410_axes_0, x = offset)[name = tensor<string, []>("op_410")];
+            tensor<bool, [1, 49]> pad_mask_off = greater_equal(x = expand_dims_5, y = var_410)[name = tensor<string, []>("pad_mask_off")];
+            tensor<bool, [1, 49]> pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = tensor<string, []>("pad_mask_3")];
+            tensor<int32, [1]> var_413_axes_0 = const()[name = tensor<string, []>("op_413_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 49]> var_413 = expand_dims(axes = var_413_axes_0, x = pad_mask_3)[name = tensor<string, []>("op_413")];
+            tensor<int32, [3]> var_414 = const()[name = tensor<string, []>("op_414"), val = tensor<int32, [3]>([1, 49, 1])];
+            tensor<bool, [1, 49, 49]> pad_mask_for_att_mask_1 = tile(reps = var_414, x = var_413)[name = tensor<string, []>("pad_mask_for_att_mask_1")];
+            tensor<int32, [3]> var_416_perm_0 = const()[name = tensor<string, []>("op_416_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<bool, [1, 49, 49]> var_416 = transpose(perm = var_416_perm_0, x = pad_mask_for_att_mask_1)[name = tensor<string, []>("transpose_365")];
+            tensor<bool, [1, 49, 49]> pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_416)[name = tensor<string, []>("pad_mask_for_att_mask")];
+            tensor<bool, [1, 49, 49]> const_71 = const()[name = tensor<string, []>("const_71"), val = tensor<bool, [1, 49, 49]>([[[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], [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], [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], [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], [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], [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], [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], [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], [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], [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], [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], [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], [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], [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], [true, true, true, true, 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], [true, true, true, true, 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], [true, true, true, true, 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], [true, true, true, true, 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], [true, true, true, true, 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], [true, true, true, true, 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], [true, true, true, true, 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], [true, true, true, true, 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], [true, true, true, true, 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], [true, true, true, true, 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], [true, true, true, true, 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], [true, true, true, true, 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], [true, true, true, true, 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], [true, true, true, true, 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], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, 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], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, 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], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, 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], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, 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], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, 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], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, 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], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, 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], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, 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], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, 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], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, 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], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, 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], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, 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], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, 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], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, 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], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, 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, 49, 49]> att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = tensor<string, []>("att_mask_9")];
+            tensor<bool, [1, 49, 49]> att_mask = logical_not(x = att_mask_9)[name = tensor<string, []>("att_mask")];
+            tensor<bool, [1, 49]> pad_mask_5 = logical_not(x = pad_mask_3)[name = tensor<string, []>("pad_mask_5")];
+            tensor<int32, [2]> pad_mask_begin_0 = const()[name = tensor<string, []>("pad_mask_begin_0"), val = tensor<int32, [2]>([0, 42])];
+            tensor<int32, [2]> pad_mask_end_0 = const()[name = tensor<string, []>("pad_mask_end_0"), val = tensor<int32, [2]>([1, 49])];
+            tensor<bool, [2]> pad_mask_end_mask_0 = const()[name = tensor<string, []>("pad_mask_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<bool, [1, 7]> 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 = tensor<string, []>("pad_mask")];
+            tensor<int32, [3]> mask_9_begin_0 = const()[name = tensor<string, []>("mask_9_begin_0"), val = tensor<int32, [3]>([0, 42, 0])];
+            tensor<int32, [3]> mask_9_end_0 = const()[name = tensor<string, []>("mask_9_end_0"), val = tensor<int32, [3]>([1, 49, 49])];
+            tensor<bool, [3]> mask_9_end_mask_0 = const()[name = tensor<string, []>("mask_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<bool, [1, 7, 49]> 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 = tensor<string, []>("mask_9")];
+            tensor<int32, [4]> cache_1_begin_0 = const()[name = tensor<string, []>("cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_1_end_0 = const()[name = tensor<string, []>("cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 42, 1024])];
+            tensor<bool, [4]> cache_1_end_mask_0 = const()[name = tensor<string, []>("cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_1_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cast_8")];
+            tensor<fp16, [24, 1, 42, 1024]> value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = cache_channel_to_fp16)[name = tensor<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 = tensor<string, []>("cache_1_cast_fp16")];
+            tensor<int32, [4]> cache_3_begin_0 = const()[name = tensor<string, []>("cache_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_3_end_0 = const()[name = tensor<string, []>("cache_3_end_0"), val = tensor<int32, [4]>([1, 1, 1024, 8])];
+            tensor<bool, [4]> cache_3_end_mask_0 = const()[name = tensor<string, []>("cache_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_3_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cast_7")];
+            tensor<fp16, [24, 1, 1024, 8]> value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = cache_time_to_fp16)[name = tensor<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 = tensor<string, []>("cache_3_cast_fp16")];
+            tensor<int32, [1]> input_27_axes_0 = const()[name = tensor<string, []>("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4606464)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4608576)))];
+            tensor<fp16, []> var_42_to_fp16 = const()[name = tensor<string, []>("op_42_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 7, 1024]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_342_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4610688))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8809216))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805056)))];
+            tensor<fp16, [4096]> encoder_layers_0_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8817472)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8825728))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13020096))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_0_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13022208)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<fp16, []> var_455_to_fp16 = const()[name = tensor<string, []>("op_455_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_456_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_455_to_fp16)[name = tensor<string, []>("op_456_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_37_cast_fp16 = add(x = var_342_cast_fp16, y = var_456_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
+            tensor<int32, [1]> key_1_axes_0 = const()[name = tensor<string, []>("key_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13024320)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13026432)))];
+            tensor<fp16, [1, 7, 1024]> key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")];
+            tensor<bool, []> input_39_interleave_0 = const()[name = tensor<string, []>("input_39_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_39_cast_fp16 = concat(axis = var_68, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = tensor<string, []>("input_39_cast_fp16")];
+            tensor<int32, [3]> var_478_begin_0 = const()[name = tensor<string, []>("op_478_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_478_end_0 = const()[name = tensor<string, []>("op_478_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_478_end_mask_0 = const()[name = tensor<string, []>("op_478_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_478_cast_fp16 = slice_by_index(begin = var_478_begin_0, end = var_478_end_0, end_mask = var_478_end_mask_0, x = cache_1_cast_fp16)[name = tensor<string, []>("op_478_cast_fp16")];
+            tensor<bool, []> var_484_interleave_0 = const()[name = tensor<string, []>("op_484_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_484_cast_fp16 = concat(axis = var_68, interleave = var_484_interleave_0, values = (var_478_cast_fp16, key_1_cast_fp16))[name = tensor<string, []>("op_484_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13028544))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14077184))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14079296)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_3_cast_fp16")];
+            tensor<int32, [4]> var_489 = const()[name = tensor<string, []>("op_489"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_1_cast_fp16 = reshape(shape = var_489, x = linear_3_cast_fp16)[name = tensor<string, []>("q_1_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14081408))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15130048))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15132160)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_4_cast_fp16")];
+            tensor<int32, [4]> var_494 = const()[name = tensor<string, []>("op_494"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_1_cast_fp16 = reshape(shape = var_494, x = linear_4_cast_fp16)[name = tensor<string, []>("k_1_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15134272))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16182912))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16185024)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_5_cast_fp16")];
+            tensor<int32, [4]> var_499 = const()[name = tensor<string, []>("op_499"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_1_cast_fp16 = reshape(shape = var_499, x = linear_5_cast_fp16)[name = tensor<string, []>("v_1_cast_fp16")];
+            tensor<int32, [4]> value_9_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16187136)))];
+            tensor<fp16, [1, 7, 8, 128]> var_512_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_512_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16189248)))];
+            tensor<fp16, [1, 7, 8, 128]> var_514_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_514_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_1_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_1_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_7_transpose_x_0 = const()[name = tensor<string, []>("x_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_7_transpose_y_0 = const()[name = tensor<string, []>("x_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_516_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_516_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16191360))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16290944))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16290752)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_514_cast_fp16)[name = tensor<string, []>("transpose_362")];
+            tensor<fp16, [1, 8, 7, 97]> 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_516_to_fp16_quantized)[name = tensor<string, []>("x_7_cast_fp16")];
+            tensor<int32, [8]> x_9_pad_0 = const()[name = tensor<string, []>("x_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_9_mode_0 = const()[name = tensor<string, []>("x_9_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_79_to_fp16 = const()[name = tensor<string, []>("const_79_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_9_cast_fp16")];
+            tensor<int32, [4]> var_524 = const()[name = tensor<string, []>("op_524"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_11_cast_fp16 = reshape(shape = var_524, x = x_9_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
+            tensor<int32, [4]> var_528_begin_0 = const()[name = tensor<string, []>("op_528_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_528_end_0 = const()[name = tensor<string, []>("op_528_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_528_end_mask_0 = const()[name = tensor<string, []>("op_528_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_528_cast_fp16 = slice_by_index(begin = var_528_begin_0, end = var_528_end_0, end_mask = var_528_end_mask_0, x = x_11_cast_fp16)[name = tensor<string, []>("op_528_cast_fp16")];
+            tensor<int32, [4]> var_529 = const()[name = tensor<string, []>("op_529"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_1_cast_fp16 = reshape(shape = var_529, x = var_528_cast_fp16)[name = tensor<string, []>("matrix_bd_1_cast_fp16")];
+            tensor<bool, []> matrix_ac_1_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_1_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_96_perm_0 = const()[name = tensor<string, []>("transpose_96_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_97_perm_0 = const()[name = tensor<string, []>("transpose_97_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1_cast_fp16)[name = tensor<string, []>("transpose_360")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_96 = transpose(perm = transpose_96_perm_0, x = var_512_cast_fp16)[name = tensor<string, []>("transpose_361")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_1_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_3_begin_0 = const()[name = tensor<string, []>("matrix_bd_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_3_end_0 = const()[name = tensor<string, []>("matrix_bd_3_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_3_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_3_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_538_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = tensor<string, []>("op_538_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_1_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_1_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_1_cast_fp16 = mul(x = var_538_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_1_cast_fp16")];
+            tensor<int32, [1]> mask_11_axes_0 = const()[name = tensor<string, []>("mask_11_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 7, 49]> mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = tensor<string, []>("mask_11")];
+            tensor<fp16, []> var_45_to_fp16 = const()[name = tensor<string, []>("op_45_to_fp16"), val = tensor<fp16, []>(-0x1.388p+13)];
+            tensor<fp16, [1, 8, 7, 49]> scores_3_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_3_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_544_cast_fp16 = softmax(axis = var_59, x = scores_3_cast_fp16)[name = tensor<string, []>("op_544_cast_fp16")];
+            tensor<fp16, []> var_44_to_fp16 = const()[name = tensor<string, []>("op_44_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 49]> input_41_cast_fp16 = select(a = var_44_to_fp16, b = var_544_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_41_cast_fp16")];
+            tensor<bool, []> x_13_transpose_x_0 = const()[name = tensor<string, []>("x_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_13_transpose_y_0 = const()[name = tensor<string, []>("x_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_1_cast_fp16)[name = tensor<string, []>("transpose_359")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_13_cast_fp16")];
+            tensor<int32, [4]> var_548_perm_0 = const()[name = tensor<string, []>("op_548_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_549 = const()[name = tensor<string, []>("op_549"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_548_cast_fp16 = transpose(perm = var_548_perm_0, x = x_13_cast_fp16)[name = tensor<string, []>("transpose_358")];
+            tensor<fp16, [1, 7, 1024]> input_43_cast_fp16 = reshape(shape = var_549, x = var_548_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16291264))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17339904))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17342016)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_7_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
+            tensor<int32, [1]> x_17_axes_0 = const()[name = tensor<string, []>("x_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17344128)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17346240)))];
+            tensor<fp16, [1, 7, 1024]> x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("x_17_cast_fp16")];
+            tensor<int32, [3]> input_49_perm_0 = const()[name = tensor<string, []>("input_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_51_pad_type_0 = const()[name = tensor<string, []>("input_51_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_51_strides_0 = const()[name = tensor<string, []>("input_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_51_pad_0 = const()[name = tensor<string, []>("input_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_51_dilations_0 = const()[name = tensor<string, []>("input_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_51_groups_0 = const()[name = tensor<string, []>("input_51_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17348352))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19447680))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = tensor<string, []>("transpose_357")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_51_cast_fp16")];
+            tensor<int32, []> x_19_split_num_splits_0 = const()[name = tensor<string, []>("x_19_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_19_split_axis_0 = const()[name = tensor<string, []>("x_19_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_19_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_19_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = tensor<string, []>("x_19_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")];
+            tensor<int32, [1]> var_575_axes_0 = const()[name = tensor<string, []>("op_575_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 7]> var_575 = expand_dims(axes = var_575_axes_0, x = pad_mask)[name = tensor<string, []>("op_575")];
+            tensor<fp16, [1, 1024, 7]> input_53_cast_fp16 = select(a = var_44_to_fp16, b = x_19_cast_fp16, cond = var_575)[name = tensor<string, []>("input_53_cast_fp16")];
+            tensor<bool, []> new_x_3_interleave_0 = const()[name = tensor<string, []>("new_x_3_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_3_cast_fp16 = concat(axis = var_59, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = tensor<string, []>("new_x_3_cast_fp16")];
+            tensor<int32, [3]> var_588_begin_0 = const()[name = tensor<string, []>("op_588_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_588_end_0 = const()[name = tensor<string, []>("op_588_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_588_end_mask_0 = const()[name = tensor<string, []>("op_588_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_588_cast_fp16 = slice_by_index(begin = var_588_begin_0, end = var_588_end_0, end_mask = var_588_end_mask_0, x = new_x_3_cast_fp16)[name = tensor<string, []>("op_588_cast_fp16")];
+            tensor<string, []> x_21_pad_type_0 = const()[name = tensor<string, []>("x_21_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_21_groups_0 = const()[name = tensor<string, []>("x_21_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_21_strides_0 = const()[name = tensor<string, []>("x_21_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_21_pad_0 = const()[name = tensor<string, []>("x_21_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_21_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19451840))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19461120))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_21_cast_fp16")];
+            tensor<int32, [3]> input_55_perm_0 = const()[name = tensor<string, []>("input_55_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_23_axes_0 = const()[name = tensor<string, []>("x_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19463232)))];
+            tensor<fp16, [1024]> encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19465344)))];
+            tensor<fp16, [1, 7, 1024]> input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = tensor<string, []>("transpose_356")];
+            tensor<fp16, [1, 7, 1024]> x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
+            tensor<int32, [3]> input_57_perm_0 = const()[name = tensor<string, []>("input_57_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = tensor<string, []>("transpose_355")];
+            tensor<fp16, [1, 1024, 7]> input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
+            tensor<string, []> x_25_pad_type_0 = const()[name = tensor<string, []>("x_25_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_25_strides_0 = const()[name = tensor<string, []>("x_25_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_25_pad_0 = const()[name = tensor<string, []>("x_25_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_25_dilations_0 = const()[name = tensor<string, []>("x_25_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_25_groups_0 = const()[name = tensor<string, []>("x_25_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19467456))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20516096))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_25_cast_fp16")];
+            tensor<int32, [3]> input_61_perm_0 = const()[name = tensor<string, []>("input_61_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = tensor<string, []>("transpose_354")];
+            tensor<fp16, [1, 7, 1024]> input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
+            tensor<int32, [1]> input_65_axes_0 = const()[name = tensor<string, []>("input_65_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20518208)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20520320)))];
+            tensor<fp16, [1, 7, 1024]> input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20522432))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24716800))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805056)))];
+            tensor<fp16, [4096]> encoder_layers_0_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24725056)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_8_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24733312))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28927680))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_0_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28929792)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_9_cast_fp16")];
+            tensor<fp16, []> var_631_to_fp16 = const()[name = tensor<string, []>("op_631_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_632_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_631_to_fp16)[name = tensor<string, []>("op_632_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_632_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
+            tensor<int32, [1]> input_77_axes_0 = const()[name = tensor<string, []>("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28931904)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28934016)))];
+            tensor<fp16, [1, 7, 1024]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
+            tensor<int32, [4]> cache_5_begin_0 = const()[name = tensor<string, []>("cache_5_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_5_end_0 = const()[name = tensor<string, []>("cache_5_end_0"), val = tensor<int32, [4]>([2, 1, 42, 1024])];
+            tensor<bool, [4]> cache_5_end_mask_0 = const()[name = tensor<string, []>("cache_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_5_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_5_cast_fp16")];
+            tensor<int32, [4]> cache_7_begin_0 = const()[name = tensor<string, []>("cache_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_7_end_0 = const()[name = tensor<string, []>("cache_7_end_0"), val = tensor<int32, [4]>([2, 1, 1024, 8])];
+            tensor<bool, [4]> cache_7_end_mask_0 = const()[name = tensor<string, []>("cache_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_7_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_7_cast_fp16")];
+            tensor<int32, [1]> input_79_axes_0 = const()[name = tensor<string, []>("input_79_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28936128)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28938240)))];
+            tensor<fp16, [1, 7, 1024]> input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28940352))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33134720))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805056)))];
+            tensor<fp16, [4096]> encoder_layers_1_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33142976)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_10_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33151232))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37345600))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_1_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37347712)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_11_cast_fp16")];
+            tensor<fp16, []> var_668_to_fp16 = const()[name = tensor<string, []>("op_668_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_669_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_668_to_fp16)[name = tensor<string, []>("op_669_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_669_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
+            tensor<int32, [1]> key_3_axes_0 = const()[name = tensor<string, []>("key_3_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37349824)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37351936)))];
+            tensor<fp16, [1, 7, 1024]> key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("key_3_cast_fp16")];
+            tensor<bool, []> input_91_interleave_0 = const()[name = tensor<string, []>("input_91_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_91_cast_fp16 = concat(axis = var_68, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = tensor<string, []>("input_91_cast_fp16")];
+            tensor<int32, [3]> var_691_begin_0 = const()[name = tensor<string, []>("op_691_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_691_end_0 = const()[name = tensor<string, []>("op_691_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_691_end_mask_0 = const()[name = tensor<string, []>("op_691_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_691_cast_fp16 = slice_by_index(begin = var_691_begin_0, end = var_691_end_0, end_mask = var_691_end_mask_0, x = cache_5_cast_fp16)[name = tensor<string, []>("op_691_cast_fp16")];
+            tensor<bool, []> var_697_interleave_0 = const()[name = tensor<string, []>("op_697_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_697_cast_fp16 = concat(axis = var_68, interleave = var_697_interleave_0, values = (var_691_cast_fp16, key_3_cast_fp16))[name = tensor<string, []>("op_697_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37354048))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38402688))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38404800)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_12_cast_fp16")];
+            tensor<int32, [4]> var_702 = const()[name = tensor<string, []>("op_702"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_7_cast_fp16 = reshape(shape = var_702, x = linear_12_cast_fp16)[name = tensor<string, []>("q_7_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38406912))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39455552))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39457664)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_13_cast_fp16")];
+            tensor<int32, [4]> var_707 = const()[name = tensor<string, []>("op_707"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_5_cast_fp16 = reshape(shape = var_707, x = linear_13_cast_fp16)[name = tensor<string, []>("k_5_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39459776))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40508416))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40510528)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_14_cast_fp16")];
+            tensor<int32, [4]> var_712 = const()[name = tensor<string, []>("op_712"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_3_cast_fp16 = reshape(shape = var_712, x = linear_14_cast_fp16)[name = tensor<string, []>("v_3_cast_fp16")];
+            tensor<int32, [4]> value_11_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40512640)))];
+            tensor<fp16, [1, 7, 8, 128]> var_725_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_725_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40514752)))];
+            tensor<fp16, [1, 7, 8, 128]> var_727_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_727_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_3_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_33_transpose_x_0 = const()[name = tensor<string, []>("x_33_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_33_transpose_y_0 = const()[name = tensor<string, []>("x_33_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_729_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_729_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40516864))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40616448))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40616256)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_727_cast_fp16)[name = tensor<string, []>("transpose_353")];
+            tensor<fp16, [1, 8, 7, 97]> 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_729_to_fp16_quantized)[name = tensor<string, []>("x_33_cast_fp16")];
+            tensor<int32, [8]> x_35_pad_0 = const()[name = tensor<string, []>("x_35_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_35_mode_0 = const()[name = tensor<string, []>("x_35_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_92_to_fp16 = const()[name = tensor<string, []>("const_92_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_35_cast_fp16")];
+            tensor<int32, [4]> var_737 = const()[name = tensor<string, []>("op_737"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_37_cast_fp16 = reshape(shape = var_737, x = x_35_cast_fp16)[name = tensor<string, []>("x_37_cast_fp16")];
+            tensor<int32, [4]> var_741_begin_0 = const()[name = tensor<string, []>("op_741_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_741_end_0 = const()[name = tensor<string, []>("op_741_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_741_end_mask_0 = const()[name = tensor<string, []>("op_741_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_741_cast_fp16 = slice_by_index(begin = var_741_begin_0, end = var_741_end_0, end_mask = var_741_end_mask_0, x = x_37_cast_fp16)[name = tensor<string, []>("op_741_cast_fp16")];
+            tensor<int32, [4]> var_742 = const()[name = tensor<string, []>("op_742"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_5_cast_fp16 = reshape(shape = var_742, x = var_741_cast_fp16)[name = tensor<string, []>("matrix_bd_5_cast_fp16")];
+            tensor<bool, []> matrix_ac_3_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_3_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_3_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_3_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_98_perm_0 = const()[name = tensor<string, []>("transpose_98_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_99_perm_0 = const()[name = tensor<string, []>("transpose_99_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5_cast_fp16)[name = tensor<string, []>("transpose_351")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_98 = transpose(perm = transpose_98_perm_0, x = var_725_cast_fp16)[name = tensor<string, []>("transpose_352")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_3_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_7_begin_0 = const()[name = tensor<string, []>("matrix_bd_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_7_end_0 = const()[name = tensor<string, []>("matrix_bd_7_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_7_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_7_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_751_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = tensor<string, []>("op_751_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_5_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_5_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_5_cast_fp16 = mul(x = var_751_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_5_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_7_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_7_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_757_cast_fp16 = softmax(axis = var_59, x = scores_7_cast_fp16)[name = tensor<string, []>("op_757_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_93_cast_fp16 = select(a = var_44_to_fp16, b = var_757_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_93_cast_fp16")];
+            tensor<bool, []> x_39_transpose_x_0 = const()[name = tensor<string, []>("x_39_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_39_transpose_y_0 = const()[name = tensor<string, []>("x_39_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_3_cast_fp16)[name = tensor<string, []>("transpose_350")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_39_cast_fp16")];
+            tensor<int32, [4]> var_761_perm_0 = const()[name = tensor<string, []>("op_761_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_762 = const()[name = tensor<string, []>("op_762"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_761_cast_fp16 = transpose(perm = var_761_perm_0, x = x_39_cast_fp16)[name = tensor<string, []>("transpose_349")];
+            tensor<fp16, [1, 7, 1024]> input_95_cast_fp16 = reshape(shape = var_762, x = var_761_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40616768))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41665408))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41667520)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_16_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
+            tensor<int32, [1]> x_43_axes_0 = const()[name = tensor<string, []>("x_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41669632)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41671744)))];
+            tensor<fp16, [1, 7, 1024]> x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")];
+            tensor<int32, [3]> input_101_perm_0 = const()[name = tensor<string, []>("input_101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_103_pad_type_0 = const()[name = tensor<string, []>("input_103_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_103_strides_0 = const()[name = tensor<string, []>("input_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_103_pad_0 = const()[name = tensor<string, []>("input_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_103_dilations_0 = const()[name = tensor<string, []>("input_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_103_groups_0 = const()[name = tensor<string, []>("input_103_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41673856))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43771072))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = tensor<string, []>("transpose_348")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_103_cast_fp16")];
+            tensor<int32, []> x_45_split_num_splits_0 = const()[name = tensor<string, []>("x_45_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_45_split_axis_0 = const()[name = tensor<string, []>("x_45_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_45_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_45_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = tensor<string, []>("x_45_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_45_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_105_cast_fp16 = select(a = var_44_to_fp16, b = x_45_cast_fp16, cond = var_575)[name = tensor<string, []>("input_105_cast_fp16")];
+            tensor<bool, []> new_x_7_interleave_0 = const()[name = tensor<string, []>("new_x_7_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_7_cast_fp16 = concat(axis = var_59, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = tensor<string, []>("new_x_7_cast_fp16")];
+            tensor<int32, [3]> var_801_begin_0 = const()[name = tensor<string, []>("op_801_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_801_end_0 = const()[name = tensor<string, []>("op_801_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_801_end_mask_0 = const()[name = tensor<string, []>("op_801_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_801_cast_fp16 = slice_by_index(begin = var_801_begin_0, end = var_801_end_0, end_mask = var_801_end_mask_0, x = new_x_7_cast_fp16)[name = tensor<string, []>("op_801_cast_fp16")];
+            tensor<string, []> x_47_pad_type_0 = const()[name = tensor<string, []>("x_47_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_47_groups_0 = const()[name = tensor<string, []>("x_47_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_47_strides_0 = const()[name = tensor<string, []>("x_47_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_47_pad_0 = const()[name = tensor<string, []>("x_47_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_47_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43775232))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43784512))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_47_cast_fp16")];
+            tensor<int32, [3]> input_107_perm_0 = const()[name = tensor<string, []>("input_107_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_49_axes_0 = const()[name = tensor<string, []>("x_49_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43786624)))];
+            tensor<fp16, [1024]> encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43788736)))];
+            tensor<fp16, [1, 7, 1024]> input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = tensor<string, []>("transpose_347")];
+            tensor<fp16, [1, 7, 1024]> x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("x_49_cast_fp16")];
+            tensor<int32, [3]> input_109_perm_0 = const()[name = tensor<string, []>("input_109_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = tensor<string, []>("transpose_346")];
+            tensor<fp16, [1, 1024, 7]> input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
+            tensor<string, []> x_51_pad_type_0 = const()[name = tensor<string, []>("x_51_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_51_strides_0 = const()[name = tensor<string, []>("x_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_51_pad_0 = const()[name = tensor<string, []>("x_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_51_dilations_0 = const()[name = tensor<string, []>("x_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_51_groups_0 = const()[name = tensor<string, []>("x_51_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43790848))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44839488))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_51_cast_fp16")];
+            tensor<int32, [3]> input_113_perm_0 = const()[name = tensor<string, []>("input_113_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = tensor<string, []>("transpose_345")];
+            tensor<fp16, [1, 7, 1024]> input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
+            tensor<int32, [1]> input_117_axes_0 = const()[name = tensor<string, []>("input_117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44841600)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44843712)))];
+            tensor<fp16, [1, 7, 1024]> input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44845824))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49040192))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805056)))];
+            tensor<fp16, [4096]> encoder_layers_1_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49048448)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_17_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49056704))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53251072))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_1_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53253184)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_18_cast_fp16")];
+            tensor<fp16, []> var_844_to_fp16 = const()[name = tensor<string, []>("op_844_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_845_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_844_to_fp16)[name = tensor<string, []>("op_845_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_845_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
+            tensor<int32, [1]> input_129_axes_0 = const()[name = tensor<string, []>("input_129_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53255296)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53257408)))];
+            tensor<fp16, [1, 7, 1024]> input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
+            tensor<int32, [4]> cache_9_begin_0 = const()[name = tensor<string, []>("cache_9_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_9_end_0 = const()[name = tensor<string, []>("cache_9_end_0"), val = tensor<int32, [4]>([3, 1, 42, 1024])];
+            tensor<bool, [4]> cache_9_end_mask_0 = const()[name = tensor<string, []>("cache_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_9_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_9_cast_fp16")];
+            tensor<int32, [4]> cache_11_begin_0 = const()[name = tensor<string, []>("cache_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_11_end_0 = const()[name = tensor<string, []>("cache_11_end_0"), val = tensor<int32, [4]>([3, 1, 1024, 8])];
+            tensor<bool, [4]> cache_11_end_mask_0 = const()[name = tensor<string, []>("cache_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_11_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_11_cast_fp16")];
+            tensor<int32, [1]> input_131_axes_0 = const()[name = tensor<string, []>("input_131_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53259520)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53261632)))];
+            tensor<fp16, [1, 7, 1024]> input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53263744))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57458112))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805056)))];
+            tensor<fp16, [4096]> encoder_layers_2_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57466368)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_19_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57474624))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61668992))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_2_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61671104)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_20_cast_fp16")];
+            tensor<fp16, []> var_881_to_fp16 = const()[name = tensor<string, []>("op_881_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_882_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_881_to_fp16)[name = tensor<string, []>("op_882_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_882_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
+            tensor<int32, [1]> key_5_axes_0 = const()[name = tensor<string, []>("key_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61673216)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61675328)))];
+            tensor<fp16, [1, 7, 1024]> key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")];
+            tensor<bool, []> input_143_interleave_0 = const()[name = tensor<string, []>("input_143_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_143_cast_fp16 = concat(axis = var_68, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = tensor<string, []>("input_143_cast_fp16")];
+            tensor<int32, [3]> var_904_begin_0 = const()[name = tensor<string, []>("op_904_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_904_end_0 = const()[name = tensor<string, []>("op_904_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_904_end_mask_0 = const()[name = tensor<string, []>("op_904_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_904_cast_fp16 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = cache_9_cast_fp16)[name = tensor<string, []>("op_904_cast_fp16")];
+            tensor<bool, []> var_910_interleave_0 = const()[name = tensor<string, []>("op_910_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_910_cast_fp16 = concat(axis = var_68, interleave = var_910_interleave_0, values = (var_904_cast_fp16, key_5_cast_fp16))[name = tensor<string, []>("op_910_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61677440))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62726080))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62728192)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_21_cast_fp16")];
+            tensor<int32, [4]> var_915 = const()[name = tensor<string, []>("op_915"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_13_cast_fp16 = reshape(shape = var_915, x = linear_21_cast_fp16)[name = tensor<string, []>("q_13_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62730304))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63778944))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63781056)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_22_cast_fp16")];
+            tensor<int32, [4]> var_920 = const()[name = tensor<string, []>("op_920"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_9_cast_fp16 = reshape(shape = var_920, x = linear_22_cast_fp16)[name = tensor<string, []>("k_9_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63783168))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64831808))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64833920)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_23_cast_fp16")];
+            tensor<int32, [4]> var_925 = const()[name = tensor<string, []>("op_925"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_5_cast_fp16 = reshape(shape = var_925, x = linear_23_cast_fp16)[name = tensor<string, []>("v_5_cast_fp16")];
+            tensor<int32, [4]> value_13_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64836032)))];
+            tensor<fp16, [1, 7, 8, 128]> var_938_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_938_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64838144)))];
+            tensor<fp16, [1, 7, 8, 128]> var_940_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_940_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_5_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_5_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_59_transpose_x_0 = const()[name = tensor<string, []>("x_59_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_59_transpose_y_0 = const()[name = tensor<string, []>("x_59_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_942_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_942_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64840256))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64939840))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64939648)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_940_cast_fp16)[name = tensor<string, []>("transpose_344")];
+            tensor<fp16, [1, 8, 7, 97]> 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_942_to_fp16_quantized)[name = tensor<string, []>("x_59_cast_fp16")];
+            tensor<int32, [8]> x_61_pad_0 = const()[name = tensor<string, []>("x_61_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_61_mode_0 = const()[name = tensor<string, []>("x_61_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_105_to_fp16 = const()[name = tensor<string, []>("const_105_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_61_cast_fp16")];
+            tensor<int32, [4]> var_950 = const()[name = tensor<string, []>("op_950"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_63_cast_fp16 = reshape(shape = var_950, x = x_61_cast_fp16)[name = tensor<string, []>("x_63_cast_fp16")];
+            tensor<int32, [4]> var_954_begin_0 = const()[name = tensor<string, []>("op_954_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_954_end_0 = const()[name = tensor<string, []>("op_954_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_954_end_mask_0 = const()[name = tensor<string, []>("op_954_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_954_cast_fp16 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, x = x_63_cast_fp16)[name = tensor<string, []>("op_954_cast_fp16")];
+            tensor<int32, [4]> var_955 = const()[name = tensor<string, []>("op_955"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_9_cast_fp16 = reshape(shape = var_955, x = var_954_cast_fp16)[name = tensor<string, []>("matrix_bd_9_cast_fp16")];
+            tensor<bool, []> matrix_ac_5_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_5_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_5_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_5_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_100_perm_0 = const()[name = tensor<string, []>("transpose_100_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_101_perm_0 = const()[name = tensor<string, []>("transpose_101_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9_cast_fp16)[name = tensor<string, []>("transpose_342")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_100 = transpose(perm = transpose_100_perm_0, x = var_938_cast_fp16)[name = tensor<string, []>("transpose_343")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_5_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_11_begin_0 = const()[name = tensor<string, []>("matrix_bd_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_11_end_0 = const()[name = tensor<string, []>("matrix_bd_11_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_11_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_11_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_964_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = tensor<string, []>("op_964_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_9_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_9_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_9_cast_fp16 = mul(x = var_964_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_9_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_11_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_11_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_970_cast_fp16 = softmax(axis = var_59, x = scores_11_cast_fp16)[name = tensor<string, []>("op_970_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_145_cast_fp16 = select(a = var_44_to_fp16, b = var_970_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_145_cast_fp16")];
+            tensor<bool, []> x_65_transpose_x_0 = const()[name = tensor<string, []>("x_65_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_65_transpose_y_0 = const()[name = tensor<string, []>("x_65_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_5_cast_fp16)[name = tensor<string, []>("transpose_341")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_65_cast_fp16")];
+            tensor<int32, [4]> var_974_perm_0 = const()[name = tensor<string, []>("op_974_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_975 = const()[name = tensor<string, []>("op_975"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_974_cast_fp16 = transpose(perm = var_974_perm_0, x = x_65_cast_fp16)[name = tensor<string, []>("transpose_340")];
+            tensor<fp16, [1, 7, 1024]> input_147_cast_fp16 = reshape(shape = var_975, x = var_974_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64940160))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65726656))), name = tensor<string, []>("encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65726848)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_25_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")];
+            tensor<int32, [1]> x_69_axes_0 = const()[name = tensor<string, []>("x_69_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65728960)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65731072)))];
+            tensor<fp16, [1, 7, 1024]> x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = tensor<string, []>("x_69_cast_fp16")];
+            tensor<int32, [3]> input_153_perm_0 = const()[name = tensor<string, []>("input_153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_155_pad_type_0 = const()[name = tensor<string, []>("input_155_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_155_strides_0 = const()[name = tensor<string, []>("input_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_155_pad_0 = const()[name = tensor<string, []>("input_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_155_dilations_0 = const()[name = tensor<string, []>("input_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_155_groups_0 = const()[name = tensor<string, []>("input_155_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65733184))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67830400))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = tensor<string, []>("transpose_339")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_155_cast_fp16")];
+            tensor<int32, []> x_71_split_num_splits_0 = const()[name = tensor<string, []>("x_71_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_71_split_axis_0 = const()[name = tensor<string, []>("x_71_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_71_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_71_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = tensor<string, []>("x_71_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_71_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_157_cast_fp16 = select(a = var_44_to_fp16, b = x_71_cast_fp16, cond = var_575)[name = tensor<string, []>("input_157_cast_fp16")];
+            tensor<bool, []> new_x_11_interleave_0 = const()[name = tensor<string, []>("new_x_11_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_11_cast_fp16 = concat(axis = var_59, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = tensor<string, []>("new_x_11_cast_fp16")];
+            tensor<int32, [3]> var_1014_begin_0 = const()[name = tensor<string, []>("op_1014_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_1014_end_0 = const()[name = tensor<string, []>("op_1014_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_1014_end_mask_0 = const()[name = tensor<string, []>("op_1014_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1014_cast_fp16 = slice_by_index(begin = var_1014_begin_0, end = var_1014_end_0, end_mask = var_1014_end_mask_0, x = new_x_11_cast_fp16)[name = tensor<string, []>("op_1014_cast_fp16")];
+            tensor<string, []> x_73_pad_type_0 = const()[name = tensor<string, []>("x_73_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_73_groups_0 = const()[name = tensor<string, []>("x_73_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_73_strides_0 = const()[name = tensor<string, []>("x_73_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_73_pad_0 = const()[name = tensor<string, []>("x_73_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_73_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67834560))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67843840))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_73_cast_fp16")];
+            tensor<int32, [3]> input_159_perm_0 = const()[name = tensor<string, []>("input_159_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_75_axes_0 = const()[name = tensor<string, []>("x_75_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67845952)))];
+            tensor<fp16, [1024]> encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67848064)))];
+            tensor<fp16, [1, 7, 1024]> input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = tensor<string, []>("transpose_338")];
+            tensor<fp16, [1, 7, 1024]> x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = tensor<string, []>("x_75_cast_fp16")];
+            tensor<int32, [3]> input_161_perm_0 = const()[name = tensor<string, []>("input_161_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = tensor<string, []>("transpose_337")];
+            tensor<fp16, [1, 1024, 7]> input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
+            tensor<string, []> x_77_pad_type_0 = const()[name = tensor<string, []>("x_77_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_77_strides_0 = const()[name = tensor<string, []>("x_77_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_77_pad_0 = const()[name = tensor<string, []>("x_77_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_77_dilations_0 = const()[name = tensor<string, []>("x_77_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_77_groups_0 = const()[name = tensor<string, []>("x_77_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67850176))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68898816))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_77_cast_fp16")];
+            tensor<int32, [3]> input_165_perm_0 = const()[name = tensor<string, []>("input_165_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = tensor<string, []>("transpose_336")];
+            tensor<fp16, [1, 7, 1024]> input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
+            tensor<int32, [1]> input_169_axes_0 = const()[name = tensor<string, []>("input_169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68900928)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68903040)))];
+            tensor<fp16, [1, 7, 1024]> input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68905152))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72050944))), name = tensor<string, []>("encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_2_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72051136)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_26_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72059392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75205184))), name = tensor<string, []>("encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_2_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75205376)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_27_cast_fp16")];
+            tensor<fp16, []> var_1057_to_fp16 = const()[name = tensor<string, []>("op_1057_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1058_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1057_to_fp16)[name = tensor<string, []>("op_1058_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_1058_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
+            tensor<int32, [1]> input_181_axes_0 = const()[name = tensor<string, []>("input_181_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75207488)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75209600)))];
+            tensor<fp16, [1, 7, 1024]> input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
+            tensor<int32, [4]> cache_13_begin_0 = const()[name = tensor<string, []>("cache_13_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_13_end_0 = const()[name = tensor<string, []>("cache_13_end_0"), val = tensor<int32, [4]>([4, 1, 42, 1024])];
+            tensor<bool, [4]> cache_13_end_mask_0 = const()[name = tensor<string, []>("cache_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_13_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_13_cast_fp16")];
+            tensor<int32, [4]> cache_15_begin_0 = const()[name = tensor<string, []>("cache_15_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_15_end_0 = const()[name = tensor<string, []>("cache_15_end_0"), val = tensor<int32, [4]>([4, 1, 1024, 8])];
+            tensor<bool, [4]> cache_15_end_mask_0 = const()[name = tensor<string, []>("cache_15_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_15_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_15_cast_fp16")];
+            tensor<int32, [1]> input_183_axes_0 = const()[name = tensor<string, []>("input_183_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75211712)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75213824)))];
+            tensor<fp16, [1, 7, 1024]> input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75215936))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78361728))), name = tensor<string, []>("encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_3_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78361920)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_28_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78370176))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81515968))), name = tensor<string, []>("encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_3_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81516160)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_29_cast_fp16")];
+            tensor<fp16, []> var_1094_to_fp16 = const()[name = tensor<string, []>("op_1094_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1095_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1094_to_fp16)[name = tensor<string, []>("op_1095_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_1095_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
+            tensor<int32, [1]> key_7_axes_0 = const()[name = tensor<string, []>("key_7_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81518272)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81520384)))];
+            tensor<fp16, [1, 7, 1024]> key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = tensor<string, []>("key_7_cast_fp16")];
+            tensor<bool, []> input_195_interleave_0 = const()[name = tensor<string, []>("input_195_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_195_cast_fp16 = concat(axis = var_68, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = tensor<string, []>("input_195_cast_fp16")];
+            tensor<int32, [3]> var_1117_begin_0 = const()[name = tensor<string, []>("op_1117_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_1117_end_0 = const()[name = tensor<string, []>("op_1117_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1117_end_mask_0 = const()[name = tensor<string, []>("op_1117_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_1117_cast_fp16 = slice_by_index(begin = var_1117_begin_0, end = var_1117_end_0, end_mask = var_1117_end_mask_0, x = cache_13_cast_fp16)[name = tensor<string, []>("op_1117_cast_fp16")];
+            tensor<bool, []> var_1123_interleave_0 = const()[name = tensor<string, []>("op_1123_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_1123_cast_fp16 = concat(axis = var_68, interleave = var_1123_interleave_0, values = (var_1117_cast_fp16, key_7_cast_fp16))[name = tensor<string, []>("op_1123_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81522496))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82308992))), name = tensor<string, []>("encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82309184)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_30_cast_fp16")];
+            tensor<int32, [4]> var_1128 = const()[name = tensor<string, []>("op_1128"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_19_cast_fp16 = reshape(shape = var_1128, x = linear_30_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82311296))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83097792))), name = tensor<string, []>("encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83097984)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_31_cast_fp16")];
+            tensor<int32, [4]> var_1133 = const()[name = tensor<string, []>("op_1133"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_13_cast_fp16 = reshape(shape = var_1133, x = linear_31_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83100096))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83886592))), name = tensor<string, []>("encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83886784)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_32_cast_fp16")];
+            tensor<int32, [4]> var_1138 = const()[name = tensor<string, []>("op_1138"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_7_cast_fp16 = reshape(shape = var_1138, x = linear_32_cast_fp16)[name = tensor<string, []>("v_7_cast_fp16")];
+            tensor<int32, [4]> value_15_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83888896)))];
+            tensor<fp16, [1, 7, 8, 128]> var_1151_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1151_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83891008)))];
+            tensor<fp16, [1, 7, 8, 128]> var_1153_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1153_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_7_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_85_transpose_x_0 = const()[name = tensor<string, []>("x_85_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_85_transpose_y_0 = const()[name = tensor<string, []>("x_85_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_1155_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1155_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83893120))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83992704))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83992512)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1153_cast_fp16)[name = tensor<string, []>("transpose_335")];
+            tensor<fp16, [1, 8, 7, 97]> 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_1155_to_fp16_quantized)[name = tensor<string, []>("x_85_cast_fp16")];
+            tensor<int32, [8]> x_87_pad_0 = const()[name = tensor<string, []>("x_87_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_87_mode_0 = const()[name = tensor<string, []>("x_87_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_118_to_fp16 = const()[name = tensor<string, []>("const_118_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_87_cast_fp16")];
+            tensor<int32, [4]> var_1163 = const()[name = tensor<string, []>("op_1163"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_89_cast_fp16 = reshape(shape = var_1163, x = x_87_cast_fp16)[name = tensor<string, []>("x_89_cast_fp16")];
+            tensor<int32, [4]> var_1167_begin_0 = const()[name = tensor<string, []>("op_1167_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1167_end_0 = const()[name = tensor<string, []>("op_1167_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_1167_end_mask_0 = const()[name = tensor<string, []>("op_1167_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_1167_cast_fp16 = slice_by_index(begin = var_1167_begin_0, end = var_1167_end_0, end_mask = var_1167_end_mask_0, x = x_89_cast_fp16)[name = tensor<string, []>("op_1167_cast_fp16")];
+            tensor<int32, [4]> var_1168 = const()[name = tensor<string, []>("op_1168"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_13_cast_fp16 = reshape(shape = var_1168, x = var_1167_cast_fp16)[name = tensor<string, []>("matrix_bd_13_cast_fp16")];
+            tensor<bool, []> matrix_ac_7_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_7_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_102_perm_0 = const()[name = tensor<string, []>("transpose_102_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_103_perm_0 = const()[name = tensor<string, []>("transpose_103_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13_cast_fp16)[name = tensor<string, []>("transpose_333")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1151_cast_fp16)[name = tensor<string, []>("transpose_334")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_7_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_15_begin_0 = const()[name = tensor<string, []>("matrix_bd_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_15_end_0 = const()[name = tensor<string, []>("matrix_bd_15_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_15_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_15_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_1177_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = tensor<string, []>("op_1177_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_13_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_13_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_13_cast_fp16 = mul(x = var_1177_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_13_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_15_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_15_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_1183_cast_fp16 = softmax(axis = var_59, x = scores_15_cast_fp16)[name = tensor<string, []>("op_1183_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_197_cast_fp16 = select(a = var_44_to_fp16, b = var_1183_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_197_cast_fp16")];
+            tensor<bool, []> x_91_transpose_x_0 = const()[name = tensor<string, []>("x_91_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_91_transpose_y_0 = const()[name = tensor<string, []>("x_91_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_7_cast_fp16)[name = tensor<string, []>("transpose_332")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_91_cast_fp16")];
+            tensor<int32, [4]> var_1187_perm_0 = const()[name = tensor<string, []>("op_1187_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1188 = const()[name = tensor<string, []>("op_1188"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_1187_cast_fp16 = transpose(perm = var_1187_perm_0, x = x_91_cast_fp16)[name = tensor<string, []>("transpose_331")];
+            tensor<fp16, [1, 7, 1024]> input_199_cast_fp16 = reshape(shape = var_1188, x = var_1187_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83993024))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84779520))), name = tensor<string, []>("encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84779712)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_34_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
+            tensor<int32, [1]> x_95_axes_0 = const()[name = tensor<string, []>("x_95_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84781824)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84783936)))];
+            tensor<fp16, [1, 7, 1024]> x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = tensor<string, []>("x_95_cast_fp16")];
+            tensor<int32, [3]> input_205_perm_0 = const()[name = tensor<string, []>("input_205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_207_pad_type_0 = const()[name = tensor<string, []>("input_207_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_207_strides_0 = const()[name = tensor<string, []>("input_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_207_pad_0 = const()[name = tensor<string, []>("input_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_207_dilations_0 = const()[name = tensor<string, []>("input_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_207_groups_0 = const()[name = tensor<string, []>("input_207_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84786048))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86883264))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = tensor<string, []>("transpose_330")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_207_cast_fp16")];
+            tensor<int32, []> x_97_split_num_splits_0 = const()[name = tensor<string, []>("x_97_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_97_split_axis_0 = const()[name = tensor<string, []>("x_97_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_97_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_97_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = tensor<string, []>("x_97_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_97_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_209_cast_fp16 = select(a = var_44_to_fp16, b = x_97_cast_fp16, cond = var_575)[name = tensor<string, []>("input_209_cast_fp16")];
+            tensor<bool, []> new_x_15_interleave_0 = const()[name = tensor<string, []>("new_x_15_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_15_cast_fp16 = concat(axis = var_59, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = tensor<string, []>("new_x_15_cast_fp16")];
+            tensor<int32, [3]> var_1227_begin_0 = const()[name = tensor<string, []>("op_1227_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_1227_end_0 = const()[name = tensor<string, []>("op_1227_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_1227_end_mask_0 = const()[name = tensor<string, []>("op_1227_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1227_cast_fp16 = slice_by_index(begin = var_1227_begin_0, end = var_1227_end_0, end_mask = var_1227_end_mask_0, x = new_x_15_cast_fp16)[name = tensor<string, []>("op_1227_cast_fp16")];
+            tensor<string, []> x_99_pad_type_0 = const()[name = tensor<string, []>("x_99_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_99_groups_0 = const()[name = tensor<string, []>("x_99_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_99_strides_0 = const()[name = tensor<string, []>("x_99_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_99_pad_0 = const()[name = tensor<string, []>("x_99_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_99_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86887424))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86896704))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_99_cast_fp16")];
+            tensor<int32, [3]> input_211_perm_0 = const()[name = tensor<string, []>("input_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_101_axes_0 = const()[name = tensor<string, []>("x_101_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86898816)))];
+            tensor<fp16, [1024]> encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86900928)))];
+            tensor<fp16, [1, 7, 1024]> input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = tensor<string, []>("transpose_329")];
+            tensor<fp16, [1, 7, 1024]> x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = tensor<string, []>("x_101_cast_fp16")];
+            tensor<int32, [3]> input_213_perm_0 = const()[name = tensor<string, []>("input_213_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = tensor<string, []>("transpose_328")];
+            tensor<fp16, [1, 1024, 7]> input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")];
+            tensor<string, []> x_103_pad_type_0 = const()[name = tensor<string, []>("x_103_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_103_strides_0 = const()[name = tensor<string, []>("x_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_103_pad_0 = const()[name = tensor<string, []>("x_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_103_dilations_0 = const()[name = tensor<string, []>("x_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_103_groups_0 = const()[name = tensor<string, []>("x_103_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86903040))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87951680))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_103_cast_fp16")];
+            tensor<int32, [3]> input_217_perm_0 = const()[name = tensor<string, []>("input_217_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = tensor<string, []>("transpose_327")];
+            tensor<fp16, [1, 7, 1024]> input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
+            tensor<int32, [1]> input_221_axes_0 = const()[name = tensor<string, []>("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87953792)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87955904)))];
+            tensor<fp16, [1, 7, 1024]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87958016))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91103808))), name = tensor<string, []>("encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_3_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91104000)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_35_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91112256))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94258048))), name = tensor<string, []>("encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_3_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94258240)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_36_cast_fp16")];
+            tensor<fp16, []> var_1270_to_fp16 = const()[name = tensor<string, []>("op_1270_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1271_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1270_to_fp16)[name = tensor<string, []>("op_1271_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1271_cast_fp16)[name = tensor<string, []>("input_231_cast_fp16")];
+            tensor<int32, [1]> input_233_axes_0 = const()[name = tensor<string, []>("input_233_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94260352)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94262464)))];
+            tensor<fp16, [1, 7, 1024]> input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")];
+            tensor<int32, [4]> cache_17_begin_0 = const()[name = tensor<string, []>("cache_17_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_17_end_0 = const()[name = tensor<string, []>("cache_17_end_0"), val = tensor<int32, [4]>([5, 1, 42, 1024])];
+            tensor<bool, [4]> cache_17_end_mask_0 = const()[name = tensor<string, []>("cache_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_17_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_17_cast_fp16")];
+            tensor<int32, [4]> cache_19_begin_0 = const()[name = tensor<string, []>("cache_19_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_19_end_0 = const()[name = tensor<string, []>("cache_19_end_0"), val = tensor<int32, [4]>([5, 1, 1024, 8])];
+            tensor<bool, [4]> cache_19_end_mask_0 = const()[name = tensor<string, []>("cache_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_19_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_19_cast_fp16")];
+            tensor<int32, [1]> input_235_axes_0 = const()[name = tensor<string, []>("input_235_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94264576)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94266688)))];
+            tensor<fp16, [1, 7, 1024]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94268800))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97414592))), name = tensor<string, []>("encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_4_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97414784)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_37_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97423040))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100568832))), name = tensor<string, []>("encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_4_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100569024)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_38_cast_fp16")];
+            tensor<fp16, []> var_1307_to_fp16 = const()[name = tensor<string, []>("op_1307_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1308_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1307_to_fp16)[name = tensor<string, []>("op_1308_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1308_cast_fp16)[name = tensor<string, []>("input_245_cast_fp16")];
+            tensor<int32, [1]> key_9_axes_0 = const()[name = tensor<string, []>("key_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100571136)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100573248)))];
+            tensor<fp16, [1, 7, 1024]> key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")];
+            tensor<bool, []> input_247_interleave_0 = const()[name = tensor<string, []>("input_247_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_247_cast_fp16 = concat(axis = var_68, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = tensor<string, []>("input_247_cast_fp16")];
+            tensor<int32, [3]> var_1330_begin_0 = const()[name = tensor<string, []>("op_1330_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_1330_end_0 = const()[name = tensor<string, []>("op_1330_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1330_end_mask_0 = const()[name = tensor<string, []>("op_1330_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_1330_cast_fp16 = slice_by_index(begin = var_1330_begin_0, end = var_1330_end_0, end_mask = var_1330_end_mask_0, x = cache_17_cast_fp16)[name = tensor<string, []>("op_1330_cast_fp16")];
+            tensor<bool, []> var_1336_interleave_0 = const()[name = tensor<string, []>("op_1336_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_1336_cast_fp16 = concat(axis = var_68, interleave = var_1336_interleave_0, values = (var_1330_cast_fp16, key_9_cast_fp16))[name = tensor<string, []>("op_1336_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100575360))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101361856))), name = tensor<string, []>("encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101362048)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_39_cast_fp16")];
+            tensor<int32, [4]> var_1341 = const()[name = tensor<string, []>("op_1341"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_25_cast_fp16 = reshape(shape = var_1341, x = linear_39_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101364160))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102150656))), name = tensor<string, []>("encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102150848)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_40_cast_fp16")];
+            tensor<int32, [4]> var_1346 = const()[name = tensor<string, []>("op_1346"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_17_cast_fp16 = reshape(shape = var_1346, x = linear_40_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102152960))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102939456))), name = tensor<string, []>("encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102939648)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_41_cast_fp16")];
+            tensor<int32, [4]> var_1351 = const()[name = tensor<string, []>("op_1351"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_9_cast_fp16 = reshape(shape = var_1351, x = linear_41_cast_fp16)[name = tensor<string, []>("v_9_cast_fp16")];
+            tensor<int32, [4]> value_17_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102941760)))];
+            tensor<fp16, [1, 7, 8, 128]> var_1364_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1364_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102943872)))];
+            tensor<fp16, [1, 7, 8, 128]> var_1366_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1366_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_9_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_111_transpose_x_0 = const()[name = tensor<string, []>("x_111_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_111_transpose_y_0 = const()[name = tensor<string, []>("x_111_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_1368_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1368_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102945984))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103045568))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103045376)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1366_cast_fp16)[name = tensor<string, []>("transpose_326")];
+            tensor<fp16, [1, 8, 7, 97]> 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_1368_to_fp16_quantized)[name = tensor<string, []>("x_111_cast_fp16")];
+            tensor<int32, [8]> x_113_pad_0 = const()[name = tensor<string, []>("x_113_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_113_mode_0 = const()[name = tensor<string, []>("x_113_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_131_to_fp16 = const()[name = tensor<string, []>("const_131_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_113_cast_fp16")];
+            tensor<int32, [4]> var_1376 = const()[name = tensor<string, []>("op_1376"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_115_cast_fp16 = reshape(shape = var_1376, x = x_113_cast_fp16)[name = tensor<string, []>("x_115_cast_fp16")];
+            tensor<int32, [4]> var_1380_begin_0 = const()[name = tensor<string, []>("op_1380_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1380_end_0 = const()[name = tensor<string, []>("op_1380_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_1380_end_mask_0 = const()[name = tensor<string, []>("op_1380_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_1380_cast_fp16 = slice_by_index(begin = var_1380_begin_0, end = var_1380_end_0, end_mask = var_1380_end_mask_0, x = x_115_cast_fp16)[name = tensor<string, []>("op_1380_cast_fp16")];
+            tensor<int32, [4]> var_1381 = const()[name = tensor<string, []>("op_1381"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_17_cast_fp16 = reshape(shape = var_1381, x = var_1380_cast_fp16)[name = tensor<string, []>("matrix_bd_17_cast_fp16")];
+            tensor<bool, []> matrix_ac_9_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_9_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_9_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_9_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_104_perm_0 = const()[name = tensor<string, []>("transpose_104_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_105_perm_0 = const()[name = tensor<string, []>("transpose_105_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17_cast_fp16)[name = tensor<string, []>("transpose_324")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1364_cast_fp16)[name = tensor<string, []>("transpose_325")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_9_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_19_begin_0 = const()[name = tensor<string, []>("matrix_bd_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_19_end_0 = const()[name = tensor<string, []>("matrix_bd_19_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_19_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_19_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_1390_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = tensor<string, []>("op_1390_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_17_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_17_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_17_cast_fp16 = mul(x = var_1390_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_17_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_19_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_19_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_1396_cast_fp16 = softmax(axis = var_59, x = scores_19_cast_fp16)[name = tensor<string, []>("op_1396_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_249_cast_fp16 = select(a = var_44_to_fp16, b = var_1396_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_249_cast_fp16")];
+            tensor<bool, []> x_117_transpose_x_0 = const()[name = tensor<string, []>("x_117_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_117_transpose_y_0 = const()[name = tensor<string, []>("x_117_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_9_cast_fp16)[name = tensor<string, []>("transpose_323")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_117_cast_fp16")];
+            tensor<int32, [4]> var_1400_perm_0 = const()[name = tensor<string, []>("op_1400_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1401 = const()[name = tensor<string, []>("op_1401"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_1400_cast_fp16 = transpose(perm = var_1400_perm_0, x = x_117_cast_fp16)[name = tensor<string, []>("transpose_322")];
+            tensor<fp16, [1, 7, 1024]> input_251_cast_fp16 = reshape(shape = var_1401, x = var_1400_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103045888))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103832384))), name = tensor<string, []>("encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103832576)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_43_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = tensor<string, []>("input_255_cast_fp16")];
+            tensor<int32, [1]> x_121_axes_0 = const()[name = tensor<string, []>("x_121_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103834688)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103836800)))];
+            tensor<fp16, [1, 7, 1024]> x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = tensor<string, []>("x_121_cast_fp16")];
+            tensor<int32, [3]> input_257_perm_0 = const()[name = tensor<string, []>("input_257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_259_pad_type_0 = const()[name = tensor<string, []>("input_259_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_259_strides_0 = const()[name = tensor<string, []>("input_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_259_pad_0 = const()[name = tensor<string, []>("input_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_259_dilations_0 = const()[name = tensor<string, []>("input_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_259_groups_0 = const()[name = tensor<string, []>("input_259_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103838912))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105936128))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = tensor<string, []>("transpose_321")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_259_cast_fp16")];
+            tensor<int32, []> x_123_split_num_splits_0 = const()[name = tensor<string, []>("x_123_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_123_split_axis_0 = const()[name = tensor<string, []>("x_123_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_123_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_123_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = tensor<string, []>("x_123_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_123_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_261_cast_fp16 = select(a = var_44_to_fp16, b = x_123_cast_fp16, cond = var_575)[name = tensor<string, []>("input_261_cast_fp16")];
+            tensor<bool, []> new_x_19_interleave_0 = const()[name = tensor<string, []>("new_x_19_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_19_cast_fp16 = concat(axis = var_59, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = tensor<string, []>("new_x_19_cast_fp16")];
+            tensor<int32, [3]> var_1440_begin_0 = const()[name = tensor<string, []>("op_1440_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_1440_end_0 = const()[name = tensor<string, []>("op_1440_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_1440_end_mask_0 = const()[name = tensor<string, []>("op_1440_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1440_cast_fp16 = slice_by_index(begin = var_1440_begin_0, end = var_1440_end_0, end_mask = var_1440_end_mask_0, x = new_x_19_cast_fp16)[name = tensor<string, []>("op_1440_cast_fp16")];
+            tensor<string, []> x_125_pad_type_0 = const()[name = tensor<string, []>("x_125_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_125_groups_0 = const()[name = tensor<string, []>("x_125_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_125_strides_0 = const()[name = tensor<string, []>("x_125_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_125_pad_0 = const()[name = tensor<string, []>("x_125_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_125_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105940288))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105949568))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_125_cast_fp16")];
+            tensor<int32, [3]> input_263_perm_0 = const()[name = tensor<string, []>("input_263_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_127_axes_0 = const()[name = tensor<string, []>("x_127_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105951680)))];
+            tensor<fp16, [1024]> encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105953792)))];
+            tensor<fp16, [1, 7, 1024]> input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = tensor<string, []>("transpose_320")];
+            tensor<fp16, [1, 7, 1024]> x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = tensor<string, []>("x_127_cast_fp16")];
+            tensor<int32, [3]> input_265_perm_0 = const()[name = tensor<string, []>("input_265_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = tensor<string, []>("transpose_319")];
+            tensor<fp16, [1, 1024, 7]> input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = tensor<string, []>("input_267_cast_fp16")];
+            tensor<string, []> x_129_pad_type_0 = const()[name = tensor<string, []>("x_129_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_129_strides_0 = const()[name = tensor<string, []>("x_129_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_129_pad_0 = const()[name = tensor<string, []>("x_129_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_129_dilations_0 = const()[name = tensor<string, []>("x_129_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_129_groups_0 = const()[name = tensor<string, []>("x_129_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105955904))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107004544))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_129_cast_fp16")];
+            tensor<int32, [3]> input_269_perm_0 = const()[name = tensor<string, []>("input_269_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = tensor<string, []>("transpose_318")];
+            tensor<fp16, [1, 7, 1024]> input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = tensor<string, []>("input_271_cast_fp16")];
+            tensor<int32, [1]> input_273_axes_0 = const()[name = tensor<string, []>("input_273_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107006656)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107008768)))];
+            tensor<fp16, [1, 7, 1024]> input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107010880))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110156672))), name = tensor<string, []>("encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_4_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110156864)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_44_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110165120))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113310912))), name = tensor<string, []>("encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_4_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113311104)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_45_cast_fp16")];
+            tensor<fp16, []> var_1483_to_fp16 = const()[name = tensor<string, []>("op_1483_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1484_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1483_to_fp16)[name = tensor<string, []>("op_1484_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1484_cast_fp16)[name = tensor<string, []>("input_283_cast_fp16")];
+            tensor<int32, [1]> input_285_axes_0 = const()[name = tensor<string, []>("input_285_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113313216)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113315328)))];
+            tensor<fp16, [1, 7, 1024]> input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = tensor<string, []>("input_285_cast_fp16")];
+            tensor<int32, [4]> cache_21_begin_0 = const()[name = tensor<string, []>("cache_21_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_21_end_0 = const()[name = tensor<string, []>("cache_21_end_0"), val = tensor<int32, [4]>([6, 1, 42, 1024])];
+            tensor<bool, [4]> cache_21_end_mask_0 = const()[name = tensor<string, []>("cache_21_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_21_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_21_cast_fp16")];
+            tensor<int32, [4]> cache_23_begin_0 = const()[name = tensor<string, []>("cache_23_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_23_end_0 = const()[name = tensor<string, []>("cache_23_end_0"), val = tensor<int32, [4]>([6, 1, 1024, 8])];
+            tensor<bool, [4]> cache_23_end_mask_0 = const()[name = tensor<string, []>("cache_23_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_23_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_23_cast_fp16")];
+            tensor<int32, [1]> input_287_axes_0 = const()[name = tensor<string, []>("input_287_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113317440)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113319552)))];
+            tensor<fp16, [1, 7, 1024]> input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113321664))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116467456))), name = tensor<string, []>("encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_5_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116467648)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_46_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116475904))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119621696))), name = tensor<string, []>("encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_5_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119621888)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_47_cast_fp16")];
+            tensor<fp16, []> var_1520_to_fp16 = const()[name = tensor<string, []>("op_1520_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1521_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1520_to_fp16)[name = tensor<string, []>("op_1521_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1521_cast_fp16)[name = tensor<string, []>("input_297_cast_fp16")];
+            tensor<int32, [1]> key_11_axes_0 = const()[name = tensor<string, []>("key_11_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119624000)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119626112)))];
+            tensor<fp16, [1, 7, 1024]> key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = tensor<string, []>("key_11_cast_fp16")];
+            tensor<bool, []> input_299_interleave_0 = const()[name = tensor<string, []>("input_299_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_299_cast_fp16 = concat(axis = var_68, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = tensor<string, []>("input_299_cast_fp16")];
+            tensor<int32, [3]> var_1543_begin_0 = const()[name = tensor<string, []>("op_1543_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_1543_end_0 = const()[name = tensor<string, []>("op_1543_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1543_end_mask_0 = const()[name = tensor<string, []>("op_1543_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_1543_cast_fp16 = slice_by_index(begin = var_1543_begin_0, end = var_1543_end_0, end_mask = var_1543_end_mask_0, x = cache_21_cast_fp16)[name = tensor<string, []>("op_1543_cast_fp16")];
+            tensor<bool, []> var_1549_interleave_0 = const()[name = tensor<string, []>("op_1549_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_1549_cast_fp16 = concat(axis = var_68, interleave = var_1549_interleave_0, values = (var_1543_cast_fp16, key_11_cast_fp16))[name = tensor<string, []>("op_1549_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119628224))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120414720))), name = tensor<string, []>("encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120414912)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_48_cast_fp16")];
+            tensor<int32, [4]> var_1554 = const()[name = tensor<string, []>("op_1554"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_31_cast_fp16 = reshape(shape = var_1554, x = linear_48_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120417024))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121203520))), name = tensor<string, []>("encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121203712)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_49_cast_fp16")];
+            tensor<int32, [4]> var_1559 = const()[name = tensor<string, []>("op_1559"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_21_cast_fp16 = reshape(shape = var_1559, x = linear_49_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121205824))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121992320))), name = tensor<string, []>("encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121992512)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_50_cast_fp16")];
+            tensor<int32, [4]> var_1564 = const()[name = tensor<string, []>("op_1564"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_11_cast_fp16 = reshape(shape = var_1564, x = linear_50_cast_fp16)[name = tensor<string, []>("v_11_cast_fp16")];
+            tensor<int32, [4]> value_19_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121994624)))];
+            tensor<fp16, [1, 7, 8, 128]> var_1577_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1577_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121996736)))];
+            tensor<fp16, [1, 7, 8, 128]> var_1579_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1579_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_11_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_137_transpose_x_0 = const()[name = tensor<string, []>("x_137_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_137_transpose_y_0 = const()[name = tensor<string, []>("x_137_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_1581_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1581_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121998848))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122098432))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122098240)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1579_cast_fp16)[name = tensor<string, []>("transpose_317")];
+            tensor<fp16, [1, 8, 7, 97]> 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_1581_to_fp16_quantized)[name = tensor<string, []>("x_137_cast_fp16")];
+            tensor<int32, [8]> x_139_pad_0 = const()[name = tensor<string, []>("x_139_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_139_mode_0 = const()[name = tensor<string, []>("x_139_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_144_to_fp16 = const()[name = tensor<string, []>("const_144_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_139_cast_fp16")];
+            tensor<int32, [4]> var_1589 = const()[name = tensor<string, []>("op_1589"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_141_cast_fp16 = reshape(shape = var_1589, x = x_139_cast_fp16)[name = tensor<string, []>("x_141_cast_fp16")];
+            tensor<int32, [4]> var_1593_begin_0 = const()[name = tensor<string, []>("op_1593_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1593_end_0 = const()[name = tensor<string, []>("op_1593_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_1593_end_mask_0 = const()[name = tensor<string, []>("op_1593_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_1593_cast_fp16 = slice_by_index(begin = var_1593_begin_0, end = var_1593_end_0, end_mask = var_1593_end_mask_0, x = x_141_cast_fp16)[name = tensor<string, []>("op_1593_cast_fp16")];
+            tensor<int32, [4]> var_1594 = const()[name = tensor<string, []>("op_1594"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_21_cast_fp16 = reshape(shape = var_1594, x = var_1593_cast_fp16)[name = tensor<string, []>("matrix_bd_21_cast_fp16")];
+            tensor<bool, []> matrix_ac_11_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_11_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_11_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_11_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_106_perm_0 = const()[name = tensor<string, []>("transpose_106_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_107_perm_0 = const()[name = tensor<string, []>("transpose_107_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21_cast_fp16)[name = tensor<string, []>("transpose_315")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1577_cast_fp16)[name = tensor<string, []>("transpose_316")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_11_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_23_begin_0 = const()[name = tensor<string, []>("matrix_bd_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_23_end_0 = const()[name = tensor<string, []>("matrix_bd_23_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_23_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_23_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_1603_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = tensor<string, []>("op_1603_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_21_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_21_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_21_cast_fp16 = mul(x = var_1603_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_21_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_23_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_23_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_1609_cast_fp16 = softmax(axis = var_59, x = scores_23_cast_fp16)[name = tensor<string, []>("op_1609_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_301_cast_fp16 = select(a = var_44_to_fp16, b = var_1609_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_301_cast_fp16")];
+            tensor<bool, []> x_143_transpose_x_0 = const()[name = tensor<string, []>("x_143_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_143_transpose_y_0 = const()[name = tensor<string, []>("x_143_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_11_cast_fp16)[name = tensor<string, []>("transpose_314")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_143_cast_fp16")];
+            tensor<int32, [4]> var_1613_perm_0 = const()[name = tensor<string, []>("op_1613_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1614 = const()[name = tensor<string, []>("op_1614"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_1613_cast_fp16 = transpose(perm = var_1613_perm_0, x = x_143_cast_fp16)[name = tensor<string, []>("transpose_313")];
+            tensor<fp16, [1, 7, 1024]> input_303_cast_fp16 = reshape(shape = var_1614, x = var_1613_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122098752))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122885248))), name = tensor<string, []>("encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122885440)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_52_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = tensor<string, []>("input_307_cast_fp16")];
+            tensor<int32, [1]> x_147_axes_0 = const()[name = tensor<string, []>("x_147_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122887552)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122889664)))];
+            tensor<fp16, [1, 7, 1024]> x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = tensor<string, []>("x_147_cast_fp16")];
+            tensor<int32, [3]> input_309_perm_0 = const()[name = tensor<string, []>("input_309_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_311_pad_type_0 = const()[name = tensor<string, []>("input_311_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_311_strides_0 = const()[name = tensor<string, []>("input_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_311_pad_0 = const()[name = tensor<string, []>("input_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_311_dilations_0 = const()[name = tensor<string, []>("input_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_311_groups_0 = const()[name = tensor<string, []>("input_311_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122891776))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124988992))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = tensor<string, []>("transpose_312")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_311_cast_fp16")];
+            tensor<int32, []> x_149_split_num_splits_0 = const()[name = tensor<string, []>("x_149_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_149_split_axis_0 = const()[name = tensor<string, []>("x_149_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_149_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_149_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = tensor<string, []>("x_149_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_149_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_313_cast_fp16 = select(a = var_44_to_fp16, b = x_149_cast_fp16, cond = var_575)[name = tensor<string, []>("input_313_cast_fp16")];
+            tensor<bool, []> new_x_23_interleave_0 = const()[name = tensor<string, []>("new_x_23_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_23_cast_fp16 = concat(axis = var_59, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = tensor<string, []>("new_x_23_cast_fp16")];
+            tensor<int32, [3]> var_1653_begin_0 = const()[name = tensor<string, []>("op_1653_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_1653_end_0 = const()[name = tensor<string, []>("op_1653_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_1653_end_mask_0 = const()[name = tensor<string, []>("op_1653_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1653_cast_fp16 = slice_by_index(begin = var_1653_begin_0, end = var_1653_end_0, end_mask = var_1653_end_mask_0, x = new_x_23_cast_fp16)[name = tensor<string, []>("op_1653_cast_fp16")];
+            tensor<string, []> x_151_pad_type_0 = const()[name = tensor<string, []>("x_151_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_151_groups_0 = const()[name = tensor<string, []>("x_151_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_151_strides_0 = const()[name = tensor<string, []>("x_151_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_151_pad_0 = const()[name = tensor<string, []>("x_151_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_151_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124993152))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125002432))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_151_cast_fp16")];
+            tensor<int32, [3]> input_315_perm_0 = const()[name = tensor<string, []>("input_315_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_153_axes_0 = const()[name = tensor<string, []>("x_153_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125004544)))];
+            tensor<fp16, [1024]> encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125006656)))];
+            tensor<fp16, [1, 7, 1024]> input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = tensor<string, []>("transpose_311")];
+            tensor<fp16, [1, 7, 1024]> x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = tensor<string, []>("x_153_cast_fp16")];
+            tensor<int32, [3]> input_317_perm_0 = const()[name = tensor<string, []>("input_317_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = tensor<string, []>("transpose_310")];
+            tensor<fp16, [1, 1024, 7]> input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = tensor<string, []>("input_319_cast_fp16")];
+            tensor<string, []> x_155_pad_type_0 = const()[name = tensor<string, []>("x_155_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_155_strides_0 = const()[name = tensor<string, []>("x_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_155_pad_0 = const()[name = tensor<string, []>("x_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_155_dilations_0 = const()[name = tensor<string, []>("x_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_155_groups_0 = const()[name = tensor<string, []>("x_155_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125008768))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126057408))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_155_cast_fp16")];
+            tensor<int32, [3]> input_321_perm_0 = const()[name = tensor<string, []>("input_321_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = tensor<string, []>("transpose_309")];
+            tensor<fp16, [1, 7, 1024]> input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = tensor<string, []>("input_323_cast_fp16")];
+            tensor<int32, [1]> input_325_axes_0 = const()[name = tensor<string, []>("input_325_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126059520)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126061632)))];
+            tensor<fp16, [1, 7, 1024]> input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126063744))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129209536))), name = tensor<string, []>("encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_5_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129209728)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_53_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129217984))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132363776))), name = tensor<string, []>("encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_5_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132363968)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_54_cast_fp16")];
+            tensor<fp16, []> var_1696_to_fp16 = const()[name = tensor<string, []>("op_1696_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1697_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1696_to_fp16)[name = tensor<string, []>("op_1697_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1697_cast_fp16)[name = tensor<string, []>("input_335_cast_fp16")];
+            tensor<int32, [1]> input_337_axes_0 = const()[name = tensor<string, []>("input_337_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132366080)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132368192)))];
+            tensor<fp16, [1, 7, 1024]> input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = tensor<string, []>("input_337_cast_fp16")];
+            tensor<int32, [4]> cache_25_begin_0 = const()[name = tensor<string, []>("cache_25_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_25_end_0 = const()[name = tensor<string, []>("cache_25_end_0"), val = tensor<int32, [4]>([7, 1, 42, 1024])];
+            tensor<bool, [4]> cache_25_end_mask_0 = const()[name = tensor<string, []>("cache_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_25_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_25_cast_fp16")];
+            tensor<int32, [4]> cache_27_begin_0 = const()[name = tensor<string, []>("cache_27_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_27_end_0 = const()[name = tensor<string, []>("cache_27_end_0"), val = tensor<int32, [4]>([7, 1, 1024, 8])];
+            tensor<bool, [4]> cache_27_end_mask_0 = const()[name = tensor<string, []>("cache_27_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_27_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_27_cast_fp16")];
+            tensor<int32, [1]> input_339_axes_0 = const()[name = tensor<string, []>("input_339_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132370304)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132372416)))];
+            tensor<fp16, [1, 7, 1024]> input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132374528))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135520320))), name = tensor<string, []>("encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_6_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135520512)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_55_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135528768))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138674560))), name = tensor<string, []>("encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_6_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138674752)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_56_cast_fp16")];
+            tensor<fp16, []> var_1733_to_fp16 = const()[name = tensor<string, []>("op_1733_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1734_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1733_to_fp16)[name = tensor<string, []>("op_1734_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1734_cast_fp16)[name = tensor<string, []>("input_349_cast_fp16")];
+            tensor<int32, [1]> key_13_axes_0 = const()[name = tensor<string, []>("key_13_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138676864)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138678976)))];
+            tensor<fp16, [1, 7, 1024]> key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")];
+            tensor<bool, []> input_351_interleave_0 = const()[name = tensor<string, []>("input_351_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_351_cast_fp16 = concat(axis = var_68, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = tensor<string, []>("input_351_cast_fp16")];
+            tensor<int32, [3]> var_1756_begin_0 = const()[name = tensor<string, []>("op_1756_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_1756_end_0 = const()[name = tensor<string, []>("op_1756_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1756_end_mask_0 = const()[name = tensor<string, []>("op_1756_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_1756_cast_fp16 = slice_by_index(begin = var_1756_begin_0, end = var_1756_end_0, end_mask = var_1756_end_mask_0, x = cache_25_cast_fp16)[name = tensor<string, []>("op_1756_cast_fp16")];
+            tensor<bool, []> var_1762_interleave_0 = const()[name = tensor<string, []>("op_1762_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_1762_cast_fp16 = concat(axis = var_68, interleave = var_1762_interleave_0, values = (var_1756_cast_fp16, key_13_cast_fp16))[name = tensor<string, []>("op_1762_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138681088))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139467584))), name = tensor<string, []>("encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139467776)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_57_cast_fp16")];
+            tensor<int32, [4]> var_1767 = const()[name = tensor<string, []>("op_1767"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_37_cast_fp16 = reshape(shape = var_1767, x = linear_57_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139469888))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140256384))), name = tensor<string, []>("encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140256576)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_58_cast_fp16")];
+            tensor<int32, [4]> var_1772 = const()[name = tensor<string, []>("op_1772"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_25_cast_fp16 = reshape(shape = var_1772, x = linear_58_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140258688))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141045184))), name = tensor<string, []>("encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141045376)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_59_cast_fp16")];
+            tensor<int32, [4]> var_1777 = const()[name = tensor<string, []>("op_1777"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_13_cast_fp16 = reshape(shape = var_1777, x = linear_59_cast_fp16)[name = tensor<string, []>("v_13_cast_fp16")];
+            tensor<int32, [4]> value_21_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141047488)))];
+            tensor<fp16, [1, 7, 8, 128]> var_1790_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1790_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141049600)))];
+            tensor<fp16, [1, 7, 8, 128]> var_1792_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1792_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_13_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_163_transpose_x_0 = const()[name = tensor<string, []>("x_163_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_163_transpose_y_0 = const()[name = tensor<string, []>("x_163_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_1794_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1794_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141051712))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141151296))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141151104)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1792_cast_fp16)[name = tensor<string, []>("transpose_308")];
+            tensor<fp16, [1, 8, 7, 97]> 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_1794_to_fp16_quantized)[name = tensor<string, []>("x_163_cast_fp16")];
+            tensor<int32, [8]> x_165_pad_0 = const()[name = tensor<string, []>("x_165_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_165_mode_0 = const()[name = tensor<string, []>("x_165_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_157_to_fp16 = const()[name = tensor<string, []>("const_157_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_165_cast_fp16")];
+            tensor<int32, [4]> var_1802 = const()[name = tensor<string, []>("op_1802"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_167_cast_fp16 = reshape(shape = var_1802, x = x_165_cast_fp16)[name = tensor<string, []>("x_167_cast_fp16")];
+            tensor<int32, [4]> var_1806_begin_0 = const()[name = tensor<string, []>("op_1806_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1806_end_0 = const()[name = tensor<string, []>("op_1806_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_1806_end_mask_0 = const()[name = tensor<string, []>("op_1806_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_1806_cast_fp16 = slice_by_index(begin = var_1806_begin_0, end = var_1806_end_0, end_mask = var_1806_end_mask_0, x = x_167_cast_fp16)[name = tensor<string, []>("op_1806_cast_fp16")];
+            tensor<int32, [4]> var_1807 = const()[name = tensor<string, []>("op_1807"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_25_cast_fp16 = reshape(shape = var_1807, x = var_1806_cast_fp16)[name = tensor<string, []>("matrix_bd_25_cast_fp16")];
+            tensor<bool, []> matrix_ac_13_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_13_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_108_perm_0 = const()[name = tensor<string, []>("transpose_108_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_109_perm_0 = const()[name = tensor<string, []>("transpose_109_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25_cast_fp16)[name = tensor<string, []>("transpose_306")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1790_cast_fp16)[name = tensor<string, []>("transpose_307")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_13_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_27_begin_0 = const()[name = tensor<string, []>("matrix_bd_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_27_end_0 = const()[name = tensor<string, []>("matrix_bd_27_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_27_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_27_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_1816_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = tensor<string, []>("op_1816_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_25_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_25_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_25_cast_fp16 = mul(x = var_1816_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_25_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_27_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_27_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_1822_cast_fp16 = softmax(axis = var_59, x = scores_27_cast_fp16)[name = tensor<string, []>("op_1822_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_353_cast_fp16 = select(a = var_44_to_fp16, b = var_1822_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_353_cast_fp16")];
+            tensor<bool, []> x_169_transpose_x_0 = const()[name = tensor<string, []>("x_169_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_169_transpose_y_0 = const()[name = tensor<string, []>("x_169_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_13_cast_fp16)[name = tensor<string, []>("transpose_305")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_169_cast_fp16")];
+            tensor<int32, [4]> var_1826_perm_0 = const()[name = tensor<string, []>("op_1826_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1827 = const()[name = tensor<string, []>("op_1827"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_1826_cast_fp16 = transpose(perm = var_1826_perm_0, x = x_169_cast_fp16)[name = tensor<string, []>("transpose_304")];
+            tensor<fp16, [1, 7, 1024]> input_355_cast_fp16 = reshape(shape = var_1827, x = var_1826_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141151616))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141938112))), name = tensor<string, []>("encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141938304)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_61_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = tensor<string, []>("input_359_cast_fp16")];
+            tensor<int32, [1]> x_173_axes_0 = const()[name = tensor<string, []>("x_173_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141940416)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141942528)))];
+            tensor<fp16, [1, 7, 1024]> x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = tensor<string, []>("x_173_cast_fp16")];
+            tensor<int32, [3]> input_361_perm_0 = const()[name = tensor<string, []>("input_361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_363_pad_type_0 = const()[name = tensor<string, []>("input_363_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_363_strides_0 = const()[name = tensor<string, []>("input_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_363_pad_0 = const()[name = tensor<string, []>("input_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_363_dilations_0 = const()[name = tensor<string, []>("input_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_363_groups_0 = const()[name = tensor<string, []>("input_363_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141944640))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144041856))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = tensor<string, []>("transpose_303")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_363_cast_fp16")];
+            tensor<int32, []> x_175_split_num_splits_0 = const()[name = tensor<string, []>("x_175_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_175_split_axis_0 = const()[name = tensor<string, []>("x_175_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_175_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_175_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = tensor<string, []>("x_175_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_175_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_365_cast_fp16 = select(a = var_44_to_fp16, b = x_175_cast_fp16, cond = var_575)[name = tensor<string, []>("input_365_cast_fp16")];
+            tensor<bool, []> new_x_27_interleave_0 = const()[name = tensor<string, []>("new_x_27_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_27_cast_fp16 = concat(axis = var_59, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = tensor<string, []>("new_x_27_cast_fp16")];
+            tensor<int32, [3]> var_1866_begin_0 = const()[name = tensor<string, []>("op_1866_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_1866_end_0 = const()[name = tensor<string, []>("op_1866_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_1866_end_mask_0 = const()[name = tensor<string, []>("op_1866_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1866_cast_fp16 = slice_by_index(begin = var_1866_begin_0, end = var_1866_end_0, end_mask = var_1866_end_mask_0, x = new_x_27_cast_fp16)[name = tensor<string, []>("op_1866_cast_fp16")];
+            tensor<string, []> x_177_pad_type_0 = const()[name = tensor<string, []>("x_177_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_177_groups_0 = const()[name = tensor<string, []>("x_177_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_177_strides_0 = const()[name = tensor<string, []>("x_177_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_177_pad_0 = const()[name = tensor<string, []>("x_177_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_177_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144046016))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144055296))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_177_cast_fp16")];
+            tensor<int32, [3]> input_367_perm_0 = const()[name = tensor<string, []>("input_367_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_179_axes_0 = const()[name = tensor<string, []>("x_179_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144057408)))];
+            tensor<fp16, [1024]> encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144059520)))];
+            tensor<fp16, [1, 7, 1024]> input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = tensor<string, []>("transpose_302")];
+            tensor<fp16, [1, 7, 1024]> x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = tensor<string, []>("x_179_cast_fp16")];
+            tensor<int32, [3]> input_369_perm_0 = const()[name = tensor<string, []>("input_369_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = tensor<string, []>("transpose_301")];
+            tensor<fp16, [1, 1024, 7]> input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = tensor<string, []>("input_371_cast_fp16")];
+            tensor<string, []> x_181_pad_type_0 = const()[name = tensor<string, []>("x_181_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_181_strides_0 = const()[name = tensor<string, []>("x_181_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_181_pad_0 = const()[name = tensor<string, []>("x_181_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_181_dilations_0 = const()[name = tensor<string, []>("x_181_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_181_groups_0 = const()[name = tensor<string, []>("x_181_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144061632))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145110272))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_181_cast_fp16")];
+            tensor<int32, [3]> input_373_perm_0 = const()[name = tensor<string, []>("input_373_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = tensor<string, []>("transpose_300")];
+            tensor<fp16, [1, 7, 1024]> input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = tensor<string, []>("input_375_cast_fp16")];
+            tensor<int32, [1]> input_377_axes_0 = const()[name = tensor<string, []>("input_377_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145112384)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145114496)))];
+            tensor<fp16, [1, 7, 1024]> input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145116608))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148262400))), name = tensor<string, []>("encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_6_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148262592)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_62_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148270848))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151416640))), name = tensor<string, []>("encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_6_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151416832)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_63_cast_fp16")];
+            tensor<fp16, []> var_1909_to_fp16 = const()[name = tensor<string, []>("op_1909_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1910_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1909_to_fp16)[name = tensor<string, []>("op_1910_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1910_cast_fp16)[name = tensor<string, []>("input_387_cast_fp16")];
+            tensor<int32, [1]> input_389_axes_0 = const()[name = tensor<string, []>("input_389_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151418944)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151421056)))];
+            tensor<fp16, [1, 7, 1024]> input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = tensor<string, []>("input_389_cast_fp16")];
+            tensor<int32, [4]> cache_29_begin_0 = const()[name = tensor<string, []>("cache_29_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_29_end_0 = const()[name = tensor<string, []>("cache_29_end_0"), val = tensor<int32, [4]>([8, 1, 42, 1024])];
+            tensor<bool, [4]> cache_29_end_mask_0 = const()[name = tensor<string, []>("cache_29_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_29_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_29_cast_fp16")];
+            tensor<int32, [4]> cache_31_begin_0 = const()[name = tensor<string, []>("cache_31_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_31_end_0 = const()[name = tensor<string, []>("cache_31_end_0"), val = tensor<int32, [4]>([8, 1, 1024, 8])];
+            tensor<bool, [4]> cache_31_end_mask_0 = const()[name = tensor<string, []>("cache_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_31_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_31_cast_fp16")];
+            tensor<int32, [1]> input_391_axes_0 = const()[name = tensor<string, []>("input_391_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151423168)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151425280)))];
+            tensor<fp16, [1, 7, 1024]> input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151427392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154573184))), name = tensor<string, []>("encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_7_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154573376)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_64_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154581632))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157727424))), name = tensor<string, []>("encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_7_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157727616)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_65_cast_fp16")];
+            tensor<fp16, []> var_1946_to_fp16 = const()[name = tensor<string, []>("op_1946_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1947_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1946_to_fp16)[name = tensor<string, []>("op_1947_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1947_cast_fp16)[name = tensor<string, []>("input_401_cast_fp16")];
+            tensor<int32, [1]> key_15_axes_0 = const()[name = tensor<string, []>("key_15_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157729728)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157731840)))];
+            tensor<fp16, [1, 7, 1024]> key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = tensor<string, []>("key_15_cast_fp16")];
+            tensor<bool, []> input_403_interleave_0 = const()[name = tensor<string, []>("input_403_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_403_cast_fp16 = concat(axis = var_68, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = tensor<string, []>("input_403_cast_fp16")];
+            tensor<int32, [3]> var_1969_begin_0 = const()[name = tensor<string, []>("op_1969_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_1969_end_0 = const()[name = tensor<string, []>("op_1969_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1969_end_mask_0 = const()[name = tensor<string, []>("op_1969_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_1969_cast_fp16 = slice_by_index(begin = var_1969_begin_0, end = var_1969_end_0, end_mask = var_1969_end_mask_0, x = cache_29_cast_fp16)[name = tensor<string, []>("op_1969_cast_fp16")];
+            tensor<bool, []> var_1975_interleave_0 = const()[name = tensor<string, []>("op_1975_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_1975_cast_fp16 = concat(axis = var_68, interleave = var_1975_interleave_0, values = (var_1969_cast_fp16, key_15_cast_fp16))[name = tensor<string, []>("op_1975_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157733952))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158520448))), name = tensor<string, []>("encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158520640)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_66_cast_fp16")];
+            tensor<int32, [4]> var_1980 = const()[name = tensor<string, []>("op_1980"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_43_cast_fp16 = reshape(shape = var_1980, x = linear_66_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158522752))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159309248))), name = tensor<string, []>("encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159309440)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_67_cast_fp16")];
+            tensor<int32, [4]> var_1985 = const()[name = tensor<string, []>("op_1985"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_29_cast_fp16 = reshape(shape = var_1985, x = linear_67_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159311552))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160098048))), name = tensor<string, []>("encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160098240)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_68_cast_fp16")];
+            tensor<int32, [4]> var_1990 = const()[name = tensor<string, []>("op_1990"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_15_cast_fp16 = reshape(shape = var_1990, x = linear_68_cast_fp16)[name = tensor<string, []>("v_15_cast_fp16")];
+            tensor<int32, [4]> value_23_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160100352)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2003_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2003_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160102464)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2005_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2005_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_15_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_189_transpose_x_0 = const()[name = tensor<string, []>("x_189_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_189_transpose_y_0 = const()[name = tensor<string, []>("x_189_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_2007_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2007_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160104576))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160204160))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160203968)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_2005_cast_fp16)[name = tensor<string, []>("transpose_299")];
+            tensor<fp16, [1, 8, 7, 97]> 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_2007_to_fp16_quantized)[name = tensor<string, []>("x_189_cast_fp16")];
+            tensor<int32, [8]> x_191_pad_0 = const()[name = tensor<string, []>("x_191_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_191_mode_0 = const()[name = tensor<string, []>("x_191_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_170_to_fp16 = const()[name = tensor<string, []>("const_170_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_191_cast_fp16")];
+            tensor<int32, [4]> var_2015 = const()[name = tensor<string, []>("op_2015"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_193_cast_fp16 = reshape(shape = var_2015, x = x_191_cast_fp16)[name = tensor<string, []>("x_193_cast_fp16")];
+            tensor<int32, [4]> var_2019_begin_0 = const()[name = tensor<string, []>("op_2019_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2019_end_0 = const()[name = tensor<string, []>("op_2019_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_2019_end_mask_0 = const()[name = tensor<string, []>("op_2019_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_2019_cast_fp16 = slice_by_index(begin = var_2019_begin_0, end = var_2019_end_0, end_mask = var_2019_end_mask_0, x = x_193_cast_fp16)[name = tensor<string, []>("op_2019_cast_fp16")];
+            tensor<int32, [4]> var_2020 = const()[name = tensor<string, []>("op_2020"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_29_cast_fp16 = reshape(shape = var_2020, x = var_2019_cast_fp16)[name = tensor<string, []>("matrix_bd_29_cast_fp16")];
+            tensor<bool, []> matrix_ac_15_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_15_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_15_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_15_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_110_perm_0 = const()[name = tensor<string, []>("transpose_110_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_111_perm_0 = const()[name = tensor<string, []>("transpose_111_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29_cast_fp16)[name = tensor<string, []>("transpose_297")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_110 = transpose(perm = transpose_110_perm_0, x = var_2003_cast_fp16)[name = tensor<string, []>("transpose_298")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_15_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_31_begin_0 = const()[name = tensor<string, []>("matrix_bd_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_31_end_0 = const()[name = tensor<string, []>("matrix_bd_31_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_31_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_31_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2029_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = tensor<string, []>("op_2029_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_29_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_29_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_29_cast_fp16 = mul(x = var_2029_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_29_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_31_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_31_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2035_cast_fp16 = softmax(axis = var_59, x = scores_31_cast_fp16)[name = tensor<string, []>("op_2035_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_405_cast_fp16 = select(a = var_44_to_fp16, b = var_2035_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_405_cast_fp16")];
+            tensor<bool, []> x_195_transpose_x_0 = const()[name = tensor<string, []>("x_195_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_195_transpose_y_0 = const()[name = tensor<string, []>("x_195_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_15_cast_fp16)[name = tensor<string, []>("transpose_296")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_195_cast_fp16")];
+            tensor<int32, [4]> var_2039_perm_0 = const()[name = tensor<string, []>("op_2039_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2040 = const()[name = tensor<string, []>("op_2040"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_2039_cast_fp16 = transpose(perm = var_2039_perm_0, x = x_195_cast_fp16)[name = tensor<string, []>("transpose_295")];
+            tensor<fp16, [1, 7, 1024]> input_407_cast_fp16 = reshape(shape = var_2040, x = var_2039_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160204480))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160990976))), name = tensor<string, []>("encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160991168)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_70_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = tensor<string, []>("input_411_cast_fp16")];
+            tensor<int32, [1]> x_199_axes_0 = const()[name = tensor<string, []>("x_199_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160993280)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160995392)))];
+            tensor<fp16, [1, 7, 1024]> x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = tensor<string, []>("x_199_cast_fp16")];
+            tensor<int32, [3]> input_413_perm_0 = const()[name = tensor<string, []>("input_413_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_415_pad_type_0 = const()[name = tensor<string, []>("input_415_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_415_strides_0 = const()[name = tensor<string, []>("input_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_415_pad_0 = const()[name = tensor<string, []>("input_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_415_dilations_0 = const()[name = tensor<string, []>("input_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_415_groups_0 = const()[name = tensor<string, []>("input_415_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160997504))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163094720))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = tensor<string, []>("transpose_294")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_415_cast_fp16")];
+            tensor<int32, []> x_201_split_num_splits_0 = const()[name = tensor<string, []>("x_201_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_201_split_axis_0 = const()[name = tensor<string, []>("x_201_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_201_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_201_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = tensor<string, []>("x_201_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_201_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_417_cast_fp16 = select(a = var_44_to_fp16, b = x_201_cast_fp16, cond = var_575)[name = tensor<string, []>("input_417_cast_fp16")];
+            tensor<bool, []> new_x_31_interleave_0 = const()[name = tensor<string, []>("new_x_31_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_31_cast_fp16 = concat(axis = var_59, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = tensor<string, []>("new_x_31_cast_fp16")];
+            tensor<int32, [3]> var_2079_begin_0 = const()[name = tensor<string, []>("op_2079_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_2079_end_0 = const()[name = tensor<string, []>("op_2079_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_2079_end_mask_0 = const()[name = tensor<string, []>("op_2079_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2079_cast_fp16 = slice_by_index(begin = var_2079_begin_0, end = var_2079_end_0, end_mask = var_2079_end_mask_0, x = new_x_31_cast_fp16)[name = tensor<string, []>("op_2079_cast_fp16")];
+            tensor<string, []> x_203_pad_type_0 = const()[name = tensor<string, []>("x_203_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_203_groups_0 = const()[name = tensor<string, []>("x_203_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_203_strides_0 = const()[name = tensor<string, []>("x_203_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_203_pad_0 = const()[name = tensor<string, []>("x_203_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_203_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163098880))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163108160))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_203_cast_fp16")];
+            tensor<int32, [3]> input_419_perm_0 = const()[name = tensor<string, []>("input_419_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_205_axes_0 = const()[name = tensor<string, []>("x_205_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163110272)))];
+            tensor<fp16, [1024]> encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163112384)))];
+            tensor<fp16, [1, 7, 1024]> input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = tensor<string, []>("transpose_293")];
+            tensor<fp16, [1, 7, 1024]> x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = tensor<string, []>("x_205_cast_fp16")];
+            tensor<int32, [3]> input_421_perm_0 = const()[name = tensor<string, []>("input_421_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = tensor<string, []>("transpose_292")];
+            tensor<fp16, [1, 1024, 7]> input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor<string, []>("input_423_cast_fp16")];
+            tensor<string, []> x_207_pad_type_0 = const()[name = tensor<string, []>("x_207_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_207_strides_0 = const()[name = tensor<string, []>("x_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_207_pad_0 = const()[name = tensor<string, []>("x_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_207_dilations_0 = const()[name = tensor<string, []>("x_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_207_groups_0 = const()[name = tensor<string, []>("x_207_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163114496))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164163136))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_207_cast_fp16")];
+            tensor<int32, [3]> input_425_perm_0 = const()[name = tensor<string, []>("input_425_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = tensor<string, []>("transpose_291")];
+            tensor<fp16, [1, 7, 1024]> input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = tensor<string, []>("input_427_cast_fp16")];
+            tensor<int32, [1]> input_429_axes_0 = const()[name = tensor<string, []>("input_429_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164165248)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164167360)))];
+            tensor<fp16, [1, 7, 1024]> input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164169472))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167315264))), name = tensor<string, []>("encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_7_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167315456)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_71_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167323712))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170469504))), name = tensor<string, []>("encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_7_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170469696)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_72_cast_fp16")];
+            tensor<fp16, []> var_2122_to_fp16 = const()[name = tensor<string, []>("op_2122_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2123_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2122_to_fp16)[name = tensor<string, []>("op_2123_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_2123_cast_fp16)[name = tensor<string, []>("input_439_cast_fp16")];
+            tensor<int32, [1]> input_441_axes_0 = const()[name = tensor<string, []>("input_441_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170471808)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170473920)))];
+            tensor<fp16, [1, 7, 1024]> input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = tensor<string, []>("input_441_cast_fp16")];
+            tensor<int32, [4]> cache_33_begin_0 = const()[name = tensor<string, []>("cache_33_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_33_end_0 = const()[name = tensor<string, []>("cache_33_end_0"), val = tensor<int32, [4]>([9, 1, 42, 1024])];
+            tensor<bool, [4]> cache_33_end_mask_0 = const()[name = tensor<string, []>("cache_33_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_33_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_33_cast_fp16")];
+            tensor<int32, [4]> cache_35_begin_0 = const()[name = tensor<string, []>("cache_35_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_35_end_0 = const()[name = tensor<string, []>("cache_35_end_0"), val = tensor<int32, [4]>([9, 1, 1024, 8])];
+            tensor<bool, [4]> cache_35_end_mask_0 = const()[name = tensor<string, []>("cache_35_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_35_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_35_cast_fp16")];
+            tensor<int32, [1]> input_443_axes_0 = const()[name = tensor<string, []>("input_443_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170476032)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170478144)))];
+            tensor<fp16, [1, 7, 1024]> input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170480256))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173626048))), name = tensor<string, []>("encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_8_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173626240)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_73_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173634496))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176780288))), name = tensor<string, []>("encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_8_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176780480)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_74_cast_fp16")];
+            tensor<fp16, []> var_2159_to_fp16 = const()[name = tensor<string, []>("op_2159_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2160_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2159_to_fp16)[name = tensor<string, []>("op_2160_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_2160_cast_fp16)[name = tensor<string, []>("input_453_cast_fp16")];
+            tensor<int32, [1]> key_17_axes_0 = const()[name = tensor<string, []>("key_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176782592)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176784704)))];
+            tensor<fp16, [1, 7, 1024]> key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = tensor<string, []>("key_17_cast_fp16")];
+            tensor<bool, []> input_455_interleave_0 = const()[name = tensor<string, []>("input_455_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_455_cast_fp16 = concat(axis = var_68, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = tensor<string, []>("input_455_cast_fp16")];
+            tensor<int32, [3]> var_2182_begin_0 = const()[name = tensor<string, []>("op_2182_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_2182_end_0 = const()[name = tensor<string, []>("op_2182_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2182_end_mask_0 = const()[name = tensor<string, []>("op_2182_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_2182_cast_fp16 = slice_by_index(begin = var_2182_begin_0, end = var_2182_end_0, end_mask = var_2182_end_mask_0, x = cache_33_cast_fp16)[name = tensor<string, []>("op_2182_cast_fp16")];
+            tensor<bool, []> var_2188_interleave_0 = const()[name = tensor<string, []>("op_2188_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_2188_cast_fp16 = concat(axis = var_68, interleave = var_2188_interleave_0, values = (var_2182_cast_fp16, key_17_cast_fp16))[name = tensor<string, []>("op_2188_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176786816))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177573312))), name = tensor<string, []>("encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177573504)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_75_cast_fp16")];
+            tensor<int32, [4]> var_2193 = const()[name = tensor<string, []>("op_2193"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_49_cast_fp16 = reshape(shape = var_2193, x = linear_75_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177575616))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178362112))), name = tensor<string, []>("encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178362304)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_76_cast_fp16")];
+            tensor<int32, [4]> var_2198 = const()[name = tensor<string, []>("op_2198"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_33_cast_fp16 = reshape(shape = var_2198, x = linear_76_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178364416))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179150912))), name = tensor<string, []>("encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179151104)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_77_cast_fp16")];
+            tensor<int32, [4]> var_2203 = const()[name = tensor<string, []>("op_2203"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_17_cast_fp16 = reshape(shape = var_2203, x = linear_77_cast_fp16)[name = tensor<string, []>("v_17_cast_fp16")];
+            tensor<int32, [4]> value_25_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179153216)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2216_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2216_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179155328)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2218_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2218_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_17_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_215_transpose_x_0 = const()[name = tensor<string, []>("x_215_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_215_transpose_y_0 = const()[name = tensor<string, []>("x_215_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_2220_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2220_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179157440))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179257024))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179256832)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2218_cast_fp16)[name = tensor<string, []>("transpose_290")];
+            tensor<fp16, [1, 8, 7, 97]> 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_2220_to_fp16_quantized)[name = tensor<string, []>("x_215_cast_fp16")];
+            tensor<int32, [8]> x_217_pad_0 = const()[name = tensor<string, []>("x_217_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_217_mode_0 = const()[name = tensor<string, []>("x_217_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_183_to_fp16 = const()[name = tensor<string, []>("const_183_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_217_cast_fp16")];
+            tensor<int32, [4]> var_2228 = const()[name = tensor<string, []>("op_2228"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_219_cast_fp16 = reshape(shape = var_2228, x = x_217_cast_fp16)[name = tensor<string, []>("x_219_cast_fp16")];
+            tensor<int32, [4]> var_2232_begin_0 = const()[name = tensor<string, []>("op_2232_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2232_end_0 = const()[name = tensor<string, []>("op_2232_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_2232_end_mask_0 = const()[name = tensor<string, []>("op_2232_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_2232_cast_fp16 = slice_by_index(begin = var_2232_begin_0, end = var_2232_end_0, end_mask = var_2232_end_mask_0, x = x_219_cast_fp16)[name = tensor<string, []>("op_2232_cast_fp16")];
+            tensor<int32, [4]> var_2233 = const()[name = tensor<string, []>("op_2233"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_33_cast_fp16 = reshape(shape = var_2233, x = var_2232_cast_fp16)[name = tensor<string, []>("matrix_bd_33_cast_fp16")];
+            tensor<bool, []> matrix_ac_17_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_17_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_17_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_17_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_112_perm_0 = const()[name = tensor<string, []>("transpose_112_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_113_perm_0 = const()[name = tensor<string, []>("transpose_113_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33_cast_fp16)[name = tensor<string, []>("transpose_288")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_112 = transpose(perm = transpose_112_perm_0, x = var_2216_cast_fp16)[name = tensor<string, []>("transpose_289")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_17_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_35_begin_0 = const()[name = tensor<string, []>("matrix_bd_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_35_end_0 = const()[name = tensor<string, []>("matrix_bd_35_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_35_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_35_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2242_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = tensor<string, []>("op_2242_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_33_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_33_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_33_cast_fp16 = mul(x = var_2242_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_33_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_35_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_35_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2248_cast_fp16 = softmax(axis = var_59, x = scores_35_cast_fp16)[name = tensor<string, []>("op_2248_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_457_cast_fp16 = select(a = var_44_to_fp16, b = var_2248_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_457_cast_fp16")];
+            tensor<bool, []> x_221_transpose_x_0 = const()[name = tensor<string, []>("x_221_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_221_transpose_y_0 = const()[name = tensor<string, []>("x_221_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_17_cast_fp16)[name = tensor<string, []>("transpose_287")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_221_cast_fp16")];
+            tensor<int32, [4]> var_2252_perm_0 = const()[name = tensor<string, []>("op_2252_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2253 = const()[name = tensor<string, []>("op_2253"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_2252_cast_fp16 = transpose(perm = var_2252_perm_0, x = x_221_cast_fp16)[name = tensor<string, []>("transpose_286")];
+            tensor<fp16, [1, 7, 1024]> input_459_cast_fp16 = reshape(shape = var_2253, x = var_2252_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179257344))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180043840))), name = tensor<string, []>("encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180044032)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_79_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = tensor<string, []>("input_463_cast_fp16")];
+            tensor<int32, [1]> x_225_axes_0 = const()[name = tensor<string, []>("x_225_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180046144)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180048256)))];
+            tensor<fp16, [1, 7, 1024]> x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = tensor<string, []>("x_225_cast_fp16")];
+            tensor<int32, [3]> input_465_perm_0 = const()[name = tensor<string, []>("input_465_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_467_pad_type_0 = const()[name = tensor<string, []>("input_467_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_467_strides_0 = const()[name = tensor<string, []>("input_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_467_pad_0 = const()[name = tensor<string, []>("input_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_467_dilations_0 = const()[name = tensor<string, []>("input_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_467_groups_0 = const()[name = tensor<string, []>("input_467_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180050368))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182147584))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = tensor<string, []>("transpose_285")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_467_cast_fp16")];
+            tensor<int32, []> x_227_split_num_splits_0 = const()[name = tensor<string, []>("x_227_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_227_split_axis_0 = const()[name = tensor<string, []>("x_227_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_227_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_227_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = tensor<string, []>("x_227_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_227_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_469_cast_fp16 = select(a = var_44_to_fp16, b = x_227_cast_fp16, cond = var_575)[name = tensor<string, []>("input_469_cast_fp16")];
+            tensor<bool, []> new_x_35_interleave_0 = const()[name = tensor<string, []>("new_x_35_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_35_cast_fp16 = concat(axis = var_59, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = tensor<string, []>("new_x_35_cast_fp16")];
+            tensor<int32, [3]> var_2292_begin_0 = const()[name = tensor<string, []>("op_2292_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_2292_end_0 = const()[name = tensor<string, []>("op_2292_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_2292_end_mask_0 = const()[name = tensor<string, []>("op_2292_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2292_cast_fp16 = slice_by_index(begin = var_2292_begin_0, end = var_2292_end_0, end_mask = var_2292_end_mask_0, x = new_x_35_cast_fp16)[name = tensor<string, []>("op_2292_cast_fp16")];
+            tensor<string, []> x_229_pad_type_0 = const()[name = tensor<string, []>("x_229_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_229_groups_0 = const()[name = tensor<string, []>("x_229_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_229_strides_0 = const()[name = tensor<string, []>("x_229_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_229_pad_0 = const()[name = tensor<string, []>("x_229_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_229_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182151744))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182161024))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_229_cast_fp16")];
+            tensor<int32, [3]> input_471_perm_0 = const()[name = tensor<string, []>("input_471_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_231_axes_0 = const()[name = tensor<string, []>("x_231_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182163136)))];
+            tensor<fp16, [1024]> encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182165248)))];
+            tensor<fp16, [1, 7, 1024]> input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = tensor<string, []>("transpose_284")];
+            tensor<fp16, [1, 7, 1024]> x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = tensor<string, []>("x_231_cast_fp16")];
+            tensor<int32, [3]> input_473_perm_0 = const()[name = tensor<string, []>("input_473_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = tensor<string, []>("transpose_283")];
+            tensor<fp16, [1, 1024, 7]> input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = tensor<string, []>("input_475_cast_fp16")];
+            tensor<string, []> x_233_pad_type_0 = const()[name = tensor<string, []>("x_233_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_233_strides_0 = const()[name = tensor<string, []>("x_233_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_233_pad_0 = const()[name = tensor<string, []>("x_233_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_233_dilations_0 = const()[name = tensor<string, []>("x_233_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_233_groups_0 = const()[name = tensor<string, []>("x_233_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182167360))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183216000))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_233_cast_fp16")];
+            tensor<int32, [3]> input_477_perm_0 = const()[name = tensor<string, []>("input_477_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = tensor<string, []>("transpose_282")];
+            tensor<fp16, [1, 7, 1024]> input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = tensor<string, []>("input_479_cast_fp16")];
+            tensor<int32, [1]> input_481_axes_0 = const()[name = tensor<string, []>("input_481_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183218112)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183220224)))];
+            tensor<fp16, [1, 7, 1024]> input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183222336))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186368128))), name = tensor<string, []>("encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_8_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186368320)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_80_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186376576))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189522368))), name = tensor<string, []>("encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_8_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189522560)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_81_cast_fp16")];
+            tensor<fp16, []> var_2335_to_fp16 = const()[name = tensor<string, []>("op_2335_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2336_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2335_to_fp16)[name = tensor<string, []>("op_2336_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2336_cast_fp16)[name = tensor<string, []>("input_491_cast_fp16")];
+            tensor<int32, [1]> input_493_axes_0 = const()[name = tensor<string, []>("input_493_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189524672)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189526784)))];
+            tensor<fp16, [1, 7, 1024]> input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = tensor<string, []>("input_493_cast_fp16")];
+            tensor<int32, [4]> cache_37_begin_0 = const()[name = tensor<string, []>("cache_37_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_37_end_0 = const()[name = tensor<string, []>("cache_37_end_0"), val = tensor<int32, [4]>([10, 1, 42, 1024])];
+            tensor<bool, [4]> cache_37_end_mask_0 = const()[name = tensor<string, []>("cache_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_37_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_37_cast_fp16")];
+            tensor<int32, [4]> cache_39_begin_0 = const()[name = tensor<string, []>("cache_39_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_39_end_0 = const()[name = tensor<string, []>("cache_39_end_0"), val = tensor<int32, [4]>([10, 1, 1024, 8])];
+            tensor<bool, [4]> cache_39_end_mask_0 = const()[name = tensor<string, []>("cache_39_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_39_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_39_cast_fp16")];
+            tensor<int32, [1]> input_495_axes_0 = const()[name = tensor<string, []>("input_495_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189528896)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189531008)))];
+            tensor<fp16, [1, 7, 1024]> input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189533120))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192678912))), name = tensor<string, []>("encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_9_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192679104)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_82_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192687360))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195833152))), name = tensor<string, []>("encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_9_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195833344)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_83_cast_fp16")];
+            tensor<fp16, []> var_2372_to_fp16 = const()[name = tensor<string, []>("op_2372_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2373_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2372_to_fp16)[name = tensor<string, []>("op_2373_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2373_cast_fp16)[name = tensor<string, []>("input_505_cast_fp16")];
+            tensor<int32, [1]> key_19_axes_0 = const()[name = tensor<string, []>("key_19_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195835456)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195837568)))];
+            tensor<fp16, [1, 7, 1024]> key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = tensor<string, []>("key_19_cast_fp16")];
+            tensor<bool, []> input_507_interleave_0 = const()[name = tensor<string, []>("input_507_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_507_cast_fp16 = concat(axis = var_68, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = tensor<string, []>("input_507_cast_fp16")];
+            tensor<int32, [3]> var_2395_begin_0 = const()[name = tensor<string, []>("op_2395_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_2395_end_0 = const()[name = tensor<string, []>("op_2395_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2395_end_mask_0 = const()[name = tensor<string, []>("op_2395_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_2395_cast_fp16 = slice_by_index(begin = var_2395_begin_0, end = var_2395_end_0, end_mask = var_2395_end_mask_0, x = cache_37_cast_fp16)[name = tensor<string, []>("op_2395_cast_fp16")];
+            tensor<bool, []> var_2401_interleave_0 = const()[name = tensor<string, []>("op_2401_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_2401_cast_fp16 = concat(axis = var_68, interleave = var_2401_interleave_0, values = (var_2395_cast_fp16, key_19_cast_fp16))[name = tensor<string, []>("op_2401_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195839680))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196626176))), name = tensor<string, []>("encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196626368)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_84_cast_fp16")];
+            tensor<int32, [4]> var_2406 = const()[name = tensor<string, []>("op_2406"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_55_cast_fp16 = reshape(shape = var_2406, x = linear_84_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196628480))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197414976))), name = tensor<string, []>("encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197415168)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_85_cast_fp16")];
+            tensor<int32, [4]> var_2411 = const()[name = tensor<string, []>("op_2411"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_37_cast_fp16 = reshape(shape = var_2411, x = linear_85_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197417280))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198203776))), name = tensor<string, []>("encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198203968)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_86_cast_fp16")];
+            tensor<int32, [4]> var_2416 = const()[name = tensor<string, []>("op_2416"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_19_cast_fp16 = reshape(shape = var_2416, x = linear_86_cast_fp16)[name = tensor<string, []>("v_19_cast_fp16")];
+            tensor<int32, [4]> value_27_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198206080)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2429_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2429_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198208192)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2431_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2431_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_19_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_241_transpose_x_0 = const()[name = tensor<string, []>("x_241_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_241_transpose_y_0 = const()[name = tensor<string, []>("x_241_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_2433_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2433_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198210304))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198309888))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198309696)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2431_cast_fp16)[name = tensor<string, []>("transpose_281")];
+            tensor<fp16, [1, 8, 7, 97]> 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_2433_to_fp16_quantized)[name = tensor<string, []>("x_241_cast_fp16")];
+            tensor<int32, [8]> x_243_pad_0 = const()[name = tensor<string, []>("x_243_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_243_mode_0 = const()[name = tensor<string, []>("x_243_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_196_to_fp16 = const()[name = tensor<string, []>("const_196_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_243_cast_fp16")];
+            tensor<int32, [4]> var_2441 = const()[name = tensor<string, []>("op_2441"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_245_cast_fp16 = reshape(shape = var_2441, x = x_243_cast_fp16)[name = tensor<string, []>("x_245_cast_fp16")];
+            tensor<int32, [4]> var_2445_begin_0 = const()[name = tensor<string, []>("op_2445_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2445_end_0 = const()[name = tensor<string, []>("op_2445_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_2445_end_mask_0 = const()[name = tensor<string, []>("op_2445_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_2445_cast_fp16 = slice_by_index(begin = var_2445_begin_0, end = var_2445_end_0, end_mask = var_2445_end_mask_0, x = x_245_cast_fp16)[name = tensor<string, []>("op_2445_cast_fp16")];
+            tensor<int32, [4]> var_2446 = const()[name = tensor<string, []>("op_2446"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_37_cast_fp16 = reshape(shape = var_2446, x = var_2445_cast_fp16)[name = tensor<string, []>("matrix_bd_37_cast_fp16")];
+            tensor<bool, []> matrix_ac_19_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_19_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_19_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_19_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_114_perm_0 = const()[name = tensor<string, []>("transpose_114_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_115_perm_0 = const()[name = tensor<string, []>("transpose_115_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37_cast_fp16)[name = tensor<string, []>("transpose_279")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_114 = transpose(perm = transpose_114_perm_0, x = var_2429_cast_fp16)[name = tensor<string, []>("transpose_280")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_19_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_39_begin_0 = const()[name = tensor<string, []>("matrix_bd_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_39_end_0 = const()[name = tensor<string, []>("matrix_bd_39_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_39_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_39_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2455_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = tensor<string, []>("op_2455_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_37_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_37_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_37_cast_fp16 = mul(x = var_2455_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_37_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_39_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_39_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2461_cast_fp16 = softmax(axis = var_59, x = scores_39_cast_fp16)[name = tensor<string, []>("op_2461_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_509_cast_fp16 = select(a = var_44_to_fp16, b = var_2461_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_509_cast_fp16")];
+            tensor<bool, []> x_247_transpose_x_0 = const()[name = tensor<string, []>("x_247_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_247_transpose_y_0 = const()[name = tensor<string, []>("x_247_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_19_cast_fp16)[name = tensor<string, []>("transpose_278")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_247_cast_fp16")];
+            tensor<int32, [4]> var_2465_perm_0 = const()[name = tensor<string, []>("op_2465_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2466 = const()[name = tensor<string, []>("op_2466"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_2465_cast_fp16 = transpose(perm = var_2465_perm_0, x = x_247_cast_fp16)[name = tensor<string, []>("transpose_277")];
+            tensor<fp16, [1, 7, 1024]> input_511_cast_fp16 = reshape(shape = var_2466, x = var_2465_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198310208))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199096704))), name = tensor<string, []>("encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199096896)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_88_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = tensor<string, []>("input_515_cast_fp16")];
+            tensor<int32, [1]> x_251_axes_0 = const()[name = tensor<string, []>("x_251_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199099008)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199101120)))];
+            tensor<fp16, [1, 7, 1024]> x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = tensor<string, []>("x_251_cast_fp16")];
+            tensor<int32, [3]> input_517_perm_0 = const()[name = tensor<string, []>("input_517_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_519_pad_type_0 = const()[name = tensor<string, []>("input_519_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_519_strides_0 = const()[name = tensor<string, []>("input_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_519_pad_0 = const()[name = tensor<string, []>("input_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_519_dilations_0 = const()[name = tensor<string, []>("input_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_519_groups_0 = const()[name = tensor<string, []>("input_519_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199103232))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201200448))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = tensor<string, []>("transpose_276")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_519_cast_fp16")];
+            tensor<int32, []> x_253_split_num_splits_0 = const()[name = tensor<string, []>("x_253_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_253_split_axis_0 = const()[name = tensor<string, []>("x_253_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_253_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_253_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = tensor<string, []>("x_253_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_253_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_521_cast_fp16 = select(a = var_44_to_fp16, b = x_253_cast_fp16, cond = var_575)[name = tensor<string, []>("input_521_cast_fp16")];
+            tensor<bool, []> new_x_39_interleave_0 = const()[name = tensor<string, []>("new_x_39_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_39_cast_fp16 = concat(axis = var_59, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = tensor<string, []>("new_x_39_cast_fp16")];
+            tensor<int32, [3]> var_2505_begin_0 = const()[name = tensor<string, []>("op_2505_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_2505_end_0 = const()[name = tensor<string, []>("op_2505_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_2505_end_mask_0 = const()[name = tensor<string, []>("op_2505_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2505_cast_fp16 = slice_by_index(begin = var_2505_begin_0, end = var_2505_end_0, end_mask = var_2505_end_mask_0, x = new_x_39_cast_fp16)[name = tensor<string, []>("op_2505_cast_fp16")];
+            tensor<string, []> x_255_pad_type_0 = const()[name = tensor<string, []>("x_255_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_255_groups_0 = const()[name = tensor<string, []>("x_255_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_255_strides_0 = const()[name = tensor<string, []>("x_255_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_255_pad_0 = const()[name = tensor<string, []>("x_255_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_255_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201204608))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201213888))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_255_cast_fp16")];
+            tensor<int32, [3]> input_523_perm_0 = const()[name = tensor<string, []>("input_523_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_257_axes_0 = const()[name = tensor<string, []>("x_257_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201216000)))];
+            tensor<fp16, [1024]> encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201218112)))];
+            tensor<fp16, [1, 7, 1024]> input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = tensor<string, []>("transpose_275")];
+            tensor<fp16, [1, 7, 1024]> x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = tensor<string, []>("x_257_cast_fp16")];
+            tensor<int32, [3]> input_525_perm_0 = const()[name = tensor<string, []>("input_525_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = tensor<string, []>("transpose_274")];
+            tensor<fp16, [1, 1024, 7]> input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = tensor<string, []>("input_527_cast_fp16")];
+            tensor<string, []> x_259_pad_type_0 = const()[name = tensor<string, []>("x_259_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_259_strides_0 = const()[name = tensor<string, []>("x_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_259_pad_0 = const()[name = tensor<string, []>("x_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_259_dilations_0 = const()[name = tensor<string, []>("x_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_259_groups_0 = const()[name = tensor<string, []>("x_259_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201220224))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202268864))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_259_cast_fp16")];
+            tensor<int32, [3]> input_529_perm_0 = const()[name = tensor<string, []>("input_529_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = tensor<string, []>("transpose_273")];
+            tensor<fp16, [1, 7, 1024]> input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = tensor<string, []>("input_531_cast_fp16")];
+            tensor<int32, [1]> input_533_axes_0 = const()[name = tensor<string, []>("input_533_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202270976)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202273088)))];
+            tensor<fp16, [1, 7, 1024]> input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202275200))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205420992))), name = tensor<string, []>("encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_9_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205421184)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_89_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205429440))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208575232))), name = tensor<string, []>("encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_9_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208575424)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_90_cast_fp16")];
+            tensor<fp16, []> var_2548_to_fp16 = const()[name = tensor<string, []>("op_2548_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2549_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2548_to_fp16)[name = tensor<string, []>("op_2549_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2549_cast_fp16)[name = tensor<string, []>("input_543_cast_fp16")];
+            tensor<int32, [1]> input_545_axes_0 = const()[name = tensor<string, []>("input_545_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208577536)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208579648)))];
+            tensor<fp16, [1, 7, 1024]> input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = tensor<string, []>("input_545_cast_fp16")];
+            tensor<int32, [4]> cache_41_begin_0 = const()[name = tensor<string, []>("cache_41_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_41_end_0 = const()[name = tensor<string, []>("cache_41_end_0"), val = tensor<int32, [4]>([11, 1, 42, 1024])];
+            tensor<bool, [4]> cache_41_end_mask_0 = const()[name = tensor<string, []>("cache_41_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_41_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_41_cast_fp16")];
+            tensor<int32, [4]> cache_43_begin_0 = const()[name = tensor<string, []>("cache_43_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_43_end_0 = const()[name = tensor<string, []>("cache_43_end_0"), val = tensor<int32, [4]>([11, 1, 1024, 8])];
+            tensor<bool, [4]> cache_43_end_mask_0 = const()[name = tensor<string, []>("cache_43_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_43_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_43_cast_fp16")];
+            tensor<int32, [1]> input_547_axes_0 = const()[name = tensor<string, []>("input_547_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208581760)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208583872)))];
+            tensor<fp16, [1, 7, 1024]> input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208585984))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211731776))), name = tensor<string, []>("encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_10_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211731968)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_91_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211740224))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214886016))), name = tensor<string, []>("encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_10_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214886208)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_92_cast_fp16")];
+            tensor<fp16, []> var_2585_to_fp16 = const()[name = tensor<string, []>("op_2585_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2586_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2585_to_fp16)[name = tensor<string, []>("op_2586_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2586_cast_fp16)[name = tensor<string, []>("input_557_cast_fp16")];
+            tensor<int32, [1]> key_21_axes_0 = const()[name = tensor<string, []>("key_21_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214888320)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214890432)))];
+            tensor<fp16, [1, 7, 1024]> key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = tensor<string, []>("key_21_cast_fp16")];
+            tensor<bool, []> input_559_interleave_0 = const()[name = tensor<string, []>("input_559_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_559_cast_fp16 = concat(axis = var_68, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = tensor<string, []>("input_559_cast_fp16")];
+            tensor<int32, [3]> var_2608_begin_0 = const()[name = tensor<string, []>("op_2608_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_2608_end_0 = const()[name = tensor<string, []>("op_2608_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2608_end_mask_0 = const()[name = tensor<string, []>("op_2608_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_2608_cast_fp16 = slice_by_index(begin = var_2608_begin_0, end = var_2608_end_0, end_mask = var_2608_end_mask_0, x = cache_41_cast_fp16)[name = tensor<string, []>("op_2608_cast_fp16")];
+            tensor<bool, []> var_2614_interleave_0 = const()[name = tensor<string, []>("op_2614_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_2614_cast_fp16 = concat(axis = var_68, interleave = var_2614_interleave_0, values = (var_2608_cast_fp16, key_21_cast_fp16))[name = tensor<string, []>("op_2614_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214892544))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215679040))), name = tensor<string, []>("encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215679232)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_93_cast_fp16")];
+            tensor<int32, [4]> var_2619 = const()[name = tensor<string, []>("op_2619"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_61_cast_fp16 = reshape(shape = var_2619, x = linear_93_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215681344))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216467840))), name = tensor<string, []>("encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216468032)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_94_cast_fp16")];
+            tensor<int32, [4]> var_2624 = const()[name = tensor<string, []>("op_2624"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_41_cast_fp16 = reshape(shape = var_2624, x = linear_94_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216470144))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217256640))), name = tensor<string, []>("encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217256832)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_95_cast_fp16")];
+            tensor<int32, [4]> var_2629 = const()[name = tensor<string, []>("op_2629"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_21_cast_fp16 = reshape(shape = var_2629, x = linear_95_cast_fp16)[name = tensor<string, []>("v_21_cast_fp16")];
+            tensor<int32, [4]> value_29_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217258944)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2642_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2642_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217261056)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2644_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2644_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_21_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_267_transpose_x_0 = const()[name = tensor<string, []>("x_267_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_267_transpose_y_0 = const()[name = tensor<string, []>("x_267_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_2646_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2646_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217263168))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217362752))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217362560)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2644_cast_fp16)[name = tensor<string, []>("transpose_272")];
+            tensor<fp16, [1, 8, 7, 97]> 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_2646_to_fp16_quantized)[name = tensor<string, []>("x_267_cast_fp16")];
+            tensor<int32, [8]> x_269_pad_0 = const()[name = tensor<string, []>("x_269_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_269_mode_0 = const()[name = tensor<string, []>("x_269_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_209_to_fp16 = const()[name = tensor<string, []>("const_209_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_269_cast_fp16")];
+            tensor<int32, [4]> var_2654 = const()[name = tensor<string, []>("op_2654"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_271_cast_fp16 = reshape(shape = var_2654, x = x_269_cast_fp16)[name = tensor<string, []>("x_271_cast_fp16")];
+            tensor<int32, [4]> var_2658_begin_0 = const()[name = tensor<string, []>("op_2658_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2658_end_0 = const()[name = tensor<string, []>("op_2658_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_2658_end_mask_0 = const()[name = tensor<string, []>("op_2658_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_2658_cast_fp16 = slice_by_index(begin = var_2658_begin_0, end = var_2658_end_0, end_mask = var_2658_end_mask_0, x = x_271_cast_fp16)[name = tensor<string, []>("op_2658_cast_fp16")];
+            tensor<int32, [4]> var_2659 = const()[name = tensor<string, []>("op_2659"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_41_cast_fp16 = reshape(shape = var_2659, x = var_2658_cast_fp16)[name = tensor<string, []>("matrix_bd_41_cast_fp16")];
+            tensor<bool, []> matrix_ac_21_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_21_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_21_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_21_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_116_perm_0 = const()[name = tensor<string, []>("transpose_116_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_117_perm_0 = const()[name = tensor<string, []>("transpose_117_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41_cast_fp16)[name = tensor<string, []>("transpose_270")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_116 = transpose(perm = transpose_116_perm_0, x = var_2642_cast_fp16)[name = tensor<string, []>("transpose_271")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_21_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_43_begin_0 = const()[name = tensor<string, []>("matrix_bd_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_43_end_0 = const()[name = tensor<string, []>("matrix_bd_43_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_43_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_43_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2668_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = tensor<string, []>("op_2668_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_41_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_41_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_41_cast_fp16 = mul(x = var_2668_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_41_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_43_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_43_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2674_cast_fp16 = softmax(axis = var_59, x = scores_43_cast_fp16)[name = tensor<string, []>("op_2674_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_561_cast_fp16 = select(a = var_44_to_fp16, b = var_2674_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_561_cast_fp16")];
+            tensor<bool, []> x_273_transpose_x_0 = const()[name = tensor<string, []>("x_273_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_273_transpose_y_0 = const()[name = tensor<string, []>("x_273_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_21_cast_fp16)[name = tensor<string, []>("transpose_269")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_273_cast_fp16")];
+            tensor<int32, [4]> var_2678_perm_0 = const()[name = tensor<string, []>("op_2678_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2679 = const()[name = tensor<string, []>("op_2679"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_2678_cast_fp16 = transpose(perm = var_2678_perm_0, x = x_273_cast_fp16)[name = tensor<string, []>("transpose_268")];
+            tensor<fp16, [1, 7, 1024]> input_563_cast_fp16 = reshape(shape = var_2679, x = var_2678_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217363072))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218149568))), name = tensor<string, []>("encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218149760)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_97_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = tensor<string, []>("input_567_cast_fp16")];
+            tensor<int32, [1]> x_277_axes_0 = const()[name = tensor<string, []>("x_277_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218151872)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218153984)))];
+            tensor<fp16, [1, 7, 1024]> x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = tensor<string, []>("x_277_cast_fp16")];
+            tensor<int32, [3]> input_569_perm_0 = const()[name = tensor<string, []>("input_569_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_571_pad_type_0 = const()[name = tensor<string, []>("input_571_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_571_strides_0 = const()[name = tensor<string, []>("input_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_571_pad_0 = const()[name = tensor<string, []>("input_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_571_dilations_0 = const()[name = tensor<string, []>("input_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_571_groups_0 = const()[name = tensor<string, []>("input_571_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218156096))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220253312))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = tensor<string, []>("transpose_267")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_571_cast_fp16")];
+            tensor<int32, []> x_279_split_num_splits_0 = const()[name = tensor<string, []>("x_279_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_279_split_axis_0 = const()[name = tensor<string, []>("x_279_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_279_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_279_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = tensor<string, []>("x_279_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_279_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_573_cast_fp16 = select(a = var_44_to_fp16, b = x_279_cast_fp16, cond = var_575)[name = tensor<string, []>("input_573_cast_fp16")];
+            tensor<bool, []> new_x_43_interleave_0 = const()[name = tensor<string, []>("new_x_43_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_43_cast_fp16 = concat(axis = var_59, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = tensor<string, []>("new_x_43_cast_fp16")];
+            tensor<int32, [3]> var_2718_begin_0 = const()[name = tensor<string, []>("op_2718_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_2718_end_0 = const()[name = tensor<string, []>("op_2718_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_2718_end_mask_0 = const()[name = tensor<string, []>("op_2718_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2718_cast_fp16 = slice_by_index(begin = var_2718_begin_0, end = var_2718_end_0, end_mask = var_2718_end_mask_0, x = new_x_43_cast_fp16)[name = tensor<string, []>("op_2718_cast_fp16")];
+            tensor<string, []> x_281_pad_type_0 = const()[name = tensor<string, []>("x_281_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_281_groups_0 = const()[name = tensor<string, []>("x_281_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_281_strides_0 = const()[name = tensor<string, []>("x_281_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_281_pad_0 = const()[name = tensor<string, []>("x_281_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_281_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220257472))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220266752))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_281_cast_fp16")];
+            tensor<int32, [3]> input_575_perm_0 = const()[name = tensor<string, []>("input_575_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_283_axes_0 = const()[name = tensor<string, []>("x_283_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220268864)))];
+            tensor<fp16, [1024]> encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220270976)))];
+            tensor<fp16, [1, 7, 1024]> input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = tensor<string, []>("transpose_266")];
+            tensor<fp16, [1, 7, 1024]> x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = tensor<string, []>("x_283_cast_fp16")];
+            tensor<int32, [3]> input_577_perm_0 = const()[name = tensor<string, []>("input_577_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = tensor<string, []>("transpose_265")];
+            tensor<fp16, [1, 1024, 7]> input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = tensor<string, []>("input_579_cast_fp16")];
+            tensor<string, []> x_285_pad_type_0 = const()[name = tensor<string, []>("x_285_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_285_strides_0 = const()[name = tensor<string, []>("x_285_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_285_pad_0 = const()[name = tensor<string, []>("x_285_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_285_dilations_0 = const()[name = tensor<string, []>("x_285_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_285_groups_0 = const()[name = tensor<string, []>("x_285_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220273088))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221321728))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_285_cast_fp16")];
+            tensor<int32, [3]> input_581_perm_0 = const()[name = tensor<string, []>("input_581_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = tensor<string, []>("transpose_264")];
+            tensor<fp16, [1, 7, 1024]> input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = tensor<string, []>("input_583_cast_fp16")];
+            tensor<int32, [1]> input_585_axes_0 = const()[name = tensor<string, []>("input_585_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221323840)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221325952)))];
+            tensor<fp16, [1, 7, 1024]> input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221328064))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224473856))), name = tensor<string, []>("encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_10_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224474048)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_98_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224482304))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227628096))), name = tensor<string, []>("encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_10_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227628288)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_99_cast_fp16")];
+            tensor<fp16, []> var_2761_to_fp16 = const()[name = tensor<string, []>("op_2761_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2762_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2761_to_fp16)[name = tensor<string, []>("op_2762_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2762_cast_fp16)[name = tensor<string, []>("input_595_cast_fp16")];
+            tensor<int32, [1]> input_597_axes_0 = const()[name = tensor<string, []>("input_597_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227630400)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227632512)))];
+            tensor<fp16, [1, 7, 1024]> input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = tensor<string, []>("input_597_cast_fp16")];
+            tensor<int32, [4]> cache_45_begin_0 = const()[name = tensor<string, []>("cache_45_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_45_end_0 = const()[name = tensor<string, []>("cache_45_end_0"), val = tensor<int32, [4]>([12, 1, 42, 1024])];
+            tensor<bool, [4]> cache_45_end_mask_0 = const()[name = tensor<string, []>("cache_45_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_45_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_45_cast_fp16")];
+            tensor<int32, [4]> cache_47_begin_0 = const()[name = tensor<string, []>("cache_47_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_47_end_0 = const()[name = tensor<string, []>("cache_47_end_0"), val = tensor<int32, [4]>([12, 1, 1024, 8])];
+            tensor<bool, [4]> cache_47_end_mask_0 = const()[name = tensor<string, []>("cache_47_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_47_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_47_cast_fp16")];
+            tensor<int32, [1]> input_599_axes_0 = const()[name = tensor<string, []>("input_599_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227634624)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227636736)))];
+            tensor<fp16, [1, 7, 1024]> input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227638848))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(230784640))), name = tensor<string, []>("encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_11_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(230784832)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_100_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(230793088))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233938880))), name = tensor<string, []>("encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_11_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233939072)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_101_cast_fp16")];
+            tensor<fp16, []> var_2798_to_fp16 = const()[name = tensor<string, []>("op_2798_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2799_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2798_to_fp16)[name = tensor<string, []>("op_2799_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2799_cast_fp16)[name = tensor<string, []>("input_609_cast_fp16")];
+            tensor<int32, [1]> key_23_axes_0 = const()[name = tensor<string, []>("key_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233941184)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233943296)))];
+            tensor<fp16, [1, 7, 1024]> key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = tensor<string, []>("key_23_cast_fp16")];
+            tensor<bool, []> input_611_interleave_0 = const()[name = tensor<string, []>("input_611_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_611_cast_fp16 = concat(axis = var_68, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = tensor<string, []>("input_611_cast_fp16")];
+            tensor<int32, [3]> var_2821_begin_0 = const()[name = tensor<string, []>("op_2821_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_2821_end_0 = const()[name = tensor<string, []>("op_2821_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2821_end_mask_0 = const()[name = tensor<string, []>("op_2821_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_2821_cast_fp16 = slice_by_index(begin = var_2821_begin_0, end = var_2821_end_0, end_mask = var_2821_end_mask_0, x = cache_45_cast_fp16)[name = tensor<string, []>("op_2821_cast_fp16")];
+            tensor<bool, []> var_2827_interleave_0 = const()[name = tensor<string, []>("op_2827_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_2827_cast_fp16 = concat(axis = var_68, interleave = var_2827_interleave_0, values = (var_2821_cast_fp16, key_23_cast_fp16))[name = tensor<string, []>("op_2827_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233945408))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234731904))), name = tensor<string, []>("encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234732096)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_102_cast_fp16")];
+            tensor<int32, [4]> var_2832 = const()[name = tensor<string, []>("op_2832"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_67_cast_fp16 = reshape(shape = var_2832, x = linear_102_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234734208))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235520704))), name = tensor<string, []>("encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235520896)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_103_cast_fp16")];
+            tensor<int32, [4]> var_2837 = const()[name = tensor<string, []>("op_2837"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_45_cast_fp16 = reshape(shape = var_2837, x = linear_103_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235523008))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236309504))), name = tensor<string, []>("encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236309696)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_104_cast_fp16")];
+            tensor<int32, [4]> var_2842 = const()[name = tensor<string, []>("op_2842"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_23_cast_fp16 = reshape(shape = var_2842, x = linear_104_cast_fp16)[name = tensor<string, []>("v_23_cast_fp16")];
+            tensor<int32, [4]> value_31_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236311808)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2855_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2855_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236313920)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2857_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2857_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_23_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_293_transpose_x_0 = const()[name = tensor<string, []>("x_293_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_293_transpose_y_0 = const()[name = tensor<string, []>("x_293_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_2859_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2859_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236316032))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236415616))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236415424)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2857_cast_fp16)[name = tensor<string, []>("transpose_263")];
+            tensor<fp16, [1, 8, 7, 97]> 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_2859_to_fp16_quantized)[name = tensor<string, []>("x_293_cast_fp16")];
+            tensor<int32, [8]> x_295_pad_0 = const()[name = tensor<string, []>("x_295_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_295_mode_0 = const()[name = tensor<string, []>("x_295_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_222_to_fp16 = const()[name = tensor<string, []>("const_222_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_295_cast_fp16")];
+            tensor<int32, [4]> var_2867 = const()[name = tensor<string, []>("op_2867"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_297_cast_fp16 = reshape(shape = var_2867, x = x_295_cast_fp16)[name = tensor<string, []>("x_297_cast_fp16")];
+            tensor<int32, [4]> var_2871_begin_0 = const()[name = tensor<string, []>("op_2871_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2871_end_0 = const()[name = tensor<string, []>("op_2871_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_2871_end_mask_0 = const()[name = tensor<string, []>("op_2871_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_2871_cast_fp16 = slice_by_index(begin = var_2871_begin_0, end = var_2871_end_0, end_mask = var_2871_end_mask_0, x = x_297_cast_fp16)[name = tensor<string, []>("op_2871_cast_fp16")];
+            tensor<int32, [4]> var_2872 = const()[name = tensor<string, []>("op_2872"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_45_cast_fp16 = reshape(shape = var_2872, x = var_2871_cast_fp16)[name = tensor<string, []>("matrix_bd_45_cast_fp16")];
+            tensor<bool, []> matrix_ac_23_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_23_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_23_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_23_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_118_perm_0 = const()[name = tensor<string, []>("transpose_118_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_119_perm_0 = const()[name = tensor<string, []>("transpose_119_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45_cast_fp16)[name = tensor<string, []>("transpose_261")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_118 = transpose(perm = transpose_118_perm_0, x = var_2855_cast_fp16)[name = tensor<string, []>("transpose_262")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_23_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_47_begin_0 = const()[name = tensor<string, []>("matrix_bd_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_47_end_0 = const()[name = tensor<string, []>("matrix_bd_47_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_47_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_47_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2881_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = tensor<string, []>("op_2881_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_45_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_45_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_45_cast_fp16 = mul(x = var_2881_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_45_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_47_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_47_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2887_cast_fp16 = softmax(axis = var_59, x = scores_47_cast_fp16)[name = tensor<string, []>("op_2887_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_613_cast_fp16 = select(a = var_44_to_fp16, b = var_2887_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_613_cast_fp16")];
+            tensor<bool, []> x_299_transpose_x_0 = const()[name = tensor<string, []>("x_299_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_299_transpose_y_0 = const()[name = tensor<string, []>("x_299_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_23_cast_fp16)[name = tensor<string, []>("transpose_260")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_299_cast_fp16")];
+            tensor<int32, [4]> var_2891_perm_0 = const()[name = tensor<string, []>("op_2891_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2892 = const()[name = tensor<string, []>("op_2892"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_2891_cast_fp16 = transpose(perm = var_2891_perm_0, x = x_299_cast_fp16)[name = tensor<string, []>("transpose_259")];
+            tensor<fp16, [1, 7, 1024]> input_615_cast_fp16 = reshape(shape = var_2892, x = var_2891_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236415936))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237202432))), name = tensor<string, []>("encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237202624)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_106_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = tensor<string, []>("input_619_cast_fp16")];
+            tensor<int32, [1]> x_303_axes_0 = const()[name = tensor<string, []>("x_303_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237204736)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237206848)))];
+            tensor<fp16, [1, 7, 1024]> x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = tensor<string, []>("x_303_cast_fp16")];
+            tensor<int32, [3]> input_621_perm_0 = const()[name = tensor<string, []>("input_621_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_623_pad_type_0 = const()[name = tensor<string, []>("input_623_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_623_strides_0 = const()[name = tensor<string, []>("input_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_623_pad_0 = const()[name = tensor<string, []>("input_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_623_dilations_0 = const()[name = tensor<string, []>("input_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_623_groups_0 = const()[name = tensor<string, []>("input_623_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237208960))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239306176))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = tensor<string, []>("transpose_258")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_623_cast_fp16")];
+            tensor<int32, []> x_305_split_num_splits_0 = const()[name = tensor<string, []>("x_305_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_305_split_axis_0 = const()[name = tensor<string, []>("x_305_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_305_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_305_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = tensor<string, []>("x_305_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_305_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_625_cast_fp16 = select(a = var_44_to_fp16, b = x_305_cast_fp16, cond = var_575)[name = tensor<string, []>("input_625_cast_fp16")];
+            tensor<bool, []> new_x_47_interleave_0 = const()[name = tensor<string, []>("new_x_47_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_47_cast_fp16 = concat(axis = var_59, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = tensor<string, []>("new_x_47_cast_fp16")];
+            tensor<int32, [3]> var_2931_begin_0 = const()[name = tensor<string, []>("op_2931_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_2931_end_0 = const()[name = tensor<string, []>("op_2931_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_2931_end_mask_0 = const()[name = tensor<string, []>("op_2931_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2931_cast_fp16 = slice_by_index(begin = var_2931_begin_0, end = var_2931_end_0, end_mask = var_2931_end_mask_0, x = new_x_47_cast_fp16)[name = tensor<string, []>("op_2931_cast_fp16")];
+            tensor<string, []> x_307_pad_type_0 = const()[name = tensor<string, []>("x_307_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_307_groups_0 = const()[name = tensor<string, []>("x_307_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_307_strides_0 = const()[name = tensor<string, []>("x_307_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_307_pad_0 = const()[name = tensor<string, []>("x_307_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_307_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239310336))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239319616))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_307_cast_fp16")];
+            tensor<int32, [3]> input_627_perm_0 = const()[name = tensor<string, []>("input_627_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_309_axes_0 = const()[name = tensor<string, []>("x_309_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239321728)))];
+            tensor<fp16, [1024]> encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239323840)))];
+            tensor<fp16, [1, 7, 1024]> input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = tensor<string, []>("transpose_257")];
+            tensor<fp16, [1, 7, 1024]> x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = tensor<string, []>("x_309_cast_fp16")];
+            tensor<int32, [3]> input_629_perm_0 = const()[name = tensor<string, []>("input_629_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = tensor<string, []>("transpose_256")];
+            tensor<fp16, [1, 1024, 7]> input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = tensor<string, []>("input_631_cast_fp16")];
+            tensor<string, []> x_311_pad_type_0 = const()[name = tensor<string, []>("x_311_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_311_strides_0 = const()[name = tensor<string, []>("x_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_311_pad_0 = const()[name = tensor<string, []>("x_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_311_dilations_0 = const()[name = tensor<string, []>("x_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_311_groups_0 = const()[name = tensor<string, []>("x_311_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239325952))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240374592))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_311_cast_fp16")];
+            tensor<int32, [3]> input_633_perm_0 = const()[name = tensor<string, []>("input_633_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = tensor<string, []>("transpose_255")];
+            tensor<fp16, [1, 7, 1024]> input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = tensor<string, []>("input_635_cast_fp16")];
+            tensor<int32, [1]> input_637_axes_0 = const()[name = tensor<string, []>("input_637_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240376704)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240378816)))];
+            tensor<fp16, [1, 7, 1024]> input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240380928))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243526720))), name = tensor<string, []>("encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_11_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243526912)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_107_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243535168))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246680960))), name = tensor<string, []>("encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_11_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246681152)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_108_cast_fp16")];
+            tensor<fp16, []> var_2974_to_fp16 = const()[name = tensor<string, []>("op_2974_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2975_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2974_to_fp16)[name = tensor<string, []>("op_2975_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2975_cast_fp16)[name = tensor<string, []>("input_647_cast_fp16")];
+            tensor<int32, [1]> input_649_axes_0 = const()[name = tensor<string, []>("input_649_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246683264)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246685376)))];
+            tensor<fp16, [1, 7, 1024]> input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = tensor<string, []>("input_649_cast_fp16")];
+            tensor<int32, [4]> cache_49_begin_0 = const()[name = tensor<string, []>("cache_49_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_49_end_0 = const()[name = tensor<string, []>("cache_49_end_0"), val = tensor<int32, [4]>([13, 1, 42, 1024])];
+            tensor<bool, [4]> cache_49_end_mask_0 = const()[name = tensor<string, []>("cache_49_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_49_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_49_cast_fp16")];
+            tensor<int32, [4]> cache_51_begin_0 = const()[name = tensor<string, []>("cache_51_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_51_end_0 = const()[name = tensor<string, []>("cache_51_end_0"), val = tensor<int32, [4]>([13, 1, 1024, 8])];
+            tensor<bool, [4]> cache_51_end_mask_0 = const()[name = tensor<string, []>("cache_51_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_51_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_51_cast_fp16")];
+            tensor<int32, [1]> input_651_axes_0 = const()[name = tensor<string, []>("input_651_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246687488)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246689600)))];
+            tensor<fp16, [1, 7, 1024]> input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246691712))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(249837504))), name = tensor<string, []>("encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_12_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(249837696)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_109_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(249845952))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252991744))), name = tensor<string, []>("encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_12_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252991936)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_110_cast_fp16")];
+            tensor<fp16, []> var_3011_to_fp16 = const()[name = tensor<string, []>("op_3011_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3012_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_3011_to_fp16)[name = tensor<string, []>("op_3012_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_3012_cast_fp16)[name = tensor<string, []>("input_661_cast_fp16")];
+            tensor<int32, [1]> key_25_axes_0 = const()[name = tensor<string, []>("key_25_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252994048)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252996160)))];
+            tensor<fp16, [1, 7, 1024]> key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = tensor<string, []>("key_25_cast_fp16")];
+            tensor<bool, []> input_663_interleave_0 = const()[name = tensor<string, []>("input_663_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_663_cast_fp16 = concat(axis = var_68, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = tensor<string, []>("input_663_cast_fp16")];
+            tensor<int32, [3]> var_3034_begin_0 = const()[name = tensor<string, []>("op_3034_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_3034_end_0 = const()[name = tensor<string, []>("op_3034_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3034_end_mask_0 = const()[name = tensor<string, []>("op_3034_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_3034_cast_fp16 = slice_by_index(begin = var_3034_begin_0, end = var_3034_end_0, end_mask = var_3034_end_mask_0, x = cache_49_cast_fp16)[name = tensor<string, []>("op_3034_cast_fp16")];
+            tensor<bool, []> var_3040_interleave_0 = const()[name = tensor<string, []>("op_3040_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_3040_cast_fp16 = concat(axis = var_68, interleave = var_3040_interleave_0, values = (var_3034_cast_fp16, key_25_cast_fp16))[name = tensor<string, []>("op_3040_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252998272))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(253784768))), name = tensor<string, []>("encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(253784960)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_111_cast_fp16")];
+            tensor<int32, [4]> var_3045 = const()[name = tensor<string, []>("op_3045"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_73_cast_fp16 = reshape(shape = var_3045, x = linear_111_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(253787072))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254573568))), name = tensor<string, []>("encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254573760)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_112_cast_fp16")];
+            tensor<int32, [4]> var_3050 = const()[name = tensor<string, []>("op_3050"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_49_cast_fp16 = reshape(shape = var_3050, x = linear_112_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254575872))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255362368))), name = tensor<string, []>("encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255362560)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_113_cast_fp16")];
+            tensor<int32, [4]> var_3055 = const()[name = tensor<string, []>("op_3055"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_25_cast_fp16 = reshape(shape = var_3055, x = linear_113_cast_fp16)[name = tensor<string, []>("v_25_cast_fp16")];
+            tensor<int32, [4]> value_33_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255364672)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3068_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3068_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255366784)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3070_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3070_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_25_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_319_transpose_x_0 = const()[name = tensor<string, []>("x_319_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_319_transpose_y_0 = const()[name = tensor<string, []>("x_319_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_3072_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3072_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255368896))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255468480))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255468288)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_3070_cast_fp16)[name = tensor<string, []>("transpose_254")];
+            tensor<fp16, [1, 8, 7, 97]> 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_3072_to_fp16_quantized)[name = tensor<string, []>("x_319_cast_fp16")];
+            tensor<int32, [8]> x_321_pad_0 = const()[name = tensor<string, []>("x_321_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_321_mode_0 = const()[name = tensor<string, []>("x_321_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_235_to_fp16 = const()[name = tensor<string, []>("const_235_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_321_cast_fp16")];
+            tensor<int32, [4]> var_3080 = const()[name = tensor<string, []>("op_3080"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_323_cast_fp16 = reshape(shape = var_3080, x = x_321_cast_fp16)[name = tensor<string, []>("x_323_cast_fp16")];
+            tensor<int32, [4]> var_3084_begin_0 = const()[name = tensor<string, []>("op_3084_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3084_end_0 = const()[name = tensor<string, []>("op_3084_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_3084_end_mask_0 = const()[name = tensor<string, []>("op_3084_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_3084_cast_fp16 = slice_by_index(begin = var_3084_begin_0, end = var_3084_end_0, end_mask = var_3084_end_mask_0, x = x_323_cast_fp16)[name = tensor<string, []>("op_3084_cast_fp16")];
+            tensor<int32, [4]> var_3085 = const()[name = tensor<string, []>("op_3085"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_49_cast_fp16 = reshape(shape = var_3085, x = var_3084_cast_fp16)[name = tensor<string, []>("matrix_bd_49_cast_fp16")];
+            tensor<bool, []> matrix_ac_25_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_25_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_25_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_25_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_120_perm_0 = const()[name = tensor<string, []>("transpose_120_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_121_perm_0 = const()[name = tensor<string, []>("transpose_121_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49_cast_fp16)[name = tensor<string, []>("transpose_252")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_120 = transpose(perm = transpose_120_perm_0, x = var_3068_cast_fp16)[name = tensor<string, []>("transpose_253")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_25_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_51_begin_0 = const()[name = tensor<string, []>("matrix_bd_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_51_end_0 = const()[name = tensor<string, []>("matrix_bd_51_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_51_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_51_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3094_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = tensor<string, []>("op_3094_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_49_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_49_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_49_cast_fp16 = mul(x = var_3094_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_49_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_51_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_51_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3100_cast_fp16 = softmax(axis = var_59, x = scores_51_cast_fp16)[name = tensor<string, []>("op_3100_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_665_cast_fp16 = select(a = var_44_to_fp16, b = var_3100_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_665_cast_fp16")];
+            tensor<bool, []> x_325_transpose_x_0 = const()[name = tensor<string, []>("x_325_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_325_transpose_y_0 = const()[name = tensor<string, []>("x_325_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_25_cast_fp16)[name = tensor<string, []>("transpose_251")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_325_cast_fp16")];
+            tensor<int32, [4]> var_3104_perm_0 = const()[name = tensor<string, []>("op_3104_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3105 = const()[name = tensor<string, []>("op_3105"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_3104_cast_fp16 = transpose(perm = var_3104_perm_0, x = x_325_cast_fp16)[name = tensor<string, []>("transpose_250")];
+            tensor<fp16, [1, 7, 1024]> input_667_cast_fp16 = reshape(shape = var_3105, x = var_3104_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255468800))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256255296))), name = tensor<string, []>("encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256255488)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_115_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = tensor<string, []>("input_671_cast_fp16")];
+            tensor<int32, [1]> x_329_axes_0 = const()[name = tensor<string, []>("x_329_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256257600)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256259712)))];
+            tensor<fp16, [1, 7, 1024]> x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = tensor<string, []>("x_329_cast_fp16")];
+            tensor<int32, [3]> input_673_perm_0 = const()[name = tensor<string, []>("input_673_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_675_pad_type_0 = const()[name = tensor<string, []>("input_675_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_675_strides_0 = const()[name = tensor<string, []>("input_675_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_675_pad_0 = const()[name = tensor<string, []>("input_675_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_675_dilations_0 = const()[name = tensor<string, []>("input_675_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_675_groups_0 = const()[name = tensor<string, []>("input_675_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256261824))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258359040))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = tensor<string, []>("transpose_249")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_675_cast_fp16")];
+            tensor<int32, []> x_331_split_num_splits_0 = const()[name = tensor<string, []>("x_331_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_331_split_axis_0 = const()[name = tensor<string, []>("x_331_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_331_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_331_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = tensor<string, []>("x_331_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_331_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_677_cast_fp16 = select(a = var_44_to_fp16, b = x_331_cast_fp16, cond = var_575)[name = tensor<string, []>("input_677_cast_fp16")];
+            tensor<bool, []> new_x_51_interleave_0 = const()[name = tensor<string, []>("new_x_51_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_51_cast_fp16 = concat(axis = var_59, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = tensor<string, []>("new_x_51_cast_fp16")];
+            tensor<int32, [3]> var_3144_begin_0 = const()[name = tensor<string, []>("op_3144_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_3144_end_0 = const()[name = tensor<string, []>("op_3144_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_3144_end_mask_0 = const()[name = tensor<string, []>("op_3144_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3144_cast_fp16 = slice_by_index(begin = var_3144_begin_0, end = var_3144_end_0, end_mask = var_3144_end_mask_0, x = new_x_51_cast_fp16)[name = tensor<string, []>("op_3144_cast_fp16")];
+            tensor<string, []> x_333_pad_type_0 = const()[name = tensor<string, []>("x_333_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_333_groups_0 = const()[name = tensor<string, []>("x_333_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_333_strides_0 = const()[name = tensor<string, []>("x_333_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_333_pad_0 = const()[name = tensor<string, []>("x_333_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_333_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258363200))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258372480))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_333_cast_fp16")];
+            tensor<int32, [3]> input_679_perm_0 = const()[name = tensor<string, []>("input_679_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_335_axes_0 = const()[name = tensor<string, []>("x_335_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258374592)))];
+            tensor<fp16, [1024]> encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258376704)))];
+            tensor<fp16, [1, 7, 1024]> input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = tensor<string, []>("transpose_248")];
+            tensor<fp16, [1, 7, 1024]> x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = tensor<string, []>("x_335_cast_fp16")];
+            tensor<int32, [3]> input_681_perm_0 = const()[name = tensor<string, []>("input_681_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = tensor<string, []>("transpose_247")];
+            tensor<fp16, [1, 1024, 7]> input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = tensor<string, []>("input_683_cast_fp16")];
+            tensor<string, []> x_337_pad_type_0 = const()[name = tensor<string, []>("x_337_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_337_strides_0 = const()[name = tensor<string, []>("x_337_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_337_pad_0 = const()[name = tensor<string, []>("x_337_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_337_dilations_0 = const()[name = tensor<string, []>("x_337_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_337_groups_0 = const()[name = tensor<string, []>("x_337_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258378816))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259427456))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_337_cast_fp16")];
+            tensor<int32, [3]> input_685_perm_0 = const()[name = tensor<string, []>("input_685_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = tensor<string, []>("transpose_246")];
+            tensor<fp16, [1, 7, 1024]> input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = tensor<string, []>("input_687_cast_fp16")];
+            tensor<int32, [1]> input_689_axes_0 = const()[name = tensor<string, []>("input_689_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259429568)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259431680)))];
+            tensor<fp16, [1, 7, 1024]> input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259433792))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262579584))), name = tensor<string, []>("encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_12_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262579776)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_116_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262588032))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265733824))), name = tensor<string, []>("encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_12_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265734016)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_117_cast_fp16")];
+            tensor<fp16, []> var_3187_to_fp16 = const()[name = tensor<string, []>("op_3187_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3188_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3187_to_fp16)[name = tensor<string, []>("op_3188_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_3188_cast_fp16)[name = tensor<string, []>("input_699_cast_fp16")];
+            tensor<int32, [1]> input_701_axes_0 = const()[name = tensor<string, []>("input_701_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265736128)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265738240)))];
+            tensor<fp16, [1, 7, 1024]> input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = tensor<string, []>("input_701_cast_fp16")];
+            tensor<int32, [4]> cache_53_begin_0 = const()[name = tensor<string, []>("cache_53_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_53_end_0 = const()[name = tensor<string, []>("cache_53_end_0"), val = tensor<int32, [4]>([14, 1, 42, 1024])];
+            tensor<bool, [4]> cache_53_end_mask_0 = const()[name = tensor<string, []>("cache_53_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_53_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_53_cast_fp16")];
+            tensor<int32, [4]> cache_55_begin_0 = const()[name = tensor<string, []>("cache_55_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_55_end_0 = const()[name = tensor<string, []>("cache_55_end_0"), val = tensor<int32, [4]>([14, 1, 1024, 8])];
+            tensor<bool, [4]> cache_55_end_mask_0 = const()[name = tensor<string, []>("cache_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_55_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_55_cast_fp16")];
+            tensor<int32, [1]> input_703_axes_0 = const()[name = tensor<string, []>("input_703_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265740352)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265742464)))];
+            tensor<fp16, [1, 7, 1024]> input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265744576))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268890368))), name = tensor<string, []>("encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_13_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268890560)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_118_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268898816))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272044608))), name = tensor<string, []>("encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_13_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272044800)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_119_cast_fp16")];
+            tensor<fp16, []> var_3224_to_fp16 = const()[name = tensor<string, []>("op_3224_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3225_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3224_to_fp16)[name = tensor<string, []>("op_3225_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_3225_cast_fp16)[name = tensor<string, []>("input_713_cast_fp16")];
+            tensor<int32, [1]> key_27_axes_0 = const()[name = tensor<string, []>("key_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272046912)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272049024)))];
+            tensor<fp16, [1, 7, 1024]> key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = tensor<string, []>("key_27_cast_fp16")];
+            tensor<bool, []> input_715_interleave_0 = const()[name = tensor<string, []>("input_715_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_715_cast_fp16 = concat(axis = var_68, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = tensor<string, []>("input_715_cast_fp16")];
+            tensor<int32, [3]> var_3247_begin_0 = const()[name = tensor<string, []>("op_3247_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_3247_end_0 = const()[name = tensor<string, []>("op_3247_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3247_end_mask_0 = const()[name = tensor<string, []>("op_3247_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_3247_cast_fp16 = slice_by_index(begin = var_3247_begin_0, end = var_3247_end_0, end_mask = var_3247_end_mask_0, x = cache_53_cast_fp16)[name = tensor<string, []>("op_3247_cast_fp16")];
+            tensor<bool, []> var_3253_interleave_0 = const()[name = tensor<string, []>("op_3253_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_3253_cast_fp16 = concat(axis = var_68, interleave = var_3253_interleave_0, values = (var_3247_cast_fp16, key_27_cast_fp16))[name = tensor<string, []>("op_3253_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272051136))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272837632))), name = tensor<string, []>("encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272837824)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_120_cast_fp16")];
+            tensor<int32, [4]> var_3258 = const()[name = tensor<string, []>("op_3258"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_79_cast_fp16 = reshape(shape = var_3258, x = linear_120_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272839936))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273626432))), name = tensor<string, []>("encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273626624)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_121_cast_fp16")];
+            tensor<int32, [4]> var_3263 = const()[name = tensor<string, []>("op_3263"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_53_cast_fp16 = reshape(shape = var_3263, x = linear_121_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273628736))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274415232))), name = tensor<string, []>("encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274415424)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_122_cast_fp16")];
+            tensor<int32, [4]> var_3268 = const()[name = tensor<string, []>("op_3268"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_27_cast_fp16 = reshape(shape = var_3268, x = linear_122_cast_fp16)[name = tensor<string, []>("v_27_cast_fp16")];
+            tensor<int32, [4]> value_35_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274417536)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3281_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3281_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274419648)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3283_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3283_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_27_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_345_transpose_x_0 = const()[name = tensor<string, []>("x_345_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_345_transpose_y_0 = const()[name = tensor<string, []>("x_345_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_3285_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3285_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274421760))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274521344))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274521152)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3283_cast_fp16)[name = tensor<string, []>("transpose_245")];
+            tensor<fp16, [1, 8, 7, 97]> 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_3285_to_fp16_quantized)[name = tensor<string, []>("x_345_cast_fp16")];
+            tensor<int32, [8]> x_347_pad_0 = const()[name = tensor<string, []>("x_347_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_347_mode_0 = const()[name = tensor<string, []>("x_347_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_248_to_fp16 = const()[name = tensor<string, []>("const_248_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_347_cast_fp16")];
+            tensor<int32, [4]> var_3293 = const()[name = tensor<string, []>("op_3293"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_349_cast_fp16 = reshape(shape = var_3293, x = x_347_cast_fp16)[name = tensor<string, []>("x_349_cast_fp16")];
+            tensor<int32, [4]> var_3297_begin_0 = const()[name = tensor<string, []>("op_3297_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3297_end_0 = const()[name = tensor<string, []>("op_3297_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_3297_end_mask_0 = const()[name = tensor<string, []>("op_3297_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_3297_cast_fp16 = slice_by_index(begin = var_3297_begin_0, end = var_3297_end_0, end_mask = var_3297_end_mask_0, x = x_349_cast_fp16)[name = tensor<string, []>("op_3297_cast_fp16")];
+            tensor<int32, [4]> var_3298 = const()[name = tensor<string, []>("op_3298"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_53_cast_fp16 = reshape(shape = var_3298, x = var_3297_cast_fp16)[name = tensor<string, []>("matrix_bd_53_cast_fp16")];
+            tensor<bool, []> matrix_ac_27_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_27_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_27_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_27_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_122_perm_0 = const()[name = tensor<string, []>("transpose_122_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_123_perm_0 = const()[name = tensor<string, []>("transpose_123_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53_cast_fp16)[name = tensor<string, []>("transpose_243")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_122 = transpose(perm = transpose_122_perm_0, x = var_3281_cast_fp16)[name = tensor<string, []>("transpose_244")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_27_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_55_begin_0 = const()[name = tensor<string, []>("matrix_bd_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_55_end_0 = const()[name = tensor<string, []>("matrix_bd_55_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_55_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_55_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3307_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = tensor<string, []>("op_3307_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_53_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_53_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_53_cast_fp16 = mul(x = var_3307_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_53_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_55_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_55_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3313_cast_fp16 = softmax(axis = var_59, x = scores_55_cast_fp16)[name = tensor<string, []>("op_3313_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_717_cast_fp16 = select(a = var_44_to_fp16, b = var_3313_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_717_cast_fp16")];
+            tensor<bool, []> x_351_transpose_x_0 = const()[name = tensor<string, []>("x_351_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_351_transpose_y_0 = const()[name = tensor<string, []>("x_351_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = v_27_cast_fp16)[name = tensor<string, []>("transpose_242")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_351_cast_fp16")];
+            tensor<int32, [4]> var_3317_perm_0 = const()[name = tensor<string, []>("op_3317_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3318 = const()[name = tensor<string, []>("op_3318"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_3317_cast_fp16 = transpose(perm = var_3317_perm_0, x = x_351_cast_fp16)[name = tensor<string, []>("transpose_241")];
+            tensor<fp16, [1, 7, 1024]> input_719_cast_fp16 = reshape(shape = var_3318, x = var_3317_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274521664))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275308160))), name = tensor<string, []>("encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275308352)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_124_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = tensor<string, []>("input_723_cast_fp16")];
+            tensor<int32, [1]> x_355_axes_0 = const()[name = tensor<string, []>("x_355_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275310464)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275312576)))];
+            tensor<fp16, [1, 7, 1024]> x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = tensor<string, []>("x_355_cast_fp16")];
+            tensor<int32, [3]> input_725_perm_0 = const()[name = tensor<string, []>("input_725_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_727_pad_type_0 = const()[name = tensor<string, []>("input_727_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_727_strides_0 = const()[name = tensor<string, []>("input_727_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_727_pad_0 = const()[name = tensor<string, []>("input_727_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_727_dilations_0 = const()[name = tensor<string, []>("input_727_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_727_groups_0 = const()[name = tensor<string, []>("input_727_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275314688))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277411904))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = tensor<string, []>("transpose_240")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_727_cast_fp16")];
+            tensor<int32, []> x_357_split_num_splits_0 = const()[name = tensor<string, []>("x_357_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_357_split_axis_0 = const()[name = tensor<string, []>("x_357_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_357_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_357_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = tensor<string, []>("x_357_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_357_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_729_cast_fp16 = select(a = var_44_to_fp16, b = x_357_cast_fp16, cond = var_575)[name = tensor<string, []>("input_729_cast_fp16")];
+            tensor<bool, []> new_x_55_interleave_0 = const()[name = tensor<string, []>("new_x_55_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_55_cast_fp16 = concat(axis = var_59, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = tensor<string, []>("new_x_55_cast_fp16")];
+            tensor<int32, [3]> var_3357_begin_0 = const()[name = tensor<string, []>("op_3357_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_3357_end_0 = const()[name = tensor<string, []>("op_3357_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_3357_end_mask_0 = const()[name = tensor<string, []>("op_3357_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3357_cast_fp16 = slice_by_index(begin = var_3357_begin_0, end = var_3357_end_0, end_mask = var_3357_end_mask_0, x = new_x_55_cast_fp16)[name = tensor<string, []>("op_3357_cast_fp16")];
+            tensor<string, []> x_359_pad_type_0 = const()[name = tensor<string, []>("x_359_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_359_groups_0 = const()[name = tensor<string, []>("x_359_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_359_strides_0 = const()[name = tensor<string, []>("x_359_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_359_pad_0 = const()[name = tensor<string, []>("x_359_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_359_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277416064))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277425344))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_359_cast_fp16")];
+            tensor<int32, [3]> input_731_perm_0 = const()[name = tensor<string, []>("input_731_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_361_axes_0 = const()[name = tensor<string, []>("x_361_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277427456)))];
+            tensor<fp16, [1024]> encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277429568)))];
+            tensor<fp16, [1, 7, 1024]> input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = tensor<string, []>("transpose_239")];
+            tensor<fp16, [1, 7, 1024]> x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = tensor<string, []>("x_361_cast_fp16")];
+            tensor<int32, [3]> input_733_perm_0 = const()[name = tensor<string, []>("input_733_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = tensor<string, []>("transpose_238")];
+            tensor<fp16, [1, 1024, 7]> input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = tensor<string, []>("input_735_cast_fp16")];
+            tensor<string, []> x_363_pad_type_0 = const()[name = tensor<string, []>("x_363_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_363_strides_0 = const()[name = tensor<string, []>("x_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_363_pad_0 = const()[name = tensor<string, []>("x_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_363_dilations_0 = const()[name = tensor<string, []>("x_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_363_groups_0 = const()[name = tensor<string, []>("x_363_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277431680))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278480320))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_363_cast_fp16")];
+            tensor<int32, [3]> input_737_perm_0 = const()[name = tensor<string, []>("input_737_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = tensor<string, []>("transpose_237")];
+            tensor<fp16, [1, 7, 1024]> input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = tensor<string, []>("input_739_cast_fp16")];
+            tensor<int32, [1]> input_741_axes_0 = const()[name = tensor<string, []>("input_741_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278482432)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278484544)))];
+            tensor<fp16, [1, 7, 1024]> input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278486656))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281632448))), name = tensor<string, []>("encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_13_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281632640)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_125_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281640896))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284786688))), name = tensor<string, []>("encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_13_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284786880)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_126_cast_fp16")];
+            tensor<fp16, []> var_3400_to_fp16 = const()[name = tensor<string, []>("op_3400_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3401_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3400_to_fp16)[name = tensor<string, []>("op_3401_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3401_cast_fp16)[name = tensor<string, []>("input_751_cast_fp16")];
+            tensor<int32, [1]> input_753_axes_0 = const()[name = tensor<string, []>("input_753_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284788992)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284791104)))];
+            tensor<fp16, [1, 7, 1024]> input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = tensor<string, []>("input_753_cast_fp16")];
+            tensor<int32, [4]> cache_57_begin_0 = const()[name = tensor<string, []>("cache_57_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_57_end_0 = const()[name = tensor<string, []>("cache_57_end_0"), val = tensor<int32, [4]>([15, 1, 42, 1024])];
+            tensor<bool, [4]> cache_57_end_mask_0 = const()[name = tensor<string, []>("cache_57_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_57_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_57_cast_fp16")];
+            tensor<int32, [4]> cache_59_begin_0 = const()[name = tensor<string, []>("cache_59_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_59_end_0 = const()[name = tensor<string, []>("cache_59_end_0"), val = tensor<int32, [4]>([15, 1, 1024, 8])];
+            tensor<bool, [4]> cache_59_end_mask_0 = const()[name = tensor<string, []>("cache_59_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_59_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_59_cast_fp16")];
+            tensor<int32, [1]> input_755_axes_0 = const()[name = tensor<string, []>("input_755_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284793216)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284795328)))];
+            tensor<fp16, [1, 7, 1024]> input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284797440))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287943232))), name = tensor<string, []>("encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_14_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287943424)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_127_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287951680))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291097472))), name = tensor<string, []>("encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_14_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291097664)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_128_cast_fp16")];
+            tensor<fp16, []> var_3437_to_fp16 = const()[name = tensor<string, []>("op_3437_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3438_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3437_to_fp16)[name = tensor<string, []>("op_3438_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3438_cast_fp16)[name = tensor<string, []>("input_765_cast_fp16")];
+            tensor<int32, [1]> key_29_axes_0 = const()[name = tensor<string, []>("key_29_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291099776)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291101888)))];
+            tensor<fp16, [1, 7, 1024]> key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = tensor<string, []>("key_29_cast_fp16")];
+            tensor<bool, []> input_767_interleave_0 = const()[name = tensor<string, []>("input_767_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_767_cast_fp16 = concat(axis = var_68, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = tensor<string, []>("input_767_cast_fp16")];
+            tensor<int32, [3]> var_3460_begin_0 = const()[name = tensor<string, []>("op_3460_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_3460_end_0 = const()[name = tensor<string, []>("op_3460_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3460_end_mask_0 = const()[name = tensor<string, []>("op_3460_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_3460_cast_fp16 = slice_by_index(begin = var_3460_begin_0, end = var_3460_end_0, end_mask = var_3460_end_mask_0, x = cache_57_cast_fp16)[name = tensor<string, []>("op_3460_cast_fp16")];
+            tensor<bool, []> var_3466_interleave_0 = const()[name = tensor<string, []>("op_3466_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_3466_cast_fp16 = concat(axis = var_68, interleave = var_3466_interleave_0, values = (var_3460_cast_fp16, key_29_cast_fp16))[name = tensor<string, []>("op_3466_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291104000))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291890496))), name = tensor<string, []>("encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291890688)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_129_cast_fp16")];
+            tensor<int32, [4]> var_3471 = const()[name = tensor<string, []>("op_3471"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_85_cast_fp16 = reshape(shape = var_3471, x = linear_129_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291892800))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292679296))), name = tensor<string, []>("encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292679488)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_130_cast_fp16")];
+            tensor<int32, [4]> var_3476 = const()[name = tensor<string, []>("op_3476"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_57_cast_fp16 = reshape(shape = var_3476, x = linear_130_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292681600))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293468096))), name = tensor<string, []>("encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293468288)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_131_cast_fp16")];
+            tensor<int32, [4]> var_3481 = const()[name = tensor<string, []>("op_3481"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_29_cast_fp16 = reshape(shape = var_3481, x = linear_131_cast_fp16)[name = tensor<string, []>("v_29_cast_fp16")];
+            tensor<int32, [4]> value_37_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293470400)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3494_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3494_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293472512)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3496_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3496_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_29_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_371_transpose_x_0 = const()[name = tensor<string, []>("x_371_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_371_transpose_y_0 = const()[name = tensor<string, []>("x_371_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_3498_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3498_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293474624))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293574208))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293574016)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3496_cast_fp16)[name = tensor<string, []>("transpose_236")];
+            tensor<fp16, [1, 8, 7, 97]> 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_3498_to_fp16_quantized)[name = tensor<string, []>("x_371_cast_fp16")];
+            tensor<int32, [8]> x_373_pad_0 = const()[name = tensor<string, []>("x_373_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_373_mode_0 = const()[name = tensor<string, []>("x_373_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_261_to_fp16 = const()[name = tensor<string, []>("const_261_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_373_cast_fp16")];
+            tensor<int32, [4]> var_3506 = const()[name = tensor<string, []>("op_3506"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_375_cast_fp16 = reshape(shape = var_3506, x = x_373_cast_fp16)[name = tensor<string, []>("x_375_cast_fp16")];
+            tensor<int32, [4]> var_3510_begin_0 = const()[name = tensor<string, []>("op_3510_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3510_end_0 = const()[name = tensor<string, []>("op_3510_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_3510_end_mask_0 = const()[name = tensor<string, []>("op_3510_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_3510_cast_fp16 = slice_by_index(begin = var_3510_begin_0, end = var_3510_end_0, end_mask = var_3510_end_mask_0, x = x_375_cast_fp16)[name = tensor<string, []>("op_3510_cast_fp16")];
+            tensor<int32, [4]> var_3511 = const()[name = tensor<string, []>("op_3511"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_57_cast_fp16 = reshape(shape = var_3511, x = var_3510_cast_fp16)[name = tensor<string, []>("matrix_bd_57_cast_fp16")];
+            tensor<bool, []> matrix_ac_29_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_29_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_29_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_29_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_124_perm_0 = const()[name = tensor<string, []>("transpose_124_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_125_perm_0 = const()[name = tensor<string, []>("transpose_125_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57_cast_fp16)[name = tensor<string, []>("transpose_234")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_124 = transpose(perm = transpose_124_perm_0, x = var_3494_cast_fp16)[name = tensor<string, []>("transpose_235")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_29_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_59_begin_0 = const()[name = tensor<string, []>("matrix_bd_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_59_end_0 = const()[name = tensor<string, []>("matrix_bd_59_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_59_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_59_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3520_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = tensor<string, []>("op_3520_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_57_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_57_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_57_cast_fp16 = mul(x = var_3520_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_57_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_59_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_59_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3526_cast_fp16 = softmax(axis = var_59, x = scores_59_cast_fp16)[name = tensor<string, []>("op_3526_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_769_cast_fp16 = select(a = var_44_to_fp16, b = var_3526_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_769_cast_fp16")];
+            tensor<bool, []> x_377_transpose_x_0 = const()[name = tensor<string, []>("x_377_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_377_transpose_y_0 = const()[name = tensor<string, []>("x_377_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = v_29_cast_fp16)[name = tensor<string, []>("transpose_233")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_377_cast_fp16")];
+            tensor<int32, [4]> var_3530_perm_0 = const()[name = tensor<string, []>("op_3530_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3531 = const()[name = tensor<string, []>("op_3531"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_3530_cast_fp16 = transpose(perm = var_3530_perm_0, x = x_377_cast_fp16)[name = tensor<string, []>("transpose_232")];
+            tensor<fp16, [1, 7, 1024]> input_771_cast_fp16 = reshape(shape = var_3531, x = var_3530_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293574528))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294361024))), name = tensor<string, []>("encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294361216)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_133_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = tensor<string, []>("input_775_cast_fp16")];
+            tensor<int32, [1]> x_381_axes_0 = const()[name = tensor<string, []>("x_381_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294363328)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294365440)))];
+            tensor<fp16, [1, 7, 1024]> x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = tensor<string, []>("x_381_cast_fp16")];
+            tensor<int32, [3]> input_777_perm_0 = const()[name = tensor<string, []>("input_777_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_779_pad_type_0 = const()[name = tensor<string, []>("input_779_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_779_strides_0 = const()[name = tensor<string, []>("input_779_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_779_pad_0 = const()[name = tensor<string, []>("input_779_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_779_dilations_0 = const()[name = tensor<string, []>("input_779_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_779_groups_0 = const()[name = tensor<string, []>("input_779_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294367552))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(296464768))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = tensor<string, []>("transpose_231")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_779_cast_fp16")];
+            tensor<int32, []> x_383_split_num_splits_0 = const()[name = tensor<string, []>("x_383_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_383_split_axis_0 = const()[name = tensor<string, []>("x_383_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_383_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_383_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = tensor<string, []>("x_383_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_383_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_781_cast_fp16 = select(a = var_44_to_fp16, b = x_383_cast_fp16, cond = var_575)[name = tensor<string, []>("input_781_cast_fp16")];
+            tensor<bool, []> new_x_59_interleave_0 = const()[name = tensor<string, []>("new_x_59_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_59_cast_fp16 = concat(axis = var_59, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = tensor<string, []>("new_x_59_cast_fp16")];
+            tensor<int32, [3]> var_3570_begin_0 = const()[name = tensor<string, []>("op_3570_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_3570_end_0 = const()[name = tensor<string, []>("op_3570_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_3570_end_mask_0 = const()[name = tensor<string, []>("op_3570_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3570_cast_fp16 = slice_by_index(begin = var_3570_begin_0, end = var_3570_end_0, end_mask = var_3570_end_mask_0, x = new_x_59_cast_fp16)[name = tensor<string, []>("op_3570_cast_fp16")];
+            tensor<string, []> x_385_pad_type_0 = const()[name = tensor<string, []>("x_385_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_385_groups_0 = const()[name = tensor<string, []>("x_385_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_385_strides_0 = const()[name = tensor<string, []>("x_385_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_385_pad_0 = const()[name = tensor<string, []>("x_385_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_385_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(296468928))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(296478208))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_385_cast_fp16")];
+            tensor<int32, [3]> input_783_perm_0 = const()[name = tensor<string, []>("input_783_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_387_axes_0 = const()[name = tensor<string, []>("x_387_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(296480320)))];
+            tensor<fp16, [1024]> encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(296482432)))];
+            tensor<fp16, [1, 7, 1024]> input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = tensor<string, []>("transpose_230")];
+            tensor<fp16, [1, 7, 1024]> x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = tensor<string, []>("x_387_cast_fp16")];
+            tensor<int32, [3]> input_785_perm_0 = const()[name = tensor<string, []>("input_785_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = tensor<string, []>("transpose_229")];
+            tensor<fp16, [1, 1024, 7]> input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = tensor<string, []>("input_787_cast_fp16")];
+            tensor<string, []> x_389_pad_type_0 = const()[name = tensor<string, []>("x_389_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_389_strides_0 = const()[name = tensor<string, []>("x_389_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_389_pad_0 = const()[name = tensor<string, []>("x_389_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_389_dilations_0 = const()[name = tensor<string, []>("x_389_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_389_groups_0 = const()[name = tensor<string, []>("x_389_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(296484544))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297533184))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_389_cast_fp16")];
+            tensor<int32, [3]> input_789_perm_0 = const()[name = tensor<string, []>("input_789_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = tensor<string, []>("transpose_228")];
+            tensor<fp16, [1, 7, 1024]> input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = tensor<string, []>("input_791_cast_fp16")];
+            tensor<int32, [1]> input_793_axes_0 = const()[name = tensor<string, []>("input_793_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297535296)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297537408)))];
+            tensor<fp16, [1, 7, 1024]> input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297539520))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(300685312))), name = tensor<string, []>("encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_14_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(300685504)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_134_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(300693760))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303839552))), name = tensor<string, []>("encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_14_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303839744)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_135_cast_fp16")];
+            tensor<fp16, []> var_3613_to_fp16 = const()[name = tensor<string, []>("op_3613_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3614_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3613_to_fp16)[name = tensor<string, []>("op_3614_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3614_cast_fp16)[name = tensor<string, []>("input_803_cast_fp16")];
+            tensor<int32, [1]> input_805_axes_0 = const()[name = tensor<string, []>("input_805_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303841856)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303843968)))];
+            tensor<fp16, [1, 7, 1024]> input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = tensor<string, []>("input_805_cast_fp16")];
+            tensor<int32, [4]> cache_61_begin_0 = const()[name = tensor<string, []>("cache_61_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_61_end_0 = const()[name = tensor<string, []>("cache_61_end_0"), val = tensor<int32, [4]>([16, 1, 42, 1024])];
+            tensor<bool, [4]> cache_61_end_mask_0 = const()[name = tensor<string, []>("cache_61_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_61_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_61_cast_fp16")];
+            tensor<int32, [4]> cache_63_begin_0 = const()[name = tensor<string, []>("cache_63_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_63_end_0 = const()[name = tensor<string, []>("cache_63_end_0"), val = tensor<int32, [4]>([16, 1, 1024, 8])];
+            tensor<bool, [4]> cache_63_end_mask_0 = const()[name = tensor<string, []>("cache_63_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_63_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_63_cast_fp16")];
+            tensor<int32, [1]> input_807_axes_0 = const()[name = tensor<string, []>("input_807_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303846080)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303848192)))];
+            tensor<fp16, [1, 7, 1024]> input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303850304))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306996096))), name = tensor<string, []>("encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_15_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306996288)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_136_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307004544))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310150336))), name = tensor<string, []>("encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_15_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310150528)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_137_cast_fp16")];
+            tensor<fp16, []> var_3650_to_fp16 = const()[name = tensor<string, []>("op_3650_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3651_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3650_to_fp16)[name = tensor<string, []>("op_3651_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3651_cast_fp16)[name = tensor<string, []>("input_817_cast_fp16")];
+            tensor<int32, [1]> key_31_axes_0 = const()[name = tensor<string, []>("key_31_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310152640)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310154752)))];
+            tensor<fp16, [1, 7, 1024]> key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = tensor<string, []>("key_31_cast_fp16")];
+            tensor<bool, []> input_819_interleave_0 = const()[name = tensor<string, []>("input_819_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_819_cast_fp16 = concat(axis = var_68, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = tensor<string, []>("input_819_cast_fp16")];
+            tensor<int32, [3]> var_3673_begin_0 = const()[name = tensor<string, []>("op_3673_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_3673_end_0 = const()[name = tensor<string, []>("op_3673_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3673_end_mask_0 = const()[name = tensor<string, []>("op_3673_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_3673_cast_fp16 = slice_by_index(begin = var_3673_begin_0, end = var_3673_end_0, end_mask = var_3673_end_mask_0, x = cache_61_cast_fp16)[name = tensor<string, []>("op_3673_cast_fp16")];
+            tensor<bool, []> var_3679_interleave_0 = const()[name = tensor<string, []>("op_3679_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_3679_cast_fp16 = concat(axis = var_68, interleave = var_3679_interleave_0, values = (var_3673_cast_fp16, key_31_cast_fp16))[name = tensor<string, []>("op_3679_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310156864))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310943360))), name = tensor<string, []>("encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310943552)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_138_cast_fp16")];
+            tensor<int32, [4]> var_3684 = const()[name = tensor<string, []>("op_3684"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_91_cast_fp16 = reshape(shape = var_3684, x = linear_138_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310945664))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311732160))), name = tensor<string, []>("encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311732352)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_139_cast_fp16")];
+            tensor<int32, [4]> var_3689 = const()[name = tensor<string, []>("op_3689"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_61_cast_fp16 = reshape(shape = var_3689, x = linear_139_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311734464))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312520960))), name = tensor<string, []>("encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312521152)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_140_cast_fp16")];
+            tensor<int32, [4]> var_3694 = const()[name = tensor<string, []>("op_3694"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_31_cast_fp16 = reshape(shape = var_3694, x = linear_140_cast_fp16)[name = tensor<string, []>("v_31_cast_fp16")];
+            tensor<int32, [4]> value_39_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312523264)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3707_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3707_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312525376)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3709_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3709_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_31_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_397_transpose_x_0 = const()[name = tensor<string, []>("x_397_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_397_transpose_y_0 = const()[name = tensor<string, []>("x_397_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_3711_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3711_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312527488))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312627072))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312626880)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3709_cast_fp16)[name = tensor<string, []>("transpose_227")];
+            tensor<fp16, [1, 8, 7, 97]> 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_3711_to_fp16_quantized)[name = tensor<string, []>("x_397_cast_fp16")];
+            tensor<int32, [8]> x_399_pad_0 = const()[name = tensor<string, []>("x_399_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_399_mode_0 = const()[name = tensor<string, []>("x_399_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_274_to_fp16 = const()[name = tensor<string, []>("const_274_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_399_cast_fp16")];
+            tensor<int32, [4]> var_3719 = const()[name = tensor<string, []>("op_3719"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_401_cast_fp16 = reshape(shape = var_3719, x = x_399_cast_fp16)[name = tensor<string, []>("x_401_cast_fp16")];
+            tensor<int32, [4]> var_3723_begin_0 = const()[name = tensor<string, []>("op_3723_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3723_end_0 = const()[name = tensor<string, []>("op_3723_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_3723_end_mask_0 = const()[name = tensor<string, []>("op_3723_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_3723_cast_fp16 = slice_by_index(begin = var_3723_begin_0, end = var_3723_end_0, end_mask = var_3723_end_mask_0, x = x_401_cast_fp16)[name = tensor<string, []>("op_3723_cast_fp16")];
+            tensor<int32, [4]> var_3724 = const()[name = tensor<string, []>("op_3724"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_61_cast_fp16 = reshape(shape = var_3724, x = var_3723_cast_fp16)[name = tensor<string, []>("matrix_bd_61_cast_fp16")];
+            tensor<bool, []> matrix_ac_31_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_31_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_31_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_31_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_126_perm_0 = const()[name = tensor<string, []>("transpose_126_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_127_perm_0 = const()[name = tensor<string, []>("transpose_127_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61_cast_fp16)[name = tensor<string, []>("transpose_225")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_126 = transpose(perm = transpose_126_perm_0, x = var_3707_cast_fp16)[name = tensor<string, []>("transpose_226")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_31_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_63_begin_0 = const()[name = tensor<string, []>("matrix_bd_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_63_end_0 = const()[name = tensor<string, []>("matrix_bd_63_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_63_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_63_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3733_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = tensor<string, []>("op_3733_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_61_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_61_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_61_cast_fp16 = mul(x = var_3733_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_61_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_63_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_63_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3739_cast_fp16 = softmax(axis = var_59, x = scores_63_cast_fp16)[name = tensor<string, []>("op_3739_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_821_cast_fp16 = select(a = var_44_to_fp16, b = var_3739_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_821_cast_fp16")];
+            tensor<bool, []> x_403_transpose_x_0 = const()[name = tensor<string, []>("x_403_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_403_transpose_y_0 = const()[name = tensor<string, []>("x_403_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = v_31_cast_fp16)[name = tensor<string, []>("transpose_224")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_403_cast_fp16")];
+            tensor<int32, [4]> var_3743_perm_0 = const()[name = tensor<string, []>("op_3743_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3744 = const()[name = tensor<string, []>("op_3744"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_3743_cast_fp16 = transpose(perm = var_3743_perm_0, x = x_403_cast_fp16)[name = tensor<string, []>("transpose_223")];
+            tensor<fp16, [1, 7, 1024]> input_823_cast_fp16 = reshape(shape = var_3744, x = var_3743_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312627392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313413888))), name = tensor<string, []>("encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313414080)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_142_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = tensor<string, []>("input_827_cast_fp16")];
+            tensor<int32, [1]> x_407_axes_0 = const()[name = tensor<string, []>("x_407_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313416192)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313418304)))];
+            tensor<fp16, [1, 7, 1024]> x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = tensor<string, []>("x_407_cast_fp16")];
+            tensor<int32, [3]> input_829_perm_0 = const()[name = tensor<string, []>("input_829_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_831_pad_type_0 = const()[name = tensor<string, []>("input_831_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_831_strides_0 = const()[name = tensor<string, []>("input_831_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_831_pad_0 = const()[name = tensor<string, []>("input_831_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_831_dilations_0 = const()[name = tensor<string, []>("input_831_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_831_groups_0 = const()[name = tensor<string, []>("input_831_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313420416))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315517632))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = tensor<string, []>("transpose_222")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_831_cast_fp16")];
+            tensor<int32, []> x_409_split_num_splits_0 = const()[name = tensor<string, []>("x_409_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_409_split_axis_0 = const()[name = tensor<string, []>("x_409_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_409_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_409_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = tensor<string, []>("x_409_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_409_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_833_cast_fp16 = select(a = var_44_to_fp16, b = x_409_cast_fp16, cond = var_575)[name = tensor<string, []>("input_833_cast_fp16")];
+            tensor<bool, []> new_x_63_interleave_0 = const()[name = tensor<string, []>("new_x_63_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_63_cast_fp16 = concat(axis = var_59, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = tensor<string, []>("new_x_63_cast_fp16")];
+            tensor<int32, [3]> var_3783_begin_0 = const()[name = tensor<string, []>("op_3783_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_3783_end_0 = const()[name = tensor<string, []>("op_3783_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_3783_end_mask_0 = const()[name = tensor<string, []>("op_3783_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3783_cast_fp16 = slice_by_index(begin = var_3783_begin_0, end = var_3783_end_0, end_mask = var_3783_end_mask_0, x = new_x_63_cast_fp16)[name = tensor<string, []>("op_3783_cast_fp16")];
+            tensor<string, []> x_411_pad_type_0 = const()[name = tensor<string, []>("x_411_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_411_groups_0 = const()[name = tensor<string, []>("x_411_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_411_strides_0 = const()[name = tensor<string, []>("x_411_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_411_pad_0 = const()[name = tensor<string, []>("x_411_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_411_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315521792))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315531072))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_411_cast_fp16")];
+            tensor<int32, [3]> input_835_perm_0 = const()[name = tensor<string, []>("input_835_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_413_axes_0 = const()[name = tensor<string, []>("x_413_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315533184)))];
+            tensor<fp16, [1024]> encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315535296)))];
+            tensor<fp16, [1, 7, 1024]> input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = tensor<string, []>("transpose_221")];
+            tensor<fp16, [1, 7, 1024]> x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = tensor<string, []>("x_413_cast_fp16")];
+            tensor<int32, [3]> input_837_perm_0 = const()[name = tensor<string, []>("input_837_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = tensor<string, []>("transpose_220")];
+            tensor<fp16, [1, 1024, 7]> input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = tensor<string, []>("input_839_cast_fp16")];
+            tensor<string, []> x_415_pad_type_0 = const()[name = tensor<string, []>("x_415_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_415_strides_0 = const()[name = tensor<string, []>("x_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_415_pad_0 = const()[name = tensor<string, []>("x_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_415_dilations_0 = const()[name = tensor<string, []>("x_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_415_groups_0 = const()[name = tensor<string, []>("x_415_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315537408))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316586048))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_415_cast_fp16")];
+            tensor<int32, [3]> input_841_perm_0 = const()[name = tensor<string, []>("input_841_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = tensor<string, []>("transpose_219")];
+            tensor<fp16, [1, 7, 1024]> input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = tensor<string, []>("input_843_cast_fp16")];
+            tensor<int32, [1]> input_845_axes_0 = const()[name = tensor<string, []>("input_845_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316588160)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316590272)))];
+            tensor<fp16, [1, 7, 1024]> input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316592384))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319738176))), name = tensor<string, []>("encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_15_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319738368)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_143_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319746624))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322892416))), name = tensor<string, []>("encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_15_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322892608)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_144_cast_fp16")];
+            tensor<fp16, []> var_3826_to_fp16 = const()[name = tensor<string, []>("op_3826_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3827_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3826_to_fp16)[name = tensor<string, []>("op_3827_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3827_cast_fp16)[name = tensor<string, []>("input_855_cast_fp16")];
+            tensor<int32, [1]> input_857_axes_0 = const()[name = tensor<string, []>("input_857_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322894720)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322896832)))];
+            tensor<fp16, [1, 7, 1024]> input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = tensor<string, []>("input_857_cast_fp16")];
+            tensor<int32, [4]> cache_65_begin_0 = const()[name = tensor<string, []>("cache_65_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_65_end_0 = const()[name = tensor<string, []>("cache_65_end_0"), val = tensor<int32, [4]>([17, 1, 42, 1024])];
+            tensor<bool, [4]> cache_65_end_mask_0 = const()[name = tensor<string, []>("cache_65_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_65_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_65_cast_fp16")];
+            tensor<int32, [4]> cache_67_begin_0 = const()[name = tensor<string, []>("cache_67_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_67_end_0 = const()[name = tensor<string, []>("cache_67_end_0"), val = tensor<int32, [4]>([17, 1, 1024, 8])];
+            tensor<bool, [4]> cache_67_end_mask_0 = const()[name = tensor<string, []>("cache_67_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_67_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_67_cast_fp16")];
+            tensor<int32, [1]> input_859_axes_0 = const()[name = tensor<string, []>("input_859_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322898944)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322901056)))];
+            tensor<fp16, [1, 7, 1024]> input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322903168))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(326048960))), name = tensor<string, []>("encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_16_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(326049152)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_145_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(326057408))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329203200))), name = tensor<string, []>("encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_16_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329203392)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_146_cast_fp16")];
+            tensor<fp16, []> var_3863_to_fp16 = const()[name = tensor<string, []>("op_3863_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3864_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3863_to_fp16)[name = tensor<string, []>("op_3864_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3864_cast_fp16)[name = tensor<string, []>("input_869_cast_fp16")];
+            tensor<int32, [1]> key_33_axes_0 = const()[name = tensor<string, []>("key_33_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329205504)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329207616)))];
+            tensor<fp16, [1, 7, 1024]> key_33_cast_fp16 = layer_norm(axes = key_33_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = tensor<string, []>("key_33_cast_fp16")];
+            tensor<bool, []> input_871_interleave_0 = const()[name = tensor<string, []>("input_871_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_871_cast_fp16 = concat(axis = var_68, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_33_cast_fp16))[name = tensor<string, []>("input_871_cast_fp16")];
+            tensor<int32, [3]> var_3886_begin_0 = const()[name = tensor<string, []>("op_3886_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_3886_end_0 = const()[name = tensor<string, []>("op_3886_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3886_end_mask_0 = const()[name = tensor<string, []>("op_3886_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_3886_cast_fp16 = slice_by_index(begin = var_3886_begin_0, end = var_3886_end_0, end_mask = var_3886_end_mask_0, x = cache_65_cast_fp16)[name = tensor<string, []>("op_3886_cast_fp16")];
+            tensor<bool, []> var_3892_interleave_0 = const()[name = tensor<string, []>("op_3892_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_3892_cast_fp16 = concat(axis = var_68, interleave = var_3892_interleave_0, values = (var_3886_cast_fp16, key_33_cast_fp16))[name = tensor<string, []>("op_3892_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329209728))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329996224))), name = tensor<string, []>("encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329996416)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_147_cast_fp16")];
+            tensor<int32, [4]> var_3897 = const()[name = tensor<string, []>("op_3897"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_97_cast_fp16 = reshape(shape = var_3897, x = linear_147_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329998528))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330785024))), name = tensor<string, []>("encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330785216)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_148_cast_fp16")];
+            tensor<int32, [4]> var_3902 = const()[name = tensor<string, []>("op_3902"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_65_cast_fp16 = reshape(shape = var_3902, x = linear_148_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330787328))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331573824))), name = tensor<string, []>("encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331574016)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_149_cast_fp16")];
+            tensor<int32, [4]> var_3907 = const()[name = tensor<string, []>("op_3907"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_33_cast_fp16 = reshape(shape = var_3907, x = linear_149_cast_fp16)[name = tensor<string, []>("v_33_cast_fp16")];
+            tensor<int32, [4]> value_41_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331576128)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3920_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3920_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331578240)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3922_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3922_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_33_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_423_transpose_x_0 = const()[name = tensor<string, []>("x_423_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_423_transpose_y_0 = const()[name = tensor<string, []>("x_423_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_3924_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3924_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331580352))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331679936))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331679744)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_33_cast_fp16 = transpose(perm = q_with_bias_v_33_perm_0, x = var_3922_cast_fp16)[name = tensor<string, []>("transpose_218")];
+            tensor<fp16, [1, 8, 7, 97]> 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_3924_to_fp16_quantized)[name = tensor<string, []>("x_423_cast_fp16")];
+            tensor<int32, [8]> x_425_pad_0 = const()[name = tensor<string, []>("x_425_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_425_mode_0 = const()[name = tensor<string, []>("x_425_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_287_to_fp16 = const()[name = tensor<string, []>("const_287_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_425_cast_fp16")];
+            tensor<int32, [4]> var_3932 = const()[name = tensor<string, []>("op_3932"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_427_cast_fp16 = reshape(shape = var_3932, x = x_425_cast_fp16)[name = tensor<string, []>("x_427_cast_fp16")];
+            tensor<int32, [4]> var_3936_begin_0 = const()[name = tensor<string, []>("op_3936_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3936_end_0 = const()[name = tensor<string, []>("op_3936_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_3936_end_mask_0 = const()[name = tensor<string, []>("op_3936_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_3936_cast_fp16 = slice_by_index(begin = var_3936_begin_0, end = var_3936_end_0, end_mask = var_3936_end_mask_0, x = x_427_cast_fp16)[name = tensor<string, []>("op_3936_cast_fp16")];
+            tensor<int32, [4]> var_3937 = const()[name = tensor<string, []>("op_3937"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_65_cast_fp16 = reshape(shape = var_3937, x = var_3936_cast_fp16)[name = tensor<string, []>("matrix_bd_65_cast_fp16")];
+            tensor<bool, []> matrix_ac_33_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_33_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_33_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_33_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_128_perm_0 = const()[name = tensor<string, []>("transpose_128_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_129_perm_0 = const()[name = tensor<string, []>("transpose_129_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65_cast_fp16)[name = tensor<string, []>("transpose_216")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_128 = transpose(perm = transpose_128_perm_0, x = var_3920_cast_fp16)[name = tensor<string, []>("transpose_217")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_33_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_67_begin_0 = const()[name = tensor<string, []>("matrix_bd_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_67_end_0 = const()[name = tensor<string, []>("matrix_bd_67_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_67_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_67_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3946_cast_fp16 = add(x = matrix_ac_33_cast_fp16, y = matrix_bd_67_cast_fp16)[name = tensor<string, []>("op_3946_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_65_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_65_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_65_cast_fp16 = mul(x = var_3946_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_65_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_67_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_67_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3952_cast_fp16 = softmax(axis = var_59, x = scores_67_cast_fp16)[name = tensor<string, []>("op_3952_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_873_cast_fp16 = select(a = var_44_to_fp16, b = var_3952_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_873_cast_fp16")];
+            tensor<bool, []> x_429_transpose_x_0 = const()[name = tensor<string, []>("x_429_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_429_transpose_y_0 = const()[name = tensor<string, []>("x_429_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = v_33_cast_fp16)[name = tensor<string, []>("transpose_215")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_429_cast_fp16")];
+            tensor<int32, [4]> var_3956_perm_0 = const()[name = tensor<string, []>("op_3956_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3957 = const()[name = tensor<string, []>("op_3957"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_3956_cast_fp16 = transpose(perm = var_3956_perm_0, x = x_429_cast_fp16)[name = tensor<string, []>("transpose_214")];
+            tensor<fp16, [1, 7, 1024]> input_875_cast_fp16 = reshape(shape = var_3957, x = var_3956_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331680256))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332466752))), name = tensor<string, []>("encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332466944)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_151_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = tensor<string, []>("input_879_cast_fp16")];
+            tensor<int32, [1]> x_433_axes_0 = const()[name = tensor<string, []>("x_433_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332469056)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332471168)))];
+            tensor<fp16, [1, 7, 1024]> x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = tensor<string, []>("x_433_cast_fp16")];
+            tensor<int32, [3]> input_881_perm_0 = const()[name = tensor<string, []>("input_881_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_883_pad_type_0 = const()[name = tensor<string, []>("input_883_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_883_strides_0 = const()[name = tensor<string, []>("input_883_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_883_pad_0 = const()[name = tensor<string, []>("input_883_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_883_dilations_0 = const()[name = tensor<string, []>("input_883_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_883_groups_0 = const()[name = tensor<string, []>("input_883_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332473280))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334570496))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = tensor<string, []>("transpose_213")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_883_cast_fp16")];
+            tensor<int32, []> x_435_split_num_splits_0 = const()[name = tensor<string, []>("x_435_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_435_split_axis_0 = const()[name = tensor<string, []>("x_435_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_435_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_435_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = tensor<string, []>("x_435_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_435_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_885_cast_fp16 = select(a = var_44_to_fp16, b = x_435_cast_fp16, cond = var_575)[name = tensor<string, []>("input_885_cast_fp16")];
+            tensor<bool, []> new_x_67_interleave_0 = const()[name = tensor<string, []>("new_x_67_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_67_cast_fp16 = concat(axis = var_59, interleave = new_x_67_interleave_0, values = (cache_67_cast_fp16, input_885_cast_fp16))[name = tensor<string, []>("new_x_67_cast_fp16")];
+            tensor<int32, [3]> var_3996_begin_0 = const()[name = tensor<string, []>("op_3996_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_3996_end_0 = const()[name = tensor<string, []>("op_3996_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_3996_end_mask_0 = const()[name = tensor<string, []>("op_3996_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3996_cast_fp16 = slice_by_index(begin = var_3996_begin_0, end = var_3996_end_0, end_mask = var_3996_end_mask_0, x = new_x_67_cast_fp16)[name = tensor<string, []>("op_3996_cast_fp16")];
+            tensor<string, []> x_437_pad_type_0 = const()[name = tensor<string, []>("x_437_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_437_groups_0 = const()[name = tensor<string, []>("x_437_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_437_strides_0 = const()[name = tensor<string, []>("x_437_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_437_pad_0 = const()[name = tensor<string, []>("x_437_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_437_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334574656))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334583936))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_437_cast_fp16")];
+            tensor<int32, [3]> input_887_perm_0 = const()[name = tensor<string, []>("input_887_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_439_axes_0 = const()[name = tensor<string, []>("x_439_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334586048)))];
+            tensor<fp16, [1024]> encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334588160)))];
+            tensor<fp16, [1, 7, 1024]> input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = tensor<string, []>("transpose_212")];
+            tensor<fp16, [1, 7, 1024]> x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = tensor<string, []>("x_439_cast_fp16")];
+            tensor<int32, [3]> input_889_perm_0 = const()[name = tensor<string, []>("input_889_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = tensor<string, []>("transpose_211")];
+            tensor<fp16, [1, 1024, 7]> input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = tensor<string, []>("input_891_cast_fp16")];
+            tensor<string, []> x_441_pad_type_0 = const()[name = tensor<string, []>("x_441_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_441_strides_0 = const()[name = tensor<string, []>("x_441_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_441_pad_0 = const()[name = tensor<string, []>("x_441_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_441_dilations_0 = const()[name = tensor<string, []>("x_441_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_441_groups_0 = const()[name = tensor<string, []>("x_441_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334590272))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335638912))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_441_cast_fp16")];
+            tensor<int32, [3]> input_893_perm_0 = const()[name = tensor<string, []>("input_893_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = tensor<string, []>("transpose_210")];
+            tensor<fp16, [1, 7, 1024]> input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = tensor<string, []>("input_895_cast_fp16")];
+            tensor<int32, [1]> input_897_axes_0 = const()[name = tensor<string, []>("input_897_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335641024)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335643136)))];
+            tensor<fp16, [1, 7, 1024]> input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335645248))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338791040))), name = tensor<string, []>("encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_16_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338791232)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_152_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338799488))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(341945280))), name = tensor<string, []>("encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_16_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(341945472)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_153_cast_fp16")];
+            tensor<fp16, []> var_4039_to_fp16 = const()[name = tensor<string, []>("op_4039_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4040_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_4039_to_fp16)[name = tensor<string, []>("op_4040_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_907_cast_fp16 = add(x = input_895_cast_fp16, y = var_4040_cast_fp16)[name = tensor<string, []>("input_907_cast_fp16")];
+            tensor<int32, [1]> input_909_axes_0 = const()[name = tensor<string, []>("input_909_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(341947584)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(341949696)))];
+            tensor<fp16, [1, 7, 1024]> input_909_cast_fp16 = layer_norm(axes = input_909_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_907_cast_fp16)[name = tensor<string, []>("input_909_cast_fp16")];
+            tensor<int32, [4]> cache_69_begin_0 = const()[name = tensor<string, []>("cache_69_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> cache_69_end_0 = const()[name = tensor<string, []>("cache_69_end_0"), val = tensor<int32, [4]>([18, 1, 42, 1024])];
+            tensor<bool, [4]> cache_69_end_mask_0 = const()[name = tensor<string, []>("cache_69_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_69_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_69_cast_fp16")];
+            tensor<int32, [4]> cache_71_begin_0 = const()[name = tensor<string, []>("cache_71_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> cache_71_end_0 = const()[name = tensor<string, []>("cache_71_end_0"), val = tensor<int32, [4]>([18, 1, 1024, 8])];
+            tensor<bool, [4]> cache_71_end_mask_0 = const()[name = tensor<string, []>("cache_71_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_71_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_71_cast_fp16")];
+            tensor<int32, [1]> input_911_axes_0 = const()[name = tensor<string, []>("input_911_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(341951808)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(341953920)))];
+            tensor<fp16, [1, 7, 1024]> input_911_cast_fp16 = layer_norm(axes = input_911_axes_0, beta = encoder_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_feed_forward1_weight_to_fp16, x = input_909_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(341956032))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(345101824))), name = tensor<string, []>("encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_17_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(345102016)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_154_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_915_cast_fp16 = silu(x = linear_154_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(345110272))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348256064))), name = tensor<string, []>("encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_17_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348256256)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_155_cast_fp16")];
+            tensor<fp16, []> var_4076_to_fp16 = const()[name = tensor<string, []>("op_4076_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4077_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_4076_to_fp16)[name = tensor<string, []>("op_4077_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_921_cast_fp16 = add(x = input_909_cast_fp16, y = var_4077_cast_fp16)[name = tensor<string, []>("input_921_cast_fp16")];
+            tensor<int32, [1]> key_35_axes_0 = const()[name = tensor<string, []>("key_35_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348258368)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348260480)))];
+            tensor<fp16, [1, 7, 1024]> key_35_cast_fp16 = layer_norm(axes = key_35_axes_0, beta = encoder_layers_17_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_self_att_weight_to_fp16, x = input_921_cast_fp16)[name = tensor<string, []>("key_35_cast_fp16")];
+            tensor<bool, []> input_923_interleave_0 = const()[name = tensor<string, []>("input_923_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_923_cast_fp16 = concat(axis = var_68, interleave = input_923_interleave_0, values = (cache_69_cast_fp16, key_35_cast_fp16))[name = tensor<string, []>("input_923_cast_fp16")];
+            tensor<int32, [3]> var_4099_begin_0 = const()[name = tensor<string, []>("op_4099_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_4099_end_0 = const()[name = tensor<string, []>("op_4099_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4099_end_mask_0 = const()[name = tensor<string, []>("op_4099_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_4099_cast_fp16 = slice_by_index(begin = var_4099_begin_0, end = var_4099_end_0, end_mask = var_4099_end_mask_0, x = cache_69_cast_fp16)[name = tensor<string, []>("op_4099_cast_fp16")];
+            tensor<bool, []> var_4105_interleave_0 = const()[name = tensor<string, []>("op_4105_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_4105_cast_fp16 = concat(axis = var_68, interleave = var_4105_interleave_0, values = (var_4099_cast_fp16, key_35_cast_fp16))[name = tensor<string, []>("op_4105_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348262592))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349049088))), name = tensor<string, []>("encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349049280)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_156_cast_fp16")];
+            tensor<int32, [4]> var_4110 = const()[name = tensor<string, []>("op_4110"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_103_cast_fp16 = reshape(shape = var_4110, x = linear_156_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349051392))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349837888))), name = tensor<string, []>("encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349838080)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_157_cast_fp16")];
+            tensor<int32, [4]> var_4115 = const()[name = tensor<string, []>("op_4115"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_69_cast_fp16 = reshape(shape = var_4115, x = linear_157_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349840192))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350626688))), name = tensor<string, []>("encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350626880)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_158_cast_fp16")];
+            tensor<int32, [4]> var_4120 = const()[name = tensor<string, []>("op_4120"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_35_cast_fp16 = reshape(shape = var_4120, x = linear_158_cast_fp16)[name = tensor<string, []>("v_35_cast_fp16")];
+            tensor<int32, [4]> value_43_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_17_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350628992)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4133_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_4133_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_17_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350631104)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4135_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_4135_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_35_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_449_transpose_x_0 = const()[name = tensor<string, []>("x_449_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_449_transpose_y_0 = const()[name = tensor<string, []>("x_449_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_4137_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_4137_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350633216))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350732800))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350732608)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_35_cast_fp16 = transpose(perm = q_with_bias_v_35_perm_0, x = var_4135_cast_fp16)[name = tensor<string, []>("transpose_209")];
+            tensor<fp16, [1, 8, 7, 97]> 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_4137_to_fp16_quantized)[name = tensor<string, []>("x_449_cast_fp16")];
+            tensor<int32, [8]> x_451_pad_0 = const()[name = tensor<string, []>("x_451_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_451_mode_0 = const()[name = tensor<string, []>("x_451_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_300_to_fp16 = const()[name = tensor<string, []>("const_300_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_451_cast_fp16")];
+            tensor<int32, [4]> var_4145 = const()[name = tensor<string, []>("op_4145"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_453_cast_fp16 = reshape(shape = var_4145, x = x_451_cast_fp16)[name = tensor<string, []>("x_453_cast_fp16")];
+            tensor<int32, [4]> var_4149_begin_0 = const()[name = tensor<string, []>("op_4149_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4149_end_0 = const()[name = tensor<string, []>("op_4149_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_4149_end_mask_0 = const()[name = tensor<string, []>("op_4149_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_4149_cast_fp16 = slice_by_index(begin = var_4149_begin_0, end = var_4149_end_0, end_mask = var_4149_end_mask_0, x = x_453_cast_fp16)[name = tensor<string, []>("op_4149_cast_fp16")];
+            tensor<int32, [4]> var_4150 = const()[name = tensor<string, []>("op_4150"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_69_cast_fp16 = reshape(shape = var_4150, x = var_4149_cast_fp16)[name = tensor<string, []>("matrix_bd_69_cast_fp16")];
+            tensor<bool, []> matrix_ac_35_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_35_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_35_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_35_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_130_perm_0 = const()[name = tensor<string, []>("transpose_130_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_131_perm_0 = const()[name = tensor<string, []>("transpose_131_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69_cast_fp16)[name = tensor<string, []>("transpose_207")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_130 = transpose(perm = transpose_130_perm_0, x = var_4133_cast_fp16)[name = tensor<string, []>("transpose_208")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_35_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_71_begin_0 = const()[name = tensor<string, []>("matrix_bd_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_71_end_0 = const()[name = tensor<string, []>("matrix_bd_71_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_71_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_71_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_4159_cast_fp16 = add(x = matrix_ac_35_cast_fp16, y = matrix_bd_71_cast_fp16)[name = tensor<string, []>("op_4159_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_69_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_69_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_69_cast_fp16 = mul(x = var_4159_cast_fp16, y = _inversed_scores_69_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_69_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_71_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_69_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_71_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_4165_cast_fp16 = softmax(axis = var_59, x = scores_71_cast_fp16)[name = tensor<string, []>("op_4165_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_925_cast_fp16 = select(a = var_44_to_fp16, b = var_4165_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_925_cast_fp16")];
+            tensor<bool, []> x_455_transpose_x_0 = const()[name = tensor<string, []>("x_455_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_455_transpose_y_0 = const()[name = tensor<string, []>("x_455_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = v_35_cast_fp16)[name = tensor<string, []>("transpose_206")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_455_cast_fp16")];
+            tensor<int32, [4]> var_4169_perm_0 = const()[name = tensor<string, []>("op_4169_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4170 = const()[name = tensor<string, []>("op_4170"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_4169_cast_fp16 = transpose(perm = var_4169_perm_0, x = x_455_cast_fp16)[name = tensor<string, []>("transpose_205")];
+            tensor<fp16, [1, 7, 1024]> input_927_cast_fp16 = reshape(shape = var_4170, x = var_4169_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350733120))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(351519616))), name = tensor<string, []>("encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(351519808)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_160_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_931_cast_fp16 = add(x = input_921_cast_fp16, y = linear_160_cast_fp16)[name = tensor<string, []>("input_931_cast_fp16")];
+            tensor<int32, [1]> x_459_axes_0 = const()[name = tensor<string, []>("x_459_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(351521920)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(351524032)))];
+            tensor<fp16, [1, 7, 1024]> x_459_cast_fp16 = layer_norm(axes = x_459_axes_0, beta = encoder_layers_17_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_conv_weight_to_fp16, x = input_931_cast_fp16)[name = tensor<string, []>("x_459_cast_fp16")];
+            tensor<int32, [3]> input_933_perm_0 = const()[name = tensor<string, []>("input_933_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_935_pad_type_0 = const()[name = tensor<string, []>("input_935_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_935_strides_0 = const()[name = tensor<string, []>("input_935_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_935_pad_0 = const()[name = tensor<string, []>("input_935_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_935_dilations_0 = const()[name = tensor<string, []>("input_935_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_935_groups_0 = const()[name = tensor<string, []>("input_935_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(351526144))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353623360))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_933_cast_fp16 = transpose(perm = input_933_perm_0, x = x_459_cast_fp16)[name = tensor<string, []>("transpose_204")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_935_cast_fp16")];
+            tensor<int32, []> x_461_split_num_splits_0 = const()[name = tensor<string, []>("x_461_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_461_split_axis_0 = const()[name = tensor<string, []>("x_461_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_461_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_461_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_461_split_1_sigmoid_cast_fp16 = sigmoid(x = x_461_split_cast_fp16_1)[name = tensor<string, []>("x_461_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_461_cast_fp16 = mul(x = x_461_split_cast_fp16_0, y = x_461_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_461_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_937_cast_fp16 = select(a = var_44_to_fp16, b = x_461_cast_fp16, cond = var_575)[name = tensor<string, []>("input_937_cast_fp16")];
+            tensor<bool, []> new_x_71_interleave_0 = const()[name = tensor<string, []>("new_x_71_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_71_cast_fp16 = concat(axis = var_59, interleave = new_x_71_interleave_0, values = (cache_71_cast_fp16, input_937_cast_fp16))[name = tensor<string, []>("new_x_71_cast_fp16")];
+            tensor<int32, [3]> var_4209_begin_0 = const()[name = tensor<string, []>("op_4209_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_4209_end_0 = const()[name = tensor<string, []>("op_4209_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_4209_end_mask_0 = const()[name = tensor<string, []>("op_4209_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4209_cast_fp16 = slice_by_index(begin = var_4209_begin_0, end = var_4209_end_0, end_mask = var_4209_end_mask_0, x = new_x_71_cast_fp16)[name = tensor<string, []>("op_4209_cast_fp16")];
+            tensor<string, []> x_463_pad_type_0 = const()[name = tensor<string, []>("x_463_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_463_groups_0 = const()[name = tensor<string, []>("x_463_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_463_strides_0 = const()[name = tensor<string, []>("x_463_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_463_pad_0 = const()[name = tensor<string, []>("x_463_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_463_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353627520))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353636800))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_463_cast_fp16")];
+            tensor<int32, [3]> input_939_perm_0 = const()[name = tensor<string, []>("input_939_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_465_axes_0 = const()[name = tensor<string, []>("x_465_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353638912)))];
+            tensor<fp16, [1024]> encoder_layers_17_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353641024)))];
+            tensor<fp16, [1, 7, 1024]> input_939_cast_fp16 = transpose(perm = input_939_perm_0, x = x_463_cast_fp16)[name = tensor<string, []>("transpose_203")];
+            tensor<fp16, [1, 7, 1024]> x_465_cast_fp16 = layer_norm(axes = x_465_axes_0, beta = encoder_layers_17_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_conv_batch_norm_weight_to_fp16, x = input_939_cast_fp16)[name = tensor<string, []>("x_465_cast_fp16")];
+            tensor<int32, [3]> input_941_perm_0 = const()[name = tensor<string, []>("input_941_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_941_cast_fp16 = transpose(perm = input_941_perm_0, x = x_465_cast_fp16)[name = tensor<string, []>("transpose_202")];
+            tensor<fp16, [1, 1024, 7]> input_943_cast_fp16 = silu(x = input_941_cast_fp16)[name = tensor<string, []>("input_943_cast_fp16")];
+            tensor<string, []> x_467_pad_type_0 = const()[name = tensor<string, []>("x_467_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_467_strides_0 = const()[name = tensor<string, []>("x_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_467_pad_0 = const()[name = tensor<string, []>("x_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_467_dilations_0 = const()[name = tensor<string, []>("x_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_467_groups_0 = const()[name = tensor<string, []>("x_467_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353643136))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(354691776))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_467_cast_fp16")];
+            tensor<int32, [3]> input_945_perm_0 = const()[name = tensor<string, []>("input_945_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_945_cast_fp16 = transpose(perm = input_945_perm_0, x = x_467_cast_fp16)[name = tensor<string, []>("transpose_201")];
+            tensor<fp16, [1, 7, 1024]> input_947_cast_fp16 = add(x = input_931_cast_fp16, y = input_945_cast_fp16)[name = tensor<string, []>("input_947_cast_fp16")];
+            tensor<int32, [1]> input_949_axes_0 = const()[name = tensor<string, []>("input_949_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(354693888)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(354696000)))];
+            tensor<fp16, [1, 7, 1024]> input_949_cast_fp16 = layer_norm(axes = input_949_axes_0, beta = encoder_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_feed_forward2_weight_to_fp16, x = input_947_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(354698112))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(357843904))), name = tensor<string, []>("encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_17_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(357844096)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_161_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_953_cast_fp16 = silu(x = linear_161_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(357852352))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360998144))), name = tensor<string, []>("encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_17_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360998336)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_162_cast_fp16")];
+            tensor<fp16, []> var_4252_to_fp16 = const()[name = tensor<string, []>("op_4252_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4253_cast_fp16 = mul(x = linear_162_cast_fp16, y = var_4252_to_fp16)[name = tensor<string, []>("op_4253_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_959_cast_fp16 = add(x = input_947_cast_fp16, y = var_4253_cast_fp16)[name = tensor<string, []>("input_959_cast_fp16")];
+            tensor<int32, [1]> input_961_axes_0 = const()[name = tensor<string, []>("input_961_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361000448)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_17_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361002560)))];
+            tensor<fp16, [1, 7, 1024]> input_961_cast_fp16 = layer_norm(axes = input_961_axes_0, beta = encoder_layers_17_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_out_weight_to_fp16, x = input_959_cast_fp16)[name = tensor<string, []>("input_961_cast_fp16")];
+            tensor<int32, [4]> cache_73_begin_0 = const()[name = tensor<string, []>("cache_73_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> cache_73_end_0 = const()[name = tensor<string, []>("cache_73_end_0"), val = tensor<int32, [4]>([19, 1, 42, 1024])];
+            tensor<bool, [4]> cache_73_end_mask_0 = const()[name = tensor<string, []>("cache_73_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_73_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_73_cast_fp16")];
+            tensor<int32, [4]> cache_75_begin_0 = const()[name = tensor<string, []>("cache_75_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> cache_75_end_0 = const()[name = tensor<string, []>("cache_75_end_0"), val = tensor<int32, [4]>([19, 1, 1024, 8])];
+            tensor<bool, [4]> cache_75_end_mask_0 = const()[name = tensor<string, []>("cache_75_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_75_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_75_cast_fp16")];
+            tensor<int32, [1]> input_963_axes_0 = const()[name = tensor<string, []>("input_963_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361004672)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361006784)))];
+            tensor<fp16, [1, 7, 1024]> input_963_cast_fp16 = layer_norm(axes = input_963_axes_0, beta = encoder_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_feed_forward1_weight_to_fp16, x = input_961_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361008896))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(364154688))), name = tensor<string, []>("encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([4096, 1024])];
+            tensor<fp16, [4096]> encoder_layers_18_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(364154880)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_163_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_967_cast_fp16 = silu(x = linear_163_cast_fp16)[name = tensor<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<uint8, [3145728]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(364163136))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(367308928))), name = tensor<string, []>("encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 4096])];
+            tensor<fp16, [1024]> encoder_layers_18_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(367309120)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_164_cast_fp16")];
+            tensor<fp16, []> var_4289_to_fp16 = const()[name = tensor<string, []>("op_4289_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4290_cast_fp16 = mul(x = linear_164_cast_fp16, y = var_4289_to_fp16)[name = tensor<string, []>("op_4290_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_973_cast_fp16 = add(x = input_961_cast_fp16, y = var_4290_cast_fp16)[name = tensor<string, []>("input_973_cast_fp16")];
+            tensor<int32, [1]> key_37_axes_0 = const()[name = tensor<string, []>("key_37_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(367311232)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(367313344)))];
+            tensor<fp16, [1, 7, 1024]> key_37_cast_fp16 = layer_norm(axes = key_37_axes_0, beta = encoder_layers_18_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_self_att_weight_to_fp16, x = input_973_cast_fp16)[name = tensor<string, []>("key_37_cast_fp16")];
+            tensor<bool, []> input_975_interleave_0 = const()[name = tensor<string, []>("input_975_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_975_cast_fp16 = concat(axis = var_68, interleave = input_975_interleave_0, values = (cache_73_cast_fp16, key_37_cast_fp16))[name = tensor<string, []>("input_975_cast_fp16")];
+            tensor<int32, [3]> var_4312_begin_0 = const()[name = tensor<string, []>("op_4312_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_4312_end_0 = const()[name = tensor<string, []>("op_4312_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4312_end_mask_0 = const()[name = tensor<string, []>("op_4312_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_4312_cast_fp16 = slice_by_index(begin = var_4312_begin_0, end = var_4312_end_0, end_mask = var_4312_end_mask_0, x = cache_73_cast_fp16)[name = tensor<string, []>("op_4312_cast_fp16")];
+            tensor<bool, []> var_4318_interleave_0 = const()[name = tensor<string, []>("op_4318_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_4318_cast_fp16 = concat(axis = var_68, interleave = var_4318_interleave_0, values = (var_4312_cast_fp16, key_37_cast_fp16))[name = tensor<string, []>("op_4318_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(367315456))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368101952))), name = tensor<string, []>("encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368102144)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_165_cast_fp16")];
+            tensor<int32, [4]> var_4323 = const()[name = tensor<string, []>("op_4323"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_109_cast_fp16 = reshape(shape = var_4323, x = linear_165_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368104256))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368890752))), name = tensor<string, []>("encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368890944)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_166_cast_fp16")];
+            tensor<int32, [4]> var_4328 = const()[name = tensor<string, []>("op_4328"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_73_cast_fp16 = reshape(shape = var_4328, x = linear_166_cast_fp16)[name = tensor<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<uint8, [786432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368893056))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369679552))), name = tensor<string, []>("encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([1024, 1024])];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369679744)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_167_cast_fp16")];
+            tensor<int32, [4]> var_4333 = const()[name = tensor<string, []>("op_4333"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_37_cast_fp16 = reshape(shape = var_4333, x = linear_167_cast_fp16)[name = tensor<string, []>("v_37_cast_fp16")];
+            tensor<int32, [4]> value_45_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_18_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369681856)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4346_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_4346_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_18_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369683968)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4348_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_4348_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_37_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_475_transpose_x_0 = const()[name = tensor<string, []>("x_475_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_475_transpose_y_0 = const()[name = tensor<string, []>("x_475_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_4350_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_4350_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369686080))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369785664))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369785472)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_37_cast_fp16 = transpose(perm = q_with_bias_v_37_perm_0, x = var_4348_cast_fp16)[name = tensor<string, []>("transpose_200")];
+            tensor<fp16, [1, 8, 7, 97]> 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_4350_to_fp16_quantized)[name = tensor<string, []>("x_475_cast_fp16")];
+            tensor<int32, [8]> x_477_pad_0 = const()[name = tensor<string, []>("x_477_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_477_mode_0 = const()[name = tensor<string, []>("x_477_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_313_to_fp16 = const()[name = tensor<string, []>("const_313_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_477_cast_fp16")];
+            tensor<int32, [4]> var_4358 = const()[name = tensor<string, []>("op_4358"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_479_cast_fp16 = reshape(shape = var_4358, x = x_477_cast_fp16)[name = tensor<string, []>("x_479_cast_fp16")];
+            tensor<int32, [4]> var_4362_begin_0 = const()[name = tensor<string, []>("op_4362_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4362_end_0 = const()[name = tensor<string, []>("op_4362_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_4362_end_mask_0 = const()[name = tensor<string, []>("op_4362_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_4362_cast_fp16 = slice_by_index(begin = var_4362_begin_0, end = var_4362_end_0, end_mask = var_4362_end_mask_0, x = x_479_cast_fp16)[name = tensor<string, []>("op_4362_cast_fp16")];
+            tensor<int32, [4]> var_4363 = const()[name = tensor<string, []>("op_4363"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_73_cast_fp16 = reshape(shape = var_4363, x = var_4362_cast_fp16)[name = tensor<string, []>("matrix_bd_73_cast_fp16")];
+            tensor<bool, []> matrix_ac_37_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_37_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_37_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_37_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_132_perm_0 = const()[name = tensor<string, []>("transpose_132_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_133_perm_0 = const()[name = tensor<string, []>("transpose_133_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73_cast_fp16)[name = tensor<string, []>("transpose_198")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_132 = transpose(perm = transpose_132_perm_0, x = var_4346_cast_fp16)[name = tensor<string, []>("transpose_199")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_37_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_75_begin_0 = const()[name = tensor<string, []>("matrix_bd_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_75_end_0 = const()[name = tensor<string, []>("matrix_bd_75_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_75_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_75_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_4372_cast_fp16 = add(x = matrix_ac_37_cast_fp16, y = matrix_bd_75_cast_fp16)[name = tensor<string, []>("op_4372_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_73_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_73_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_73_cast_fp16 = mul(x = var_4372_cast_fp16, y = _inversed_scores_73_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_73_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_75_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_73_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_75_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_4378_cast_fp16 = softmax(axis = var_59, x = scores_75_cast_fp16)[name = tensor<string, []>("op_4378_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_977_cast_fp16 = select(a = var_44_to_fp16, b = var_4378_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_977_cast_fp16")];
+            tensor<bool, []> x_481_transpose_x_0 = const()[name = tensor<string, []>("x_481_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_481_transpose_y_0 = const()[name = tensor<string, []>("x_481_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = v_37_cast_fp16)[name = tensor<string, []>("transpose_197")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_481_cast_fp16")];
+            tensor<int32, [4]> var_4382_perm_0 = const()[name = tensor<string, []>("op_4382_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4383 = const()[name = tensor<string, []>("op_4383"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_4382_cast_fp16 = transpose(perm = var_4382_perm_0, x = x_481_cast_fp16)[name = tensor<string, []>("transpose_196")];
+            tensor<fp16, [1, 7, 1024]> input_979_cast_fp16 = reshape(shape = var_4383, x = var_4382_cast_fp16)[name = tensor<string, []>("input_979_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369785984))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(370834624))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(370836736)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_169_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_983_cast_fp16 = add(x = input_973_cast_fp16, y = linear_169_cast_fp16)[name = tensor<string, []>("input_983_cast_fp16")];
+            tensor<int32, [1]> x_485_axes_0 = const()[name = tensor<string, []>("x_485_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(370838848)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(370840960)))];
+            tensor<fp16, [1, 7, 1024]> x_485_cast_fp16 = layer_norm(axes = x_485_axes_0, beta = encoder_layers_18_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_conv_weight_to_fp16, x = input_983_cast_fp16)[name = tensor<string, []>("x_485_cast_fp16")];
+            tensor<int32, [3]> input_985_perm_0 = const()[name = tensor<string, []>("input_985_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_987_pad_type_0 = const()[name = tensor<string, []>("input_987_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_987_strides_0 = const()[name = tensor<string, []>("input_987_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_987_pad_0 = const()[name = tensor<string, []>("input_987_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_987_dilations_0 = const()[name = tensor<string, []>("input_987_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_987_groups_0 = const()[name = tensor<string, []>("input_987_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(370843072))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372940288))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_985_cast_fp16 = transpose(perm = input_985_perm_0, x = x_485_cast_fp16)[name = tensor<string, []>("transpose_195")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_987_cast_fp16")];
+            tensor<int32, []> x_487_split_num_splits_0 = const()[name = tensor<string, []>("x_487_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_487_split_axis_0 = const()[name = tensor<string, []>("x_487_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_487_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_487_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_487_split_1_sigmoid_cast_fp16 = sigmoid(x = x_487_split_cast_fp16_1)[name = tensor<string, []>("x_487_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_487_cast_fp16 = mul(x = x_487_split_cast_fp16_0, y = x_487_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_487_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_989_cast_fp16 = select(a = var_44_to_fp16, b = x_487_cast_fp16, cond = var_575)[name = tensor<string, []>("input_989_cast_fp16")];
+            tensor<bool, []> new_x_75_interleave_0 = const()[name = tensor<string, []>("new_x_75_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_75_cast_fp16 = concat(axis = var_59, interleave = new_x_75_interleave_0, values = (cache_75_cast_fp16, input_989_cast_fp16))[name = tensor<string, []>("new_x_75_cast_fp16")];
+            tensor<int32, [3]> var_4422_begin_0 = const()[name = tensor<string, []>("op_4422_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_4422_end_0 = const()[name = tensor<string, []>("op_4422_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_4422_end_mask_0 = const()[name = tensor<string, []>("op_4422_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4422_cast_fp16 = slice_by_index(begin = var_4422_begin_0, end = var_4422_end_0, end_mask = var_4422_end_mask_0, x = new_x_75_cast_fp16)[name = tensor<string, []>("op_4422_cast_fp16")];
+            tensor<string, []> x_489_pad_type_0 = const()[name = tensor<string, []>("x_489_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_489_groups_0 = const()[name = tensor<string, []>("x_489_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_489_strides_0 = const()[name = tensor<string, []>("x_489_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_489_pad_0 = const()[name = tensor<string, []>("x_489_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_489_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372944448))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372953728))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_489_cast_fp16")];
+            tensor<int32, [3]> input_991_perm_0 = const()[name = tensor<string, []>("input_991_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_491_axes_0 = const()[name = tensor<string, []>("x_491_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372955840)))];
+            tensor<fp16, [1024]> encoder_layers_18_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372957952)))];
+            tensor<fp16, [1, 7, 1024]> input_991_cast_fp16 = transpose(perm = input_991_perm_0, x = x_489_cast_fp16)[name = tensor<string, []>("transpose_194")];
+            tensor<fp16, [1, 7, 1024]> x_491_cast_fp16 = layer_norm(axes = x_491_axes_0, beta = encoder_layers_18_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_conv_batch_norm_weight_to_fp16, x = input_991_cast_fp16)[name = tensor<string, []>("x_491_cast_fp16")];
+            tensor<int32, [3]> input_993_perm_0 = const()[name = tensor<string, []>("input_993_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_993_cast_fp16 = transpose(perm = input_993_perm_0, x = x_491_cast_fp16)[name = tensor<string, []>("transpose_193")];
+            tensor<fp16, [1, 1024, 7]> input_995_cast_fp16 = silu(x = input_993_cast_fp16)[name = tensor<string, []>("input_995_cast_fp16")];
+            tensor<string, []> x_493_pad_type_0 = const()[name = tensor<string, []>("x_493_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_493_strides_0 = const()[name = tensor<string, []>("x_493_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_493_pad_0 = const()[name = tensor<string, []>("x_493_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_493_dilations_0 = const()[name = tensor<string, []>("x_493_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_493_groups_0 = const()[name = tensor<string, []>("x_493_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372960064))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(374008704))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_493_cast_fp16")];
+            tensor<int32, [3]> input_997_perm_0 = const()[name = tensor<string, []>("input_997_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_997_cast_fp16 = transpose(perm = input_997_perm_0, x = x_493_cast_fp16)[name = tensor<string, []>("transpose_192")];
+            tensor<fp16, [1, 7, 1024]> input_999_cast_fp16 = add(x = input_983_cast_fp16, y = input_997_cast_fp16)[name = tensor<string, []>("input_999_cast_fp16")];
+            tensor<int32, [1]> input_1001_axes_0 = const()[name = tensor<string, []>("input_1001_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(374010816)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(374012928)))];
+            tensor<fp16, [1, 7, 1024]> input_1001_cast_fp16 = layer_norm(axes = input_1001_axes_0, beta = encoder_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_feed_forward2_weight_to_fp16, x = input_999_cast_fp16)[name = tensor<string, []>("input_1001_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(374015040))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378209408))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805056)))];
+            tensor<fp16, [4096]> encoder_layers_18_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378217664)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_170_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1005_cast_fp16 = silu(x = linear_170_cast_fp16)[name = tensor<string, []>("input_1005_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378225920))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382420288))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_18_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382422400)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_171_cast_fp16")];
+            tensor<fp16, []> var_4465_to_fp16 = const()[name = tensor<string, []>("op_4465_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4466_cast_fp16 = mul(x = linear_171_cast_fp16, y = var_4465_to_fp16)[name = tensor<string, []>("op_4466_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1011_cast_fp16 = add(x = input_999_cast_fp16, y = var_4466_cast_fp16)[name = tensor<string, []>("input_1011_cast_fp16")];
+            tensor<int32, [1]> input_1013_axes_0 = const()[name = tensor<string, []>("input_1013_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382424512)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_18_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382426624)))];
+            tensor<fp16, [1, 7, 1024]> input_1013_cast_fp16 = layer_norm(axes = input_1013_axes_0, beta = encoder_layers_18_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_out_weight_to_fp16, x = input_1011_cast_fp16)[name = tensor<string, []>("input_1013_cast_fp16")];
+            tensor<int32, [4]> cache_77_begin_0 = const()[name = tensor<string, []>("cache_77_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> cache_77_end_0 = const()[name = tensor<string, []>("cache_77_end_0"), val = tensor<int32, [4]>([20, 1, 42, 1024])];
+            tensor<bool, [4]> cache_77_end_mask_0 = const()[name = tensor<string, []>("cache_77_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_77_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_77_cast_fp16")];
+            tensor<int32, [4]> cache_79_begin_0 = const()[name = tensor<string, []>("cache_79_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> cache_79_end_0 = const()[name = tensor<string, []>("cache_79_end_0"), val = tensor<int32, [4]>([20, 1, 1024, 8])];
+            tensor<bool, [4]> cache_79_end_mask_0 = const()[name = tensor<string, []>("cache_79_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_79_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_79_cast_fp16")];
+            tensor<int32, [1]> input_1015_axes_0 = const()[name = tensor<string, []>("input_1015_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382428736)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382430848)))];
+            tensor<fp16, [1, 7, 1024]> input_1015_cast_fp16 = layer_norm(axes = input_1015_axes_0, beta = encoder_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_feed_forward1_weight_to_fp16, x = input_1013_cast_fp16)[name = tensor<string, []>("input_1015_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382432960))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386627328))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805056)))];
+            tensor<fp16, [4096]> encoder_layers_19_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386635584)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_172_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1019_cast_fp16 = silu(x = linear_172_cast_fp16)[name = tensor<string, []>("input_1019_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386643840))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(390838208))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_19_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(390840320)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_173_cast_fp16")];
+            tensor<fp16, []> var_4502_to_fp16 = const()[name = tensor<string, []>("op_4502_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4503_cast_fp16 = mul(x = linear_173_cast_fp16, y = var_4502_to_fp16)[name = tensor<string, []>("op_4503_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1025_cast_fp16 = add(x = input_1013_cast_fp16, y = var_4503_cast_fp16)[name = tensor<string, []>("input_1025_cast_fp16")];
+            tensor<int32, [1]> key_39_axes_0 = const()[name = tensor<string, []>("key_39_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(390842432)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(390844544)))];
+            tensor<fp16, [1, 7, 1024]> key_39_cast_fp16 = layer_norm(axes = key_39_axes_0, beta = encoder_layers_19_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_self_att_weight_to_fp16, x = input_1025_cast_fp16)[name = tensor<string, []>("key_39_cast_fp16")];
+            tensor<bool, []> input_1027_interleave_0 = const()[name = tensor<string, []>("input_1027_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_1027_cast_fp16 = concat(axis = var_68, interleave = input_1027_interleave_0, values = (cache_77_cast_fp16, key_39_cast_fp16))[name = tensor<string, []>("input_1027_cast_fp16")];
+            tensor<int32, [3]> var_4525_begin_0 = const()[name = tensor<string, []>("op_4525_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_4525_end_0 = const()[name = tensor<string, []>("op_4525_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4525_end_mask_0 = const()[name = tensor<string, []>("op_4525_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_4525_cast_fp16 = slice_by_index(begin = var_4525_begin_0, end = var_4525_end_0, end_mask = var_4525_end_mask_0, x = cache_77_cast_fp16)[name = tensor<string, []>("op_4525_cast_fp16")];
+            tensor<bool, []> var_4531_interleave_0 = const()[name = tensor<string, []>("op_4531_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_4531_cast_fp16 = concat(axis = var_68, interleave = var_4531_interleave_0, values = (var_4525_cast_fp16, key_39_cast_fp16))[name = tensor<string, []>("op_4531_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(390846656))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(391895296))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(391897408)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_174_cast_fp16")];
+            tensor<int32, [4]> var_4536 = const()[name = tensor<string, []>("op_4536"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_115_cast_fp16 = reshape(shape = var_4536, x = linear_174_cast_fp16)[name = tensor<string, []>("q_115_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(391899520))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(392948160))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(392950272)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_175_cast_fp16")];
+            tensor<int32, [4]> var_4541 = const()[name = tensor<string, []>("op_4541"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_77_cast_fp16 = reshape(shape = var_4541, x = linear_175_cast_fp16)[name = tensor<string, []>("k_77_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(392952384))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(394001024))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(394003136)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_176_cast_fp16")];
+            tensor<int32, [4]> var_4546 = const()[name = tensor<string, []>("op_4546"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_39_cast_fp16 = reshape(shape = var_4546, x = linear_176_cast_fp16)[name = tensor<string, []>("v_39_cast_fp16")];
+            tensor<int32, [4]> value_47_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_19_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(394005248)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4559_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_4559_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_19_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(394007360)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4561_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_4561_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_39_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_501_transpose_x_0 = const()[name = tensor<string, []>("x_501_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_501_transpose_y_0 = const()[name = tensor<string, []>("x_501_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_4563_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_4563_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(394009472))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(394109056))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(394108864)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_39_cast_fp16 = transpose(perm = q_with_bias_v_39_perm_0, x = var_4561_cast_fp16)[name = tensor<string, []>("transpose_191")];
+            tensor<fp16, [1, 8, 7, 97]> 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_4563_to_fp16_quantized)[name = tensor<string, []>("x_501_cast_fp16")];
+            tensor<int32, [8]> x_503_pad_0 = const()[name = tensor<string, []>("x_503_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_503_mode_0 = const()[name = tensor<string, []>("x_503_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_326_to_fp16 = const()[name = tensor<string, []>("const_326_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_503_cast_fp16")];
+            tensor<int32, [4]> var_4571 = const()[name = tensor<string, []>("op_4571"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_505_cast_fp16 = reshape(shape = var_4571, x = x_503_cast_fp16)[name = tensor<string, []>("x_505_cast_fp16")];
+            tensor<int32, [4]> var_4575_begin_0 = const()[name = tensor<string, []>("op_4575_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4575_end_0 = const()[name = tensor<string, []>("op_4575_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_4575_end_mask_0 = const()[name = tensor<string, []>("op_4575_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_4575_cast_fp16 = slice_by_index(begin = var_4575_begin_0, end = var_4575_end_0, end_mask = var_4575_end_mask_0, x = x_505_cast_fp16)[name = tensor<string, []>("op_4575_cast_fp16")];
+            tensor<int32, [4]> var_4576 = const()[name = tensor<string, []>("op_4576"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_77_cast_fp16 = reshape(shape = var_4576, x = var_4575_cast_fp16)[name = tensor<string, []>("matrix_bd_77_cast_fp16")];
+            tensor<bool, []> matrix_ac_39_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_39_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_39_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_39_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_134_perm_0 = const()[name = tensor<string, []>("transpose_134_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_135_perm_0 = const()[name = tensor<string, []>("transpose_135_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77_cast_fp16)[name = tensor<string, []>("transpose_189")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_134 = transpose(perm = transpose_134_perm_0, x = var_4559_cast_fp16)[name = tensor<string, []>("transpose_190")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_39_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_79_begin_0 = const()[name = tensor<string, []>("matrix_bd_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_79_end_0 = const()[name = tensor<string, []>("matrix_bd_79_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_79_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_79_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_4585_cast_fp16 = add(x = matrix_ac_39_cast_fp16, y = matrix_bd_79_cast_fp16)[name = tensor<string, []>("op_4585_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_77_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_77_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_77_cast_fp16 = mul(x = var_4585_cast_fp16, y = _inversed_scores_77_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_77_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_79_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_77_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_79_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_4591_cast_fp16 = softmax(axis = var_59, x = scores_79_cast_fp16)[name = tensor<string, []>("op_4591_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_1029_cast_fp16 = select(a = var_44_to_fp16, b = var_4591_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_1029_cast_fp16")];
+            tensor<bool, []> x_507_transpose_x_0 = const()[name = tensor<string, []>("x_507_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_507_transpose_y_0 = const()[name = tensor<string, []>("x_507_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = v_39_cast_fp16)[name = tensor<string, []>("transpose_188")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_507_cast_fp16")];
+            tensor<int32, [4]> var_4595_perm_0 = const()[name = tensor<string, []>("op_4595_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4596 = const()[name = tensor<string, []>("op_4596"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_4595_cast_fp16 = transpose(perm = var_4595_perm_0, x = x_507_cast_fp16)[name = tensor<string, []>("transpose_187")];
+            tensor<fp16, [1, 7, 1024]> input_1031_cast_fp16 = reshape(shape = var_4596, x = var_4595_cast_fp16)[name = tensor<string, []>("input_1031_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(394109376))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395158016))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395160128)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_178_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1035_cast_fp16 = add(x = input_1025_cast_fp16, y = linear_178_cast_fp16)[name = tensor<string, []>("input_1035_cast_fp16")];
+            tensor<int32, [1]> x_511_axes_0 = const()[name = tensor<string, []>("x_511_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395162240)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395164352)))];
+            tensor<fp16, [1, 7, 1024]> x_511_cast_fp16 = layer_norm(axes = x_511_axes_0, beta = encoder_layers_19_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_conv_weight_to_fp16, x = input_1035_cast_fp16)[name = tensor<string, []>("x_511_cast_fp16")];
+            tensor<int32, [3]> input_1037_perm_0 = const()[name = tensor<string, []>("input_1037_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_1039_pad_type_0 = const()[name = tensor<string, []>("input_1039_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_1039_strides_0 = const()[name = tensor<string, []>("input_1039_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1039_pad_0 = const()[name = tensor<string, []>("input_1039_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1039_dilations_0 = const()[name = tensor<string, []>("input_1039_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_1039_groups_0 = const()[name = tensor<string, []>("input_1039_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395166464))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(397263680))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_1037_cast_fp16 = transpose(perm = input_1037_perm_0, x = x_511_cast_fp16)[name = tensor<string, []>("transpose_186")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_1039_cast_fp16")];
+            tensor<int32, []> x_513_split_num_splits_0 = const()[name = tensor<string, []>("x_513_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_513_split_axis_0 = const()[name = tensor<string, []>("x_513_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_513_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_513_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_513_split_1_sigmoid_cast_fp16 = sigmoid(x = x_513_split_cast_fp16_1)[name = tensor<string, []>("x_513_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_513_cast_fp16 = mul(x = x_513_split_cast_fp16_0, y = x_513_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_513_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_1041_cast_fp16 = select(a = var_44_to_fp16, b = x_513_cast_fp16, cond = var_575)[name = tensor<string, []>("input_1041_cast_fp16")];
+            tensor<bool, []> new_x_79_interleave_0 = const()[name = tensor<string, []>("new_x_79_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_79_cast_fp16 = concat(axis = var_59, interleave = new_x_79_interleave_0, values = (cache_79_cast_fp16, input_1041_cast_fp16))[name = tensor<string, []>("new_x_79_cast_fp16")];
+            tensor<int32, [3]> var_4635_begin_0 = const()[name = tensor<string, []>("op_4635_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_4635_end_0 = const()[name = tensor<string, []>("op_4635_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_4635_end_mask_0 = const()[name = tensor<string, []>("op_4635_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4635_cast_fp16 = slice_by_index(begin = var_4635_begin_0, end = var_4635_end_0, end_mask = var_4635_end_mask_0, x = new_x_79_cast_fp16)[name = tensor<string, []>("op_4635_cast_fp16")];
+            tensor<string, []> x_515_pad_type_0 = const()[name = tensor<string, []>("x_515_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_515_groups_0 = const()[name = tensor<string, []>("x_515_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_515_strides_0 = const()[name = tensor<string, []>("x_515_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_515_pad_0 = const()[name = tensor<string, []>("x_515_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_515_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(397267840))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(397277120))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_515_cast_fp16")];
+            tensor<int32, [3]> input_1043_perm_0 = const()[name = tensor<string, []>("input_1043_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_517_axes_0 = const()[name = tensor<string, []>("x_517_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(397279232)))];
+            tensor<fp16, [1024]> encoder_layers_19_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(397281344)))];
+            tensor<fp16, [1, 7, 1024]> input_1043_cast_fp16 = transpose(perm = input_1043_perm_0, x = x_515_cast_fp16)[name = tensor<string, []>("transpose_185")];
+            tensor<fp16, [1, 7, 1024]> x_517_cast_fp16 = layer_norm(axes = x_517_axes_0, beta = encoder_layers_19_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_conv_batch_norm_weight_to_fp16, x = input_1043_cast_fp16)[name = tensor<string, []>("x_517_cast_fp16")];
+            tensor<int32, [3]> input_1045_perm_0 = const()[name = tensor<string, []>("input_1045_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_1045_cast_fp16 = transpose(perm = input_1045_perm_0, x = x_517_cast_fp16)[name = tensor<string, []>("transpose_184")];
+            tensor<fp16, [1, 1024, 7]> input_1047_cast_fp16 = silu(x = input_1045_cast_fp16)[name = tensor<string, []>("input_1047_cast_fp16")];
+            tensor<string, []> x_519_pad_type_0 = const()[name = tensor<string, []>("x_519_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_519_strides_0 = const()[name = tensor<string, []>("x_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_519_pad_0 = const()[name = tensor<string, []>("x_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_519_dilations_0 = const()[name = tensor<string, []>("x_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_519_groups_0 = const()[name = tensor<string, []>("x_519_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(397283456))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(398332096))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_519_cast_fp16")];
+            tensor<int32, [3]> input_1049_perm_0 = const()[name = tensor<string, []>("input_1049_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_1049_cast_fp16 = transpose(perm = input_1049_perm_0, x = x_519_cast_fp16)[name = tensor<string, []>("transpose_183")];
+            tensor<fp16, [1, 7, 1024]> input_1051_cast_fp16 = add(x = input_1035_cast_fp16, y = input_1049_cast_fp16)[name = tensor<string, []>("input_1051_cast_fp16")];
+            tensor<int32, [1]> input_1053_axes_0 = const()[name = tensor<string, []>("input_1053_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(398334208)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(398336320)))];
+            tensor<fp16, [1, 7, 1024]> input_1053_cast_fp16 = layer_norm(axes = input_1053_axes_0, beta = encoder_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_feed_forward2_weight_to_fp16, x = input_1051_cast_fp16)[name = tensor<string, []>("input_1053_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(398338432))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(402532800))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805056)))];
+            tensor<fp16, [4096]> encoder_layers_19_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(402541056)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_179_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1057_cast_fp16 = silu(x = linear_179_cast_fp16)[name = tensor<string, []>("input_1057_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(402549312))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406743680))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_19_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406745792)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_180_cast_fp16")];
+            tensor<fp16, []> var_4678_to_fp16 = const()[name = tensor<string, []>("op_4678_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4679_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_4678_to_fp16)[name = tensor<string, []>("op_4679_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1063_cast_fp16 = add(x = input_1051_cast_fp16, y = var_4679_cast_fp16)[name = tensor<string, []>("input_1063_cast_fp16")];
+            tensor<int32, [1]> input_1065_axes_0 = const()[name = tensor<string, []>("input_1065_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406747904)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_19_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406750016)))];
+            tensor<fp16, [1, 7, 1024]> input_1065_cast_fp16 = layer_norm(axes = input_1065_axes_0, beta = encoder_layers_19_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_out_weight_to_fp16, x = input_1063_cast_fp16)[name = tensor<string, []>("input_1065_cast_fp16")];
+            tensor<int32, [4]> cache_81_begin_0 = const()[name = tensor<string, []>("cache_81_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> cache_81_end_0 = const()[name = tensor<string, []>("cache_81_end_0"), val = tensor<int32, [4]>([21, 1, 42, 1024])];
+            tensor<bool, [4]> cache_81_end_mask_0 = const()[name = tensor<string, []>("cache_81_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_81_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_81_cast_fp16")];
+            tensor<int32, [4]> cache_83_begin_0 = const()[name = tensor<string, []>("cache_83_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> cache_83_end_0 = const()[name = tensor<string, []>("cache_83_end_0"), val = tensor<int32, [4]>([21, 1, 1024, 8])];
+            tensor<bool, [4]> cache_83_end_mask_0 = const()[name = tensor<string, []>("cache_83_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_83_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_83_cast_fp16")];
+            tensor<int32, [1]> input_1067_axes_0 = const()[name = tensor<string, []>("input_1067_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406752128)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406754240)))];
+            tensor<fp16, [1, 7, 1024]> input_1067_cast_fp16 = layer_norm(axes = input_1067_axes_0, beta = encoder_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_feed_forward1_weight_to_fp16, x = input_1065_cast_fp16)[name = tensor<string, []>("input_1067_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406756352))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(410950720))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805056)))];
+            tensor<fp16, [4096]> encoder_layers_20_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(410958976)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_181_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1071_cast_fp16 = silu(x = linear_181_cast_fp16)[name = tensor<string, []>("input_1071_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(410967232))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(415161600))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_20_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(415163712)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_182_cast_fp16")];
+            tensor<fp16, []> var_4715_to_fp16 = const()[name = tensor<string, []>("op_4715_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4716_cast_fp16 = mul(x = linear_182_cast_fp16, y = var_4715_to_fp16)[name = tensor<string, []>("op_4716_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1077_cast_fp16 = add(x = input_1065_cast_fp16, y = var_4716_cast_fp16)[name = tensor<string, []>("input_1077_cast_fp16")];
+            tensor<int32, [1]> key_41_axes_0 = const()[name = tensor<string, []>("key_41_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(415165824)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(415167936)))];
+            tensor<fp16, [1, 7, 1024]> key_41_cast_fp16 = layer_norm(axes = key_41_axes_0, beta = encoder_layers_20_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_self_att_weight_to_fp16, x = input_1077_cast_fp16)[name = tensor<string, []>("key_41_cast_fp16")];
+            tensor<bool, []> input_1079_interleave_0 = const()[name = tensor<string, []>("input_1079_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_1079_cast_fp16 = concat(axis = var_68, interleave = input_1079_interleave_0, values = (cache_81_cast_fp16, key_41_cast_fp16))[name = tensor<string, []>("input_1079_cast_fp16")];
+            tensor<int32, [3]> var_4738_begin_0 = const()[name = tensor<string, []>("op_4738_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_4738_end_0 = const()[name = tensor<string, []>("op_4738_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4738_end_mask_0 = const()[name = tensor<string, []>("op_4738_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_4738_cast_fp16 = slice_by_index(begin = var_4738_begin_0, end = var_4738_end_0, end_mask = var_4738_end_mask_0, x = cache_81_cast_fp16)[name = tensor<string, []>("op_4738_cast_fp16")];
+            tensor<bool, []> var_4744_interleave_0 = const()[name = tensor<string, []>("op_4744_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_4744_cast_fp16 = concat(axis = var_68, interleave = var_4744_interleave_0, values = (var_4738_cast_fp16, key_41_cast_fp16))[name = tensor<string, []>("op_4744_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(415170048))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(416218688))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(416220800)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_183_cast_fp16")];
+            tensor<int32, [4]> var_4749 = const()[name = tensor<string, []>("op_4749"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_121_cast_fp16 = reshape(shape = var_4749, x = linear_183_cast_fp16)[name = tensor<string, []>("q_121_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(416222912))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417271552))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417273664)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_184_cast_fp16")];
+            tensor<int32, [4]> var_4754 = const()[name = tensor<string, []>("op_4754"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_81_cast_fp16 = reshape(shape = var_4754, x = linear_184_cast_fp16)[name = tensor<string, []>("k_81_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417275776))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(418324416))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(418326528)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_185_cast_fp16")];
+            tensor<int32, [4]> var_4759 = const()[name = tensor<string, []>("op_4759"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_41_cast_fp16 = reshape(shape = var_4759, x = linear_185_cast_fp16)[name = tensor<string, []>("v_41_cast_fp16")];
+            tensor<int32, [4]> value_49_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_20_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(418328640)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4772_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_4772_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_20_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(418330752)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4774_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_4774_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_41_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_527_transpose_x_0 = const()[name = tensor<string, []>("x_527_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_527_transpose_y_0 = const()[name = tensor<string, []>("x_527_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_4776_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_4776_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(418332864))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(418432448))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(418432256)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_41_cast_fp16 = transpose(perm = q_with_bias_v_41_perm_0, x = var_4774_cast_fp16)[name = tensor<string, []>("transpose_182")];
+            tensor<fp16, [1, 8, 7, 97]> 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_4776_to_fp16_quantized)[name = tensor<string, []>("x_527_cast_fp16")];
+            tensor<int32, [8]> x_529_pad_0 = const()[name = tensor<string, []>("x_529_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_529_mode_0 = const()[name = tensor<string, []>("x_529_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_339_to_fp16 = const()[name = tensor<string, []>("const_339_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_529_cast_fp16")];
+            tensor<int32, [4]> var_4784 = const()[name = tensor<string, []>("op_4784"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_531_cast_fp16 = reshape(shape = var_4784, x = x_529_cast_fp16)[name = tensor<string, []>("x_531_cast_fp16")];
+            tensor<int32, [4]> var_4788_begin_0 = const()[name = tensor<string, []>("op_4788_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4788_end_0 = const()[name = tensor<string, []>("op_4788_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_4788_end_mask_0 = const()[name = tensor<string, []>("op_4788_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_4788_cast_fp16 = slice_by_index(begin = var_4788_begin_0, end = var_4788_end_0, end_mask = var_4788_end_mask_0, x = x_531_cast_fp16)[name = tensor<string, []>("op_4788_cast_fp16")];
+            tensor<int32, [4]> var_4789 = const()[name = tensor<string, []>("op_4789"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_81_cast_fp16 = reshape(shape = var_4789, x = var_4788_cast_fp16)[name = tensor<string, []>("matrix_bd_81_cast_fp16")];
+            tensor<bool, []> matrix_ac_41_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_41_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_41_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_41_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_136_perm_0 = const()[name = tensor<string, []>("transpose_136_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_137_perm_0 = const()[name = tensor<string, []>("transpose_137_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81_cast_fp16)[name = tensor<string, []>("transpose_180")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_136 = transpose(perm = transpose_136_perm_0, x = var_4772_cast_fp16)[name = tensor<string, []>("transpose_181")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_41_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_83_begin_0 = const()[name = tensor<string, []>("matrix_bd_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_83_end_0 = const()[name = tensor<string, []>("matrix_bd_83_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_83_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_83_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_4798_cast_fp16 = add(x = matrix_ac_41_cast_fp16, y = matrix_bd_83_cast_fp16)[name = tensor<string, []>("op_4798_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_81_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_81_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_81_cast_fp16 = mul(x = var_4798_cast_fp16, y = _inversed_scores_81_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_81_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_83_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_81_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_83_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_4804_cast_fp16 = softmax(axis = var_59, x = scores_83_cast_fp16)[name = tensor<string, []>("op_4804_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_1081_cast_fp16 = select(a = var_44_to_fp16, b = var_4804_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_1081_cast_fp16")];
+            tensor<bool, []> x_533_transpose_x_0 = const()[name = tensor<string, []>("x_533_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_533_transpose_y_0 = const()[name = tensor<string, []>("x_533_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = v_41_cast_fp16)[name = tensor<string, []>("transpose_179")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_533_cast_fp16")];
+            tensor<int32, [4]> var_4808_perm_0 = const()[name = tensor<string, []>("op_4808_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4809 = const()[name = tensor<string, []>("op_4809"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_4808_cast_fp16 = transpose(perm = var_4808_perm_0, x = x_533_cast_fp16)[name = tensor<string, []>("transpose_178")];
+            tensor<fp16, [1, 7, 1024]> input_1083_cast_fp16 = reshape(shape = var_4809, x = var_4808_cast_fp16)[name = tensor<string, []>("input_1083_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(418432768))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419481408))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419483520)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_187_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1087_cast_fp16 = add(x = input_1077_cast_fp16, y = linear_187_cast_fp16)[name = tensor<string, []>("input_1087_cast_fp16")];
+            tensor<int32, [1]> x_537_axes_0 = const()[name = tensor<string, []>("x_537_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419485632)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419487744)))];
+            tensor<fp16, [1, 7, 1024]> x_537_cast_fp16 = layer_norm(axes = x_537_axes_0, beta = encoder_layers_20_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_conv_weight_to_fp16, x = input_1087_cast_fp16)[name = tensor<string, []>("x_537_cast_fp16")];
+            tensor<int32, [3]> input_1089_perm_0 = const()[name = tensor<string, []>("input_1089_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_1091_pad_type_0 = const()[name = tensor<string, []>("input_1091_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_1091_strides_0 = const()[name = tensor<string, []>("input_1091_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1091_pad_0 = const()[name = tensor<string, []>("input_1091_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1091_dilations_0 = const()[name = tensor<string, []>("input_1091_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_1091_groups_0 = const()[name = tensor<string, []>("input_1091_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419489856))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421587072))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_1089_cast_fp16 = transpose(perm = input_1089_perm_0, x = x_537_cast_fp16)[name = tensor<string, []>("transpose_177")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_1091_cast_fp16")];
+            tensor<int32, []> x_539_split_num_splits_0 = const()[name = tensor<string, []>("x_539_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_539_split_axis_0 = const()[name = tensor<string, []>("x_539_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_539_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_539_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_539_split_1_sigmoid_cast_fp16 = sigmoid(x = x_539_split_cast_fp16_1)[name = tensor<string, []>("x_539_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_539_cast_fp16 = mul(x = x_539_split_cast_fp16_0, y = x_539_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_539_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_1093_cast_fp16 = select(a = var_44_to_fp16, b = x_539_cast_fp16, cond = var_575)[name = tensor<string, []>("input_1093_cast_fp16")];
+            tensor<bool, []> new_x_83_interleave_0 = const()[name = tensor<string, []>("new_x_83_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_83_cast_fp16 = concat(axis = var_59, interleave = new_x_83_interleave_0, values = (cache_83_cast_fp16, input_1093_cast_fp16))[name = tensor<string, []>("new_x_83_cast_fp16")];
+            tensor<int32, [3]> var_4848_begin_0 = const()[name = tensor<string, []>("op_4848_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_4848_end_0 = const()[name = tensor<string, []>("op_4848_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_4848_end_mask_0 = const()[name = tensor<string, []>("op_4848_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4848_cast_fp16 = slice_by_index(begin = var_4848_begin_0, end = var_4848_end_0, end_mask = var_4848_end_mask_0, x = new_x_83_cast_fp16)[name = tensor<string, []>("op_4848_cast_fp16")];
+            tensor<string, []> x_541_pad_type_0 = const()[name = tensor<string, []>("x_541_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_541_groups_0 = const()[name = tensor<string, []>("x_541_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_541_strides_0 = const()[name = tensor<string, []>("x_541_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_541_pad_0 = const()[name = tensor<string, []>("x_541_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_541_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421591232))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421600512))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_541_cast_fp16")];
+            tensor<int32, [3]> input_1095_perm_0 = const()[name = tensor<string, []>("input_1095_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_543_axes_0 = const()[name = tensor<string, []>("x_543_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421602624)))];
+            tensor<fp16, [1024]> encoder_layers_20_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421604736)))];
+            tensor<fp16, [1, 7, 1024]> input_1095_cast_fp16 = transpose(perm = input_1095_perm_0, x = x_541_cast_fp16)[name = tensor<string, []>("transpose_176")];
+            tensor<fp16, [1, 7, 1024]> x_543_cast_fp16 = layer_norm(axes = x_543_axes_0, beta = encoder_layers_20_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_conv_batch_norm_weight_to_fp16, x = input_1095_cast_fp16)[name = tensor<string, []>("x_543_cast_fp16")];
+            tensor<int32, [3]> input_1097_perm_0 = const()[name = tensor<string, []>("input_1097_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_1097_cast_fp16 = transpose(perm = input_1097_perm_0, x = x_543_cast_fp16)[name = tensor<string, []>("transpose_175")];
+            tensor<fp16, [1, 1024, 7]> input_1099_cast_fp16 = silu(x = input_1097_cast_fp16)[name = tensor<string, []>("input_1099_cast_fp16")];
+            tensor<string, []> x_545_pad_type_0 = const()[name = tensor<string, []>("x_545_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_545_strides_0 = const()[name = tensor<string, []>("x_545_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_545_pad_0 = const()[name = tensor<string, []>("x_545_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_545_dilations_0 = const()[name = tensor<string, []>("x_545_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_545_groups_0 = const()[name = tensor<string, []>("x_545_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421606848))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(422655488))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_545_cast_fp16")];
+            tensor<int32, [3]> input_1101_perm_0 = const()[name = tensor<string, []>("input_1101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_1101_cast_fp16 = transpose(perm = input_1101_perm_0, x = x_545_cast_fp16)[name = tensor<string, []>("transpose_174")];
+            tensor<fp16, [1, 7, 1024]> input_1103_cast_fp16 = add(x = input_1087_cast_fp16, y = input_1101_cast_fp16)[name = tensor<string, []>("input_1103_cast_fp16")];
+            tensor<int32, [1]> input_1105_axes_0 = const()[name = tensor<string, []>("input_1105_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(422657600)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(422659712)))];
+            tensor<fp16, [1, 7, 1024]> input_1105_cast_fp16 = layer_norm(axes = input_1105_axes_0, beta = encoder_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_feed_forward2_weight_to_fp16, x = input_1103_cast_fp16)[name = tensor<string, []>("input_1105_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(422661824))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(426856192))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805056)))];
+            tensor<fp16, [4096]> encoder_layers_20_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(426864448)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_188_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1109_cast_fp16 = silu(x = linear_188_cast_fp16)[name = tensor<string, []>("input_1109_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(426872704))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431067072))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_20_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431069184)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_189_cast_fp16")];
+            tensor<fp16, []> var_4891_to_fp16 = const()[name = tensor<string, []>("op_4891_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4892_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_4891_to_fp16)[name = tensor<string, []>("op_4892_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1115_cast_fp16 = add(x = input_1103_cast_fp16, y = var_4892_cast_fp16)[name = tensor<string, []>("input_1115_cast_fp16")];
+            tensor<int32, [1]> input_1117_axes_0 = const()[name = tensor<string, []>("input_1117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431071296)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_20_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431073408)))];
+            tensor<fp16, [1, 7, 1024]> input_1117_cast_fp16 = layer_norm(axes = input_1117_axes_0, beta = encoder_layers_20_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_out_weight_to_fp16, x = input_1115_cast_fp16)[name = tensor<string, []>("input_1117_cast_fp16")];
+            tensor<int32, [4]> cache_85_begin_0 = const()[name = tensor<string, []>("cache_85_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> cache_85_end_0 = const()[name = tensor<string, []>("cache_85_end_0"), val = tensor<int32, [4]>([22, 1, 42, 1024])];
+            tensor<bool, [4]> cache_85_end_mask_0 = const()[name = tensor<string, []>("cache_85_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_85_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_85_cast_fp16")];
+            tensor<int32, [4]> cache_87_begin_0 = const()[name = tensor<string, []>("cache_87_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> cache_87_end_0 = const()[name = tensor<string, []>("cache_87_end_0"), val = tensor<int32, [4]>([22, 1, 1024, 8])];
+            tensor<bool, [4]> cache_87_end_mask_0 = const()[name = tensor<string, []>("cache_87_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_87_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_87_cast_fp16")];
+            tensor<int32, [1]> input_1119_axes_0 = const()[name = tensor<string, []>("input_1119_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431075520)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431077632)))];
+            tensor<fp16, [1, 7, 1024]> input_1119_cast_fp16 = layer_norm(axes = input_1119_axes_0, beta = encoder_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_feed_forward1_weight_to_fp16, x = input_1117_cast_fp16)[name = tensor<string, []>("input_1119_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431079744))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(435274112))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8805056)))];
+            tensor<fp16, [4096]> encoder_layers_21_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(435282368)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_190_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1123_cast_fp16 = silu(x = linear_190_cast_fp16)[name = tensor<string, []>("input_1123_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(435290624))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(439484992))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_21_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(439487104)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_191_cast_fp16")];
+            tensor<fp16, []> var_4928_to_fp16 = const()[name = tensor<string, []>("op_4928_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4929_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_4928_to_fp16)[name = tensor<string, []>("op_4929_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1129_cast_fp16 = add(x = input_1117_cast_fp16, y = var_4929_cast_fp16)[name = tensor<string, []>("input_1129_cast_fp16")];
+            tensor<int32, [1]> key_43_axes_0 = const()[name = tensor<string, []>("key_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(439489216)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(439491328)))];
+            tensor<fp16, [1, 7, 1024]> key_43_cast_fp16 = layer_norm(axes = key_43_axes_0, beta = encoder_layers_21_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_self_att_weight_to_fp16, x = input_1129_cast_fp16)[name = tensor<string, []>("key_43_cast_fp16")];
+            tensor<bool, []> input_1131_interleave_0 = const()[name = tensor<string, []>("input_1131_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_1131_cast_fp16 = concat(axis = var_68, interleave = input_1131_interleave_0, values = (cache_85_cast_fp16, key_43_cast_fp16))[name = tensor<string, []>("input_1131_cast_fp16")];
+            tensor<int32, [3]> var_4951_begin_0 = const()[name = tensor<string, []>("op_4951_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_4951_end_0 = const()[name = tensor<string, []>("op_4951_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4951_end_mask_0 = const()[name = tensor<string, []>("op_4951_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_4951_cast_fp16 = slice_by_index(begin = var_4951_begin_0, end = var_4951_end_0, end_mask = var_4951_end_mask_0, x = cache_85_cast_fp16)[name = tensor<string, []>("op_4951_cast_fp16")];
+            tensor<bool, []> var_4957_interleave_0 = const()[name = tensor<string, []>("op_4957_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_4957_cast_fp16 = concat(axis = var_68, interleave = var_4957_interleave_0, values = (var_4951_cast_fp16, key_43_cast_fp16))[name = tensor<string, []>("op_4957_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(439493440))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(440542080))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(440544192)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_192_cast_fp16")];
+            tensor<int32, [4]> var_4962 = const()[name = tensor<string, []>("op_4962"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_127_cast_fp16 = reshape(shape = var_4962, x = linear_192_cast_fp16)[name = tensor<string, []>("q_127_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(440546304))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(441594944))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(441597056)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_193_cast_fp16")];
+            tensor<int32, [4]> var_4967 = const()[name = tensor<string, []>("op_4967"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_85_cast_fp16 = reshape(shape = var_4967, x = linear_193_cast_fp16)[name = tensor<string, []>("k_85_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(441599168))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442647808))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442649920)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_194_cast_fp16")];
+            tensor<int32, [4]> var_4972 = const()[name = tensor<string, []>("op_4972"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_43_cast_fp16 = reshape(shape = var_4972, x = linear_194_cast_fp16)[name = tensor<string, []>("v_43_cast_fp16")];
+            tensor<int32, [4]> value_51_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_21_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442652032)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4985_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_4985_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_21_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442654144)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4987_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_4987_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_43_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_553_transpose_x_0 = const()[name = tensor<string, []>("x_553_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_553_transpose_y_0 = const()[name = tensor<string, []>("x_553_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_4989_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_4989_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442656256))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442755840))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442755648)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_43_cast_fp16 = transpose(perm = q_with_bias_v_43_perm_0, x = var_4987_cast_fp16)[name = tensor<string, []>("transpose_173")];
+            tensor<fp16, [1, 8, 7, 97]> 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_4989_to_fp16_quantized)[name = tensor<string, []>("x_553_cast_fp16")];
+            tensor<int32, [8]> x_555_pad_0 = const()[name = tensor<string, []>("x_555_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_555_mode_0 = const()[name = tensor<string, []>("x_555_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_352_to_fp16 = const()[name = tensor<string, []>("const_352_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_555_cast_fp16")];
+            tensor<int32, [4]> var_4997 = const()[name = tensor<string, []>("op_4997"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_557_cast_fp16 = reshape(shape = var_4997, x = x_555_cast_fp16)[name = tensor<string, []>("x_557_cast_fp16")];
+            tensor<int32, [4]> var_5001_begin_0 = const()[name = tensor<string, []>("op_5001_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5001_end_0 = const()[name = tensor<string, []>("op_5001_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_5001_end_mask_0 = const()[name = tensor<string, []>("op_5001_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_5001_cast_fp16 = slice_by_index(begin = var_5001_begin_0, end = var_5001_end_0, end_mask = var_5001_end_mask_0, x = x_557_cast_fp16)[name = tensor<string, []>("op_5001_cast_fp16")];
+            tensor<int32, [4]> var_5002 = const()[name = tensor<string, []>("op_5002"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_85_cast_fp16 = reshape(shape = var_5002, x = var_5001_cast_fp16)[name = tensor<string, []>("matrix_bd_85_cast_fp16")];
+            tensor<bool, []> matrix_ac_43_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_43_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_43_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_43_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_138_perm_0 = const()[name = tensor<string, []>("transpose_138_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_139_perm_0 = const()[name = tensor<string, []>("transpose_139_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85_cast_fp16)[name = tensor<string, []>("transpose_171")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_138 = transpose(perm = transpose_138_perm_0, x = var_4985_cast_fp16)[name = tensor<string, []>("transpose_172")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_43_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_87_begin_0 = const()[name = tensor<string, []>("matrix_bd_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_87_end_0 = const()[name = tensor<string, []>("matrix_bd_87_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_87_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_87_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_5011_cast_fp16 = add(x = matrix_ac_43_cast_fp16, y = matrix_bd_87_cast_fp16)[name = tensor<string, []>("op_5011_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_85_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_85_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_85_cast_fp16 = mul(x = var_5011_cast_fp16, y = _inversed_scores_85_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_85_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_87_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_85_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_87_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_5017_cast_fp16 = softmax(axis = var_59, x = scores_87_cast_fp16)[name = tensor<string, []>("op_5017_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_1133_cast_fp16 = select(a = var_44_to_fp16, b = var_5017_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_1133_cast_fp16")];
+            tensor<bool, []> x_559_transpose_x_0 = const()[name = tensor<string, []>("x_559_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_559_transpose_y_0 = const()[name = tensor<string, []>("x_559_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = v_43_cast_fp16)[name = tensor<string, []>("transpose_170")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_559_cast_fp16")];
+            tensor<int32, [4]> var_5021_perm_0 = const()[name = tensor<string, []>("op_5021_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5022 = const()[name = tensor<string, []>("op_5022"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_5021_cast_fp16 = transpose(perm = var_5021_perm_0, x = x_559_cast_fp16)[name = tensor<string, []>("transpose_169")];
+            tensor<fp16, [1, 7, 1024]> input_1135_cast_fp16 = reshape(shape = var_5022, x = var_5021_cast_fp16)[name = tensor<string, []>("input_1135_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442756160)))];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444853376)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_196_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1139_cast_fp16 = add(x = input_1129_cast_fp16, y = linear_196_cast_fp16)[name = tensor<string, []>("input_1139_cast_fp16")];
+            tensor<int32, [1]> x_563_axes_0 = const()[name = tensor<string, []>("x_563_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444855488)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444857600)))];
+            tensor<fp16, [1, 7, 1024]> x_563_cast_fp16 = layer_norm(axes = x_563_axes_0, beta = encoder_layers_21_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_conv_weight_to_fp16, x = input_1139_cast_fp16)[name = tensor<string, []>("x_563_cast_fp16")];
+            tensor<int32, [3]> input_1141_perm_0 = const()[name = tensor<string, []>("input_1141_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_1143_pad_type_0 = const()[name = tensor<string, []>("input_1143_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_1143_strides_0 = const()[name = tensor<string, []>("input_1143_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1143_pad_0 = const()[name = tensor<string, []>("input_1143_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1143_dilations_0 = const()[name = tensor<string, []>("input_1143_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_1143_groups_0 = const()[name = tensor<string, []>("input_1143_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444859712))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(446956928))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_1141_cast_fp16 = transpose(perm = input_1141_perm_0, x = x_563_cast_fp16)[name = tensor<string, []>("transpose_168")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_1143_cast_fp16")];
+            tensor<int32, []> x_565_split_num_splits_0 = const()[name = tensor<string, []>("x_565_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_565_split_axis_0 = const()[name = tensor<string, []>("x_565_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_565_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_565_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_565_split_1_sigmoid_cast_fp16 = sigmoid(x = x_565_split_cast_fp16_1)[name = tensor<string, []>("x_565_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_565_cast_fp16 = mul(x = x_565_split_cast_fp16_0, y = x_565_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_565_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_1145_cast_fp16 = select(a = var_44_to_fp16, b = x_565_cast_fp16, cond = var_575)[name = tensor<string, []>("input_1145_cast_fp16")];
+            tensor<bool, []> new_x_87_interleave_0 = const()[name = tensor<string, []>("new_x_87_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_87_cast_fp16 = concat(axis = var_59, interleave = new_x_87_interleave_0, values = (cache_87_cast_fp16, input_1145_cast_fp16))[name = tensor<string, []>("new_x_87_cast_fp16")];
+            tensor<int32, [3]> var_5061_begin_0 = const()[name = tensor<string, []>("op_5061_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_5061_end_0 = const()[name = tensor<string, []>("op_5061_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_5061_end_mask_0 = const()[name = tensor<string, []>("op_5061_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_5061_cast_fp16 = slice_by_index(begin = var_5061_begin_0, end = var_5061_end_0, end_mask = var_5061_end_mask_0, x = new_x_87_cast_fp16)[name = tensor<string, []>("op_5061_cast_fp16")];
+            tensor<string, []> x_567_pad_type_0 = const()[name = tensor<string, []>("x_567_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_567_groups_0 = const()[name = tensor<string, []>("x_567_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_567_strides_0 = const()[name = tensor<string, []>("x_567_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_567_pad_0 = const()[name = tensor<string, []>("x_567_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_567_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(446961088))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(446970368))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_567_cast_fp16")];
+            tensor<int32, [3]> input_1147_perm_0 = const()[name = tensor<string, []>("input_1147_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_569_axes_0 = const()[name = tensor<string, []>("x_569_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(446972480)))];
+            tensor<fp16, [1024]> encoder_layers_21_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(446974592)))];
+            tensor<fp16, [1, 7, 1024]> input_1147_cast_fp16 = transpose(perm = input_1147_perm_0, x = x_567_cast_fp16)[name = tensor<string, []>("transpose_167")];
+            tensor<fp16, [1, 7, 1024]> x_569_cast_fp16 = layer_norm(axes = x_569_axes_0, beta = encoder_layers_21_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_conv_batch_norm_weight_to_fp16, x = input_1147_cast_fp16)[name = tensor<string, []>("x_569_cast_fp16")];
+            tensor<int32, [3]> input_1149_perm_0 = const()[name = tensor<string, []>("input_1149_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_1149_cast_fp16 = transpose(perm = input_1149_perm_0, x = x_569_cast_fp16)[name = tensor<string, []>("transpose_166")];
+            tensor<fp16, [1, 1024, 7]> input_1151_cast_fp16 = silu(x = input_1149_cast_fp16)[name = tensor<string, []>("input_1151_cast_fp16")];
+            tensor<string, []> x_571_pad_type_0 = const()[name = tensor<string, []>("x_571_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_571_strides_0 = const()[name = tensor<string, []>("x_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_571_pad_0 = const()[name = tensor<string, []>("x_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_571_dilations_0 = const()[name = tensor<string, []>("x_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_571_groups_0 = const()[name = tensor<string, []>("x_571_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(446976704))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(448025344))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_571_cast_fp16")];
+            tensor<int32, [3]> input_1153_perm_0 = const()[name = tensor<string, []>("input_1153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_1153_cast_fp16 = transpose(perm = input_1153_perm_0, x = x_571_cast_fp16)[name = tensor<string, []>("transpose_165")];
+            tensor<fp16, [1, 7, 1024]> input_1155_cast_fp16 = add(x = input_1139_cast_fp16, y = input_1153_cast_fp16)[name = tensor<string, []>("input_1155_cast_fp16")];
+            tensor<int32, [1]> input_1157_axes_0 = const()[name = tensor<string, []>("input_1157_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(448027456)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(448029568)))];
+            tensor<fp16, [1, 7, 1024]> input_1157_cast_fp16 = layer_norm(axes = input_1157_axes_0, beta = encoder_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_feed_forward2_weight_to_fp16, x = input_1155_cast_fp16)[name = tensor<string, []>("input_1157_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_21_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(448031680)))];
+            tensor<fp16, [4096]> encoder_layers_21_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456420352)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_197_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1161_cast_fp16 = silu(x = linear_197_cast_fp16)[name = tensor<string, []>("input_1161_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_21_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456428608)))];
+            tensor<fp16, [1024]> encoder_layers_21_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(464817280)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_198_cast_fp16")];
+            tensor<fp16, []> var_5104_to_fp16 = const()[name = tensor<string, []>("op_5104_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_5105_cast_fp16 = mul(x = linear_198_cast_fp16, y = var_5104_to_fp16)[name = tensor<string, []>("op_5105_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1167_cast_fp16 = add(x = input_1155_cast_fp16, y = var_5105_cast_fp16)[name = tensor<string, []>("input_1167_cast_fp16")];
+            tensor<int32, [1]> input_1169_axes_0 = const()[name = tensor<string, []>("input_1169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(464819392)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_21_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(464821504)))];
+            tensor<fp16, [1, 7, 1024]> input_1169_cast_fp16 = layer_norm(axes = input_1169_axes_0, beta = encoder_layers_21_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_out_weight_to_fp16, x = input_1167_cast_fp16)[name = tensor<string, []>("input_1169_cast_fp16")];
+            tensor<int32, [4]> cache_89_begin_0 = const()[name = tensor<string, []>("cache_89_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> cache_89_end_0 = const()[name = tensor<string, []>("cache_89_end_0"), val = tensor<int32, [4]>([23, 1, 42, 1024])];
+            tensor<bool, [4]> cache_89_end_mask_0 = const()[name = tensor<string, []>("cache_89_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_89_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_89_cast_fp16")];
+            tensor<int32, [4]> cache_91_begin_0 = const()[name = tensor<string, []>("cache_91_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> cache_91_end_0 = const()[name = tensor<string, []>("cache_91_end_0"), val = tensor<int32, [4]>([23, 1, 1024, 8])];
+            tensor<bool, [4]> cache_91_end_mask_0 = const()[name = tensor<string, []>("cache_91_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_91_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_91_cast_fp16")];
+            tensor<int32, [1]> input_1171_axes_0 = const()[name = tensor<string, []>("input_1171_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(464823616)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(464825728)))];
+            tensor<fp16, [1, 7, 1024]> input_1171_cast_fp16 = layer_norm(axes = input_1171_axes_0, beta = encoder_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_feed_forward1_weight_to_fp16, x = input_1169_cast_fp16)[name = tensor<string, []>("input_1171_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_22_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(464827840)))];
+            tensor<fp16, [4096]> encoder_layers_22_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(473216512)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_199_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1175_cast_fp16 = silu(x = linear_199_cast_fp16)[name = tensor<string, []>("input_1175_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_22_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(473224768)))];
+            tensor<fp16, [1024]> encoder_layers_22_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481613440)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_200_cast_fp16")];
+            tensor<fp16, []> var_5141_to_fp16 = const()[name = tensor<string, []>("op_5141_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_5142_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_5141_to_fp16)[name = tensor<string, []>("op_5142_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1181_cast_fp16 = add(x = input_1169_cast_fp16, y = var_5142_cast_fp16)[name = tensor<string, []>("input_1181_cast_fp16")];
+            tensor<int32, [1]> key_45_axes_0 = const()[name = tensor<string, []>("key_45_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481615552)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481617664)))];
+            tensor<fp16, [1, 7, 1024]> key_45_cast_fp16 = layer_norm(axes = key_45_axes_0, beta = encoder_layers_22_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_self_att_weight_to_fp16, x = input_1181_cast_fp16)[name = tensor<string, []>("key_45_cast_fp16")];
+            tensor<bool, []> input_1183_interleave_0 = const()[name = tensor<string, []>("input_1183_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_1183_cast_fp16 = concat(axis = var_68, interleave = input_1183_interleave_0, values = (cache_89_cast_fp16, key_45_cast_fp16))[name = tensor<string, []>("input_1183_cast_fp16")];
+            tensor<int32, [3]> var_5164_begin_0 = const()[name = tensor<string, []>("op_5164_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_5164_end_0 = const()[name = tensor<string, []>("op_5164_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_5164_end_mask_0 = const()[name = tensor<string, []>("op_5164_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_5164_cast_fp16 = slice_by_index(begin = var_5164_begin_0, end = var_5164_end_0, end_mask = var_5164_end_mask_0, x = cache_89_cast_fp16)[name = tensor<string, []>("op_5164_cast_fp16")];
+            tensor<bool, []> var_5170_interleave_0 = const()[name = tensor<string, []>("op_5170_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> var_5170_cast_fp16 = concat(axis = var_68, interleave = var_5170_interleave_0, values = (var_5164_cast_fp16, key_45_cast_fp16))[name = tensor<string, []>("op_5170_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481619776)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483716992)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_201_cast_fp16")];
+            tensor<int32, [4]> var_5175 = const()[name = tensor<string, []>("op_5175"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_133_cast_fp16 = reshape(shape = var_5175, x = linear_201_cast_fp16)[name = tensor<string, []>("q_133_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483719104)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(485816320)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_202_cast_fp16")];
+            tensor<int32, [4]> var_5180 = const()[name = tensor<string, []>("op_5180"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_89_cast_fp16 = reshape(shape = var_5180, x = linear_202_cast_fp16)[name = tensor<string, []>("k_89_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(485818432)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(487915648)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_203_cast_fp16")];
+            tensor<int32, [4]> var_5185 = const()[name = tensor<string, []>("op_5185"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_45_cast_fp16 = reshape(shape = var_5185, x = linear_203_cast_fp16)[name = tensor<string, []>("v_45_cast_fp16")];
+            tensor<int32, [4]> value_53_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_22_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(487917760)))];
+            tensor<fp16, [1, 7, 8, 128]> var_5198_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_5198_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_22_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(487919872)))];
+            tensor<fp16, [1, 7, 8, 128]> var_5200_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_5200_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_45_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_579_transpose_x_0 = const()[name = tensor<string, []>("x_579_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_579_transpose_y_0 = const()[name = tensor<string, []>("x_579_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_5202_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_5202_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(487921984))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(488021568))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(488021376)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_45_cast_fp16 = transpose(perm = q_with_bias_v_45_perm_0, x = var_5200_cast_fp16)[name = tensor<string, []>("transpose_164")];
+            tensor<fp16, [1, 8, 7, 97]> 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_5202_to_fp16_quantized)[name = tensor<string, []>("x_579_cast_fp16")];
+            tensor<int32, [8]> x_581_pad_0 = const()[name = tensor<string, []>("x_581_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_581_mode_0 = const()[name = tensor<string, []>("x_581_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_365_to_fp16 = const()[name = tensor<string, []>("const_365_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_581_cast_fp16")];
+            tensor<int32, [4]> var_5210 = const()[name = tensor<string, []>("op_5210"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_583_cast_fp16 = reshape(shape = var_5210, x = x_581_cast_fp16)[name = tensor<string, []>("x_583_cast_fp16")];
+            tensor<int32, [4]> var_5214_begin_0 = const()[name = tensor<string, []>("op_5214_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5214_end_0 = const()[name = tensor<string, []>("op_5214_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_5214_end_mask_0 = const()[name = tensor<string, []>("op_5214_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_5214_cast_fp16 = slice_by_index(begin = var_5214_begin_0, end = var_5214_end_0, end_mask = var_5214_end_mask_0, x = x_583_cast_fp16)[name = tensor<string, []>("op_5214_cast_fp16")];
+            tensor<int32, [4]> var_5215 = const()[name = tensor<string, []>("op_5215"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_89_cast_fp16 = reshape(shape = var_5215, x = var_5214_cast_fp16)[name = tensor<string, []>("matrix_bd_89_cast_fp16")];
+            tensor<bool, []> matrix_ac_45_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_45_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_45_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_45_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_140_perm_0 = const()[name = tensor<string, []>("transpose_140_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_141_perm_0 = const()[name = tensor<string, []>("transpose_141_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89_cast_fp16)[name = tensor<string, []>("transpose_162")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_140 = transpose(perm = transpose_140_perm_0, x = var_5198_cast_fp16)[name = tensor<string, []>("transpose_163")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_45_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_91_begin_0 = const()[name = tensor<string, []>("matrix_bd_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_91_end_0 = const()[name = tensor<string, []>("matrix_bd_91_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_91_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_91_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_5224_cast_fp16 = add(x = matrix_ac_45_cast_fp16, y = matrix_bd_91_cast_fp16)[name = tensor<string, []>("op_5224_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_89_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_89_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_89_cast_fp16 = mul(x = var_5224_cast_fp16, y = _inversed_scores_89_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_89_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_91_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_89_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_91_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_5230_cast_fp16 = softmax(axis = var_59, x = scores_91_cast_fp16)[name = tensor<string, []>("op_5230_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_1185_cast_fp16 = select(a = var_44_to_fp16, b = var_5230_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_1185_cast_fp16")];
+            tensor<bool, []> x_585_transpose_x_0 = const()[name = tensor<string, []>("x_585_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_585_transpose_y_0 = const()[name = tensor<string, []>("x_585_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = v_45_cast_fp16)[name = tensor<string, []>("transpose_161")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_585_cast_fp16")];
+            tensor<int32, [4]> var_5234_perm_0 = const()[name = tensor<string, []>("op_5234_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5235 = const()[name = tensor<string, []>("op_5235"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_5234_cast_fp16 = transpose(perm = var_5234_perm_0, x = x_585_cast_fp16)[name = tensor<string, []>("transpose_160")];
+            tensor<fp16, [1, 7, 1024]> input_1187_cast_fp16 = reshape(shape = var_5235, x = var_5234_cast_fp16)[name = tensor<string, []>("input_1187_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(488021888)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490119104)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_205_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1191_cast_fp16 = add(x = input_1181_cast_fp16, y = linear_205_cast_fp16)[name = tensor<string, []>("input_1191_cast_fp16")];
+            tensor<int32, [1]> x_589_axes_0 = const()[name = tensor<string, []>("x_589_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490121216)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490123328)))];
+            tensor<fp16, [1, 7, 1024]> x_589_cast_fp16 = layer_norm(axes = x_589_axes_0, beta = encoder_layers_22_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_conv_weight_to_fp16, x = input_1191_cast_fp16)[name = tensor<string, []>("x_589_cast_fp16")];
+            tensor<int32, [3]> input_1193_perm_0 = const()[name = tensor<string, []>("input_1193_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_1195_pad_type_0 = const()[name = tensor<string, []>("input_1195_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_1195_strides_0 = const()[name = tensor<string, []>("input_1195_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1195_pad_0 = const()[name = tensor<string, []>("input_1195_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1195_dilations_0 = const()[name = tensor<string, []>("input_1195_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_1195_groups_0 = const()[name = tensor<string, []>("input_1195_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490125440))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492222656))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_1193_cast_fp16 = transpose(perm = input_1193_perm_0, x = x_589_cast_fp16)[name = tensor<string, []>("transpose_159")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_1195_cast_fp16")];
+            tensor<int32, []> x_591_split_num_splits_0 = const()[name = tensor<string, []>("x_591_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_591_split_axis_0 = const()[name = tensor<string, []>("x_591_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_591_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_591_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_591_split_1_sigmoid_cast_fp16 = sigmoid(x = x_591_split_cast_fp16_1)[name = tensor<string, []>("x_591_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_591_cast_fp16 = mul(x = x_591_split_cast_fp16_0, y = x_591_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_591_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_1197_cast_fp16 = select(a = var_44_to_fp16, b = x_591_cast_fp16, cond = var_575)[name = tensor<string, []>("input_1197_cast_fp16")];
+            tensor<bool, []> new_x_91_interleave_0 = const()[name = tensor<string, []>("new_x_91_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_91_cast_fp16 = concat(axis = var_59, interleave = new_x_91_interleave_0, values = (cache_91_cast_fp16, input_1197_cast_fp16))[name = tensor<string, []>("new_x_91_cast_fp16")];
+            tensor<int32, [3]> var_5274_begin_0 = const()[name = tensor<string, []>("op_5274_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_5274_end_0 = const()[name = tensor<string, []>("op_5274_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_5274_end_mask_0 = const()[name = tensor<string, []>("op_5274_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_5274_cast_fp16 = slice_by_index(begin = var_5274_begin_0, end = var_5274_end_0, end_mask = var_5274_end_mask_0, x = new_x_91_cast_fp16)[name = tensor<string, []>("op_5274_cast_fp16")];
+            tensor<string, []> x_593_pad_type_0 = const()[name = tensor<string, []>("x_593_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_593_groups_0 = const()[name = tensor<string, []>("x_593_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_593_strides_0 = const()[name = tensor<string, []>("x_593_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_593_pad_0 = const()[name = tensor<string, []>("x_593_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_593_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492226816))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492236096))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_593_cast_fp16")];
+            tensor<int32, [3]> input_1199_perm_0 = const()[name = tensor<string, []>("input_1199_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_595_axes_0 = const()[name = tensor<string, []>("x_595_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492238208)))];
+            tensor<fp16, [1024]> encoder_layers_22_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492240320)))];
+            tensor<fp16, [1, 7, 1024]> input_1199_cast_fp16 = transpose(perm = input_1199_perm_0, x = x_593_cast_fp16)[name = tensor<string, []>("transpose_158")];
+            tensor<fp16, [1, 7, 1024]> x_595_cast_fp16 = layer_norm(axes = x_595_axes_0, beta = encoder_layers_22_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_conv_batch_norm_weight_to_fp16, x = input_1199_cast_fp16)[name = tensor<string, []>("x_595_cast_fp16")];
+            tensor<int32, [3]> input_1201_perm_0 = const()[name = tensor<string, []>("input_1201_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_1201_cast_fp16 = transpose(perm = input_1201_perm_0, x = x_595_cast_fp16)[name = tensor<string, []>("transpose_157")];
+            tensor<fp16, [1, 1024, 7]> input_1203_cast_fp16 = silu(x = input_1201_cast_fp16)[name = tensor<string, []>("input_1203_cast_fp16")];
+            tensor<string, []> x_597_pad_type_0 = const()[name = tensor<string, []>("x_597_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_597_strides_0 = const()[name = tensor<string, []>("x_597_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_597_pad_0 = const()[name = tensor<string, []>("x_597_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_597_dilations_0 = const()[name = tensor<string, []>("x_597_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_597_groups_0 = const()[name = tensor<string, []>("x_597_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492242432))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493291072))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_597_cast_fp16")];
+            tensor<int32, [3]> input_1205_perm_0 = const()[name = tensor<string, []>("input_1205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_1205_cast_fp16 = transpose(perm = input_1205_perm_0, x = x_597_cast_fp16)[name = tensor<string, []>("transpose_156")];
+            tensor<fp16, [1, 7, 1024]> input_1207_cast_fp16 = add(x = input_1191_cast_fp16, y = input_1205_cast_fp16)[name = tensor<string, []>("input_1207_cast_fp16")];
+            tensor<int32, [1]> input_1209_axes_0 = const()[name = tensor<string, []>("input_1209_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493293184)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493295296)))];
+            tensor<fp16, [1, 7, 1024]> input_1209_cast_fp16 = layer_norm(axes = input_1209_axes_0, beta = encoder_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_feed_forward2_weight_to_fp16, x = input_1207_cast_fp16)[name = tensor<string, []>("input_1209_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_22_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493297408)))];
+            tensor<fp16, [4096]> encoder_layers_22_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(501686080)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_206_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1213_cast_fp16 = silu(x = linear_206_cast_fp16)[name = tensor<string, []>("input_1213_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_22_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(501694336)))];
+            tensor<fp16, [1024]> encoder_layers_22_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(510083008)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_207_cast_fp16")];
+            tensor<fp16, []> var_5317_to_fp16 = const()[name = tensor<string, []>("op_5317_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_5318_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_5317_to_fp16)[name = tensor<string, []>("op_5318_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1219_cast_fp16 = add(x = input_1207_cast_fp16, y = var_5318_cast_fp16)[name = tensor<string, []>("input_1219_cast_fp16")];
+            tensor<int32, [1]> input_1221_axes_0 = const()[name = tensor<string, []>("input_1221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(510085120)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_22_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(510087232)))];
+            tensor<fp16, [1, 7, 1024]> input_1221_cast_fp16 = layer_norm(axes = input_1221_axes_0, beta = encoder_layers_22_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_out_weight_to_fp16, x = input_1219_cast_fp16)[name = tensor<string, []>("input_1221_cast_fp16")];
+            tensor<int32, [4]> cache_93_begin_0 = const()[name = tensor<string, []>("cache_93_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> cache_93_end_0 = const()[name = tensor<string, []>("cache_93_end_0"), val = tensor<int32, [4]>([24, 1, 42, 1024])];
+            tensor<bool, [4]> cache_93_end_mask_0 = const()[name = tensor<string, []>("cache_93_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_93_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_93_cast_fp16")];
+            tensor<int32, [4]> cache_begin_0 = const()[name = tensor<string, []>("cache_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> cache_end_0 = const()[name = tensor<string, []>("cache_end_0"), val = tensor<int32, [4]>([24, 1, 1024, 8])];
+            tensor<bool, [4]> cache_end_mask_0 = const()[name = tensor<string, []>("cache_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_squeeze_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_cast_fp16")];
+            tensor<int32, [1]> input_1223_axes_0 = const()[name = tensor<string, []>("input_1223_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(510089344)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(510091456)))];
+            tensor<fp16, [1, 7, 1024]> input_1223_cast_fp16 = layer_norm(axes = input_1223_axes_0, beta = encoder_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_feed_forward1_weight_to_fp16, x = input_1221_cast_fp16)[name = tensor<string, []>("input_1223_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_23_feed_forward1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(510093568)))];
+            tensor<fp16, [4096]> encoder_layers_23_feed_forward1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(518482240)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_208_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1227_cast_fp16 = silu(x = linear_208_cast_fp16)[name = tensor<string, []>("input_1227_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_23_feed_forward1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(518490496)))];
+            tensor<fp16, [1024]> encoder_layers_23_feed_forward1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526879168)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_209_cast_fp16")];
+            tensor<fp16, []> var_5354_to_fp16 = const()[name = tensor<string, []>("op_5354_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_5355_cast_fp16 = mul(x = linear_209_cast_fp16, y = var_5354_to_fp16)[name = tensor<string, []>("op_5355_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1233_cast_fp16 = add(x = input_1221_cast_fp16, y = var_5355_cast_fp16)[name = tensor<string, []>("input_1233_cast_fp16")];
+            tensor<int32, [1]> key_axes_0 = const()[name = tensor<string, []>("key_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526881280)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526883392)))];
+            tensor<fp16, [1, 7, 1024]> key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_23_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_self_att_weight_to_fp16, x = input_1233_cast_fp16)[name = tensor<string, []>("key_cast_fp16")];
+            tensor<bool, []> input_1235_interleave_0 = const()[name = tensor<string, []>("input_1235_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 49, 1024]> input_1235_cast_fp16 = concat(axis = var_68, interleave = input_1235_interleave_0, values = (cache_93_cast_fp16, key_cast_fp16))[name = tensor<string, []>("input_1235_cast_fp16")];
+            tensor<int32, [3]> var_5377_begin_0 = const()[name = tensor<string, []>("op_5377_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_5377_end_0 = const()[name = tensor<string, []>("op_5377_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_5377_end_mask_0 = const()[name = tensor<string, []>("op_5377_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_5377_cast_fp16 = slice_by_index(begin = var_5377_begin_0, end = var_5377_end_0, end_mask = var_5377_end_mask_0, x = cache_93_cast_fp16)[name = tensor<string, []>("op_5377_cast_fp16")];
+            tensor<bool, []> cache_last_channel_cur_interleave_0 = const()[name = tensor<string, []>("cache_last_channel_cur_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 42, 1024]> cache_last_channel_cur_cast_fp16 = concat(axis = var_68, interleave = cache_last_channel_cur_interleave_0, values = (var_5377_cast_fp16, key_cast_fp16))[name = tensor<string, []>("cache_last_channel_cur_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_q_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526885504)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_q_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(528982720)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_210_cast_fp16")];
+            tensor<int32, [4]> var_5388 = const()[name = tensor<string, []>("op_5388"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_139_cast_fp16 = reshape(shape = var_5388, x = linear_210_cast_fp16)[name = tensor<string, []>("q_139_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_k_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(528984832)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_k_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(531082048)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_211_cast_fp16")];
+            tensor<int32, [4]> var_5393 = const()[name = tensor<string, []>("op_5393"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_93_cast_fp16 = reshape(shape = var_5393, x = linear_211_cast_fp16)[name = tensor<string, []>("k_93_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_v_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(531084160)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_v_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533181376)))];
+            tensor<fp16, [1, 49, 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 = tensor<string, []>("linear_212_cast_fp16")];
+            tensor<int32, [4]> var_5398 = const()[name = tensor<string, []>("op_5398"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_cast_fp16 = reshape(shape = var_5398, x = linear_212_cast_fp16)[name = tensor<string, []>("v_cast_fp16")];
+            tensor<int32, [4]> value_perm_0 = const()[name = tensor<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 = tensor<string, []>("encoder_layers_23_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533183488)))];
+            tensor<fp16, [1, 7, 8, 128]> var_5411_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_5411_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_23_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533185600)))];
+            tensor<fp16, [1, 7, 8, 128]> var_5413_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_5413_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<bool, []> x_605_transpose_x_0 = const()[name = tensor<string, []>("x_605_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_605_transpose_y_0 = const()[name = tensor<string, []>("x_605_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_5415_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_5415_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533187712))), scale = tensor<fp16, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533287296))), zero_point = tensor<int8, [97]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533287104)))];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_5413_cast_fp16)[name = tensor<string, []>("transpose_155")];
+            tensor<fp16, [1, 8, 7, 97]> 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_5415_to_fp16_quantized)[name = tensor<string, []>("x_605_cast_fp16")];
+            tensor<int32, [8]> x_607_pad_0 = const()[name = tensor<string, []>("x_607_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_607_mode_0 = const()[name = tensor<string, []>("x_607_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_378_to_fp16 = const()[name = tensor<string, []>("const_378_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> 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 = tensor<string, []>("x_607_cast_fp16")];
+            tensor<int32, [4]> var_5423 = const()[name = tensor<string, []>("op_5423"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_609_cast_fp16 = reshape(shape = var_5423, x = x_607_cast_fp16)[name = tensor<string, []>("x_609_cast_fp16")];
+            tensor<int32, [4]> var_5427_begin_0 = const()[name = tensor<string, []>("op_5427_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5427_end_0 = const()[name = tensor<string, []>("op_5427_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_5427_end_mask_0 = const()[name = tensor<string, []>("op_5427_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_5427_cast_fp16 = slice_by_index(begin = var_5427_begin_0, end = var_5427_end_0, end_mask = var_5427_end_mask_0, x = x_609_cast_fp16)[name = tensor<string, []>("op_5427_cast_fp16")];
+            tensor<int32, [4]> var_5428 = const()[name = tensor<string, []>("op_5428"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_93_cast_fp16 = reshape(shape = var_5428, x = var_5427_cast_fp16)[name = tensor<string, []>("matrix_bd_93_cast_fp16")];
+            tensor<bool, []> matrix_ac_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_142_perm_0 = const()[name = tensor<string, []>("transpose_142_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_143_perm_0 = const()[name = tensor<string, []>("transpose_143_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93_cast_fp16)[name = tensor<string, []>("transpose_153")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_142 = transpose(perm = transpose_142_perm_0, x = var_5411_cast_fp16)[name = tensor<string, []>("transpose_154")];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_ac_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_begin_0 = const()[name = tensor<string, []>("matrix_bd_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_end_0 = const()[name = tensor<string, []>("matrix_bd_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> 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 = tensor<string, []>("matrix_bd_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_5437_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = tensor<string, []>("op_5437_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_93_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_93_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_93_cast_fp16 = mul(x = var_5437_cast_fp16, y = _inversed_scores_93_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_93_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_93_cast_fp16, cond = mask_11)[name = tensor<string, []>("scores_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_5443_cast_fp16 = softmax(axis = var_59, x = scores_cast_fp16)[name = tensor<string, []>("op_5443_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_1237_cast_fp16 = select(a = var_44_to_fp16, b = var_5443_cast_fp16, cond = mask_11)[name = tensor<string, []>("input_1237_cast_fp16")];
+            tensor<bool, []> x_611_transpose_x_0 = const()[name = tensor<string, []>("x_611_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_611_transpose_y_0 = const()[name = tensor<string, []>("x_611_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = tensor<string, []>("transpose_152")];
+            tensor<fp16, [1, 8, 7, 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 = tensor<string, []>("x_611_cast_fp16")];
+            tensor<int32, [4]> var_5447_perm_0 = const()[name = tensor<string, []>("op_5447_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5448 = const()[name = tensor<string, []>("op_5448"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_5447_cast_fp16 = transpose(perm = var_5447_perm_0, x = x_611_cast_fp16)[name = tensor<string, []>("transpose_151")];
+            tensor<fp16, [1, 7, 1024]> input_1239_cast_fp16 = reshape(shape = var_5448, x = var_5447_cast_fp16)[name = tensor<string, []>("input_1239_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533287616)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(535384832)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_214_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1243_cast_fp16 = add(x = input_1233_cast_fp16, y = linear_214_cast_fp16)[name = tensor<string, []>("input_1243_cast_fp16")];
+            tensor<int32, [1]> x_615_axes_0 = const()[name = tensor<string, []>("x_615_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(535386944)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(535389056)))];
+            tensor<fp16, [1, 7, 1024]> x_615_cast_fp16 = layer_norm(axes = x_615_axes_0, beta = encoder_layers_23_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_conv_weight_to_fp16, x = input_1243_cast_fp16)[name = tensor<string, []>("x_615_cast_fp16")];
+            tensor<int32, [3]> input_1245_perm_0 = const()[name = tensor<string, []>("input_1245_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_1247_pad_type_0 = const()[name = tensor<string, []>("input_1247_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_1247_strides_0 = const()[name = tensor<string, []>("input_1247_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1247_pad_0 = const()[name = tensor<string, []>("input_1247_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1247_dilations_0 = const()[name = tensor<string, []>("input_1247_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_1247_groups_0 = const()[name = tensor<string, []>("input_1247_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(535391168))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(537488384))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19445568)))];
+            tensor<fp16, [1, 1024, 7]> input_1245_cast_fp16 = transpose(perm = input_1245_perm_0, x = x_615_cast_fp16)[name = tensor<string, []>("transpose_150")];
+            tensor<fp16, [1, 2048, 7]> 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 = tensor<string, []>("input_1247_cast_fp16")];
+            tensor<int32, []> x_617_split_num_splits_0 = const()[name = tensor<string, []>("x_617_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_617_split_axis_0 = const()[name = tensor<string, []>("x_617_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 7]> x_617_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_617_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_617_split_1_sigmoid_cast_fp16 = sigmoid(x = x_617_split_cast_fp16_1)[name = tensor<string, []>("x_617_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_617_cast_fp16 = mul(x = x_617_split_cast_fp16_0, y = x_617_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("x_617_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_1249_cast_fp16 = select(a = var_44_to_fp16, b = x_617_cast_fp16, cond = var_575)[name = tensor<string, []>("input_1249_cast_fp16")];
+            tensor<bool, []> new_x_interleave_0 = const()[name = tensor<string, []>("new_x_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_cast_fp16 = concat(axis = var_59, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_1249_cast_fp16))[name = tensor<string, []>("new_x_cast_fp16")];
+            tensor<int32, [3]> cache_last_time_cur_begin_0 = const()[name = tensor<string, []>("cache_last_time_cur_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> cache_last_time_cur_end_0 = const()[name = tensor<string, []>("cache_last_time_cur_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> cache_last_time_cur_end_mask_0 = const()[name = tensor<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 = tensor<string, []>("cache_last_time_cur_cast_fp16")];
+            tensor<string, []> x_619_pad_type_0 = const()[name = tensor<string, []>("x_619_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_619_groups_0 = const()[name = tensor<string, []>("x_619_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_619_strides_0 = const()[name = tensor<string, []>("x_619_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_619_pad_0 = const()[name = tensor<string, []>("x_619_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_619_dilations_0 = const()[name = tensor<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_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(537492544))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(537501824))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_619_cast_fp16")];
+            tensor<int32, [3]> input_1251_perm_0 = const()[name = tensor<string, []>("input_1251_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_621_axes_0 = const()[name = tensor<string, []>("x_621_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_conv_batch_norm_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(537503936)))];
+            tensor<fp16, [1024]> encoder_layers_23_conv_batch_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(537506048)))];
+            tensor<fp16, [1, 7, 1024]> input_1251_cast_fp16 = transpose(perm = input_1251_perm_0, x = x_619_cast_fp16)[name = tensor<string, []>("transpose_149")];
+            tensor<fp16, [1, 7, 1024]> x_621_cast_fp16 = layer_norm(axes = x_621_axes_0, beta = encoder_layers_23_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_conv_batch_norm_weight_to_fp16, x = input_1251_cast_fp16)[name = tensor<string, []>("x_621_cast_fp16")];
+            tensor<int32, [3]> input_1253_perm_0 = const()[name = tensor<string, []>("input_1253_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_1253_cast_fp16 = transpose(perm = input_1253_perm_0, x = x_621_cast_fp16)[name = tensor<string, []>("transpose_148")];
+            tensor<fp16, [1, 1024, 7]> input_1255_cast_fp16 = silu(x = input_1253_cast_fp16)[name = tensor<string, []>("input_1255_cast_fp16")];
+            tensor<string, []> x_623_pad_type_0 = const()[name = tensor<string, []>("x_623_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_623_strides_0 = const()[name = tensor<string, []>("x_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_623_pad_0 = const()[name = tensor<string, []>("x_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_623_dilations_0 = const()[name = tensor<string, []>("x_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_623_groups_0 = const()[name = tensor<string, []>("x_623_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(537508160))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538556800))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4600960)))];
+            tensor<fp16, [1, 1024, 7]> 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 = tensor<string, []>("x_623_cast_fp16")];
+            tensor<int32, [3]> input_1257_perm_0 = const()[name = tensor<string, []>("input_1257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_1257_cast_fp16 = transpose(perm = input_1257_perm_0, x = x_623_cast_fp16)[name = tensor<string, []>("transpose_147")];
+            tensor<fp16, [1, 7, 1024]> input_1259_cast_fp16 = add(x = input_1243_cast_fp16, y = input_1257_cast_fp16)[name = tensor<string, []>("input_1259_cast_fp16")];
+            tensor<int32, [1]> input_1261_axes_0 = const()[name = tensor<string, []>("input_1261_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538558912)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538561024)))];
+            tensor<fp16, [1, 7, 1024]> input_1261_cast_fp16 = layer_norm(axes = input_1261_axes_0, beta = encoder_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_feed_forward2_weight_to_fp16, x = input_1259_cast_fp16)[name = tensor<string, []>("input_1261_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_23_feed_forward2_linear1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538563136)))];
+            tensor<fp16, [4096]> encoder_layers_23_feed_forward2_linear1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(546951808)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_215_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1265_cast_fp16 = silu(x = linear_215_cast_fp16)[name = tensor<string, []>("input_1265_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_23_feed_forward2_linear2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(546960064)))];
+            tensor<fp16, [1024]> encoder_layers_23_feed_forward2_linear2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555348736)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_216_cast_fp16")];
+            tensor<fp16, []> var_5530_to_fp16 = const()[name = tensor<string, []>("op_5530_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_5531_cast_fp16 = mul(x = linear_216_cast_fp16, y = var_5530_to_fp16)[name = tensor<string, []>("op_5531_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1271_cast_fp16 = add(x = input_1259_cast_fp16, y = var_5531_cast_fp16)[name = tensor<string, []>("input_1271_cast_fp16")];
+            tensor<int32, [1]> audio_signal_axes_0 = const()[name = tensor<string, []>("audio_signal_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555350848)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_23_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555352960)))];
+            tensor<fp16, [1, 7, 1024]> audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_23_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_out_weight_to_fp16, x = input_1271_cast_fp16)[name = tensor<string, []>("audio_signal_cast_fp16")];
+            tensor<int32, []> obj_5_axis_0 = const()[name = tensor<string, []>("obj_5_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [24, 1, 42, 1024]> obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_484_cast_fp16, var_697_cast_fp16, var_910_cast_fp16, var_1123_cast_fp16, var_1336_cast_fp16, var_1549_cast_fp16, var_1762_cast_fp16, var_1975_cast_fp16, var_2188_cast_fp16, var_2401_cast_fp16, var_2614_cast_fp16, var_2827_cast_fp16, var_3040_cast_fp16, var_3253_cast_fp16, var_3466_cast_fp16, var_3679_cast_fp16, var_3892_cast_fp16, var_4105_cast_fp16, var_4318_cast_fp16, var_4531_cast_fp16, var_4744_cast_fp16, var_4957_cast_fp16, var_5170_cast_fp16, cache_last_channel_cur_cast_fp16))[name = tensor<string, []>("obj_5_cast_fp16")];
+            tensor<int32, []> obj_7_axis_0 = const()[name = tensor<string, []>("obj_7_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [24, 1, 1024, 8]> obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_588_cast_fp16, var_801_cast_fp16, var_1014_cast_fp16, var_1227_cast_fp16, var_1440_cast_fp16, var_1653_cast_fp16, var_1866_cast_fp16, var_2079_cast_fp16, var_2292_cast_fp16, var_2505_cast_fp16, var_2718_cast_fp16, var_2931_cast_fp16, var_3144_cast_fp16, var_3357_cast_fp16, var_3570_cast_fp16, var_3783_cast_fp16, var_3996_cast_fp16, var_4209_cast_fp16, var_4422_cast_fp16, var_4635_cast_fp16, var_4848_cast_fp16, var_5061_cast_fp16, var_5274_cast_fp16, cache_last_time_cur_cast_fp16))[name = tensor<string, []>("obj_7_cast_fp16")];
+            tensor<int32, [1]> var_5547 = add(x = cache_len, y = max_audio_length_1)[name = tensor<string, []>("op_5547")];
+            tensor<string, []> var_5547_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_5547_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> const_384_to_fp16 = const()[name = tensor<string, []>("const_384_to_fp16"), val = tensor<fp16, []>(-inf)];
+            tensor<fp16, []> var_49_promoted_to_fp16 = const()[name = tensor<string, []>("op_49_promoted_to_fp16"), val = tensor<fp16, []>(0x1.5p+5)];
+            tensor<fp16, [1]> var_5547_to_fp16 = cast(dtype = var_5547_promoted_to_fp16_dtype_0, x = var_5547)[name = tensor<string, []>("cast_6")];
+            tensor<fp16, [1]> clip_1_cast_fp16 = clip(alpha = const_384_to_fp16, beta = var_49_promoted_to_fp16, x = var_5547_to_fp16)[name = tensor<string, []>("clip_1_cast_fp16")];
+            tensor<int32, []> var_5584 = const()[name = tensor<string, []>("op_5584"), val = tensor<int32, []>(0)];
+            tensor<int32, []> one_hot_1_batch_dims_0 = const()[name = tensor<string, []>("one_hot_1_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> one_hot_1_validate_indices_0 = const()[name = tensor<string, []>("one_hot_1_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [128, 128]> to_onehot_identity_table_to_fp16 = const()[name = tensor<string, []>("to_onehot_identity_table_to_fp16"), val = tensor<fp16, [128, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555355072)))];
+            tensor<string, []> prompt_id_to_int16_dtype_0 = const()[name = tensor<string, []>("prompt_id_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1]> prompt_id_to_int16 = cast(dtype = prompt_id_to_int16_dtype_0, x = prompt_id)[name = tensor<string, []>("cast_5")];
+            tensor<fp16, [1, 128]> one_hot_1_cast_fp16_cast_uint16 = gather(axis = var_5584, batch_dims = one_hot_1_batch_dims_0, indices = prompt_id_to_int16, validate_indices = one_hot_1_validate_indices_0, x = to_onehot_identity_table_to_fp16)[name = tensor<string, []>("one_hot_1_cast_fp16_cast_uint16")];
+            tensor<int32, [1]> var_5593_axes_0 = const()[name = tensor<string, []>("op_5593_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 128]> var_5593_cast_fp16 = expand_dims(axes = var_5593_axes_0, x = one_hot_1_cast_fp16_cast_uint16)[name = tensor<string, []>("op_5593_cast_fp16")];
+            tensor<int32, [3]> one_hot_reps_0 = const()[name = tensor<string, []>("one_hot_reps_0"), val = tensor<int32, [3]>([1, 7, 1])];
+            tensor<fp16, [1, 7, 128]> one_hot_cast_fp16 = tile(reps = one_hot_reps_0, x = var_5593_cast_fp16)[name = tensor<string, []>("one_hot_cast_fp16")];
+            tensor<int32, []> var_5602 = const()[name = tensor<string, []>("op_5602"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> input_1273_interleave_0 = const()[name = tensor<string, []>("input_1273_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 7, 1152]> input_1273_cast_fp16 = concat(axis = var_5602, interleave = input_1273_interleave_0, values = (audio_signal_cast_fp16, one_hot_cast_fp16))[name = tensor<string, []>("input_1273_cast_fp16")];
+            tensor<fp16, [2048, 1152]> prompt_kernel_0_weight_to_fp16 = const()[name = tensor<string, []>("prompt_kernel_0_weight_to_fp16"), val = tensor<fp16, [2048, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555387904)))];
+            tensor<fp16, [2048]> prompt_kernel_0_bias_to_fp16 = const()[name = tensor<string, []>("prompt_kernel_0_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(560106560)))];
+            tensor<fp16, [1, 7, 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 = tensor<string, []>("linear_217_cast_fp16")];
+            tensor<fp16, [1, 7, 2048]> input_cast_fp16 = relu(x = linear_217_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<fp16, [1024, 2048]> prompt_kernel_2_weight_to_fp16 = const()[name = tensor<string, []>("prompt_kernel_2_weight_to_fp16"), val = tensor<fp16, [1024, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(560110720)))];
+            tensor<fp16, [1024]> prompt_kernel_2_bias_to_fp16 = const()[name = tensor<string, []>("prompt_kernel_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(564305088)))];
+            tensor<fp16, [1, 7, 1024]> linear_218_cast_fp16 = linear(bias = prompt_kernel_2_bias_to_fp16, weight = prompt_kernel_2_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("linear_218_cast_fp16")];
+            tensor<int32, [3]> var_5615_perm_0 = const()[name = tensor<string, []>("op_5615_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> var_5615_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_5615_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<string, []> cast_246_dtype_0 = const()[name = tensor<string, []>("cast_246_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [4]> var_5623_perm_0 = const()[name = tensor<string, []>("op_5623_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<string, []> var_5623_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_5623_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [4]> var_5626_perm_0 = const()[name = tensor<string, []>("op_5626_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<string, []> var_5626_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_5626_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<string, []> cast_247_dtype_0 = const()[name = tensor<string, []>("cast_247_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1]> cache_len_out = cast(dtype = cast_247_dtype_0, x = clip_1_cast_fp16)[name = tensor<string, []>("cast_0")];
+            tensor<fp16, [1, 24, 1024, 8]> var_5626_cast_fp16 = transpose(perm = var_5626_perm_0, x = obj_7_cast_fp16)[name = tensor<string, []>("transpose_144")];
+            tensor<fp32, [1, 24, 1024, 8]> cache_time_out = cast(dtype = var_5626_cast_fp16_to_fp32_dtype_0, x = var_5626_cast_fp16)[name = tensor<string, []>("cast_1")];
+            tensor<fp16, [1, 24, 42, 1024]> var_5623_cast_fp16 = transpose(perm = var_5623_perm_0, x = obj_5_cast_fp16)[name = tensor<string, []>("transpose_145")];
+            tensor<fp32, [1, 24, 42, 1024]> cache_channel_out = cast(dtype = var_5623_cast_fp16_to_fp32_dtype_0, x = var_5623_cast_fp16)[name = tensor<string, []>("cast_2")];
+            tensor<int32, [1]> encoded_length = cast(dtype = cast_246_dtype_0, x = clip_0_cast_fp16)[name = tensor<string, []>("cast_3")];
+            tensor<fp16, [1, 1024, 7]> var_5615_cast_fp16 = transpose(perm = var_5615_perm_0, x = linear_218_cast_fp16)[name = tensor<string, []>("transpose_146")];
+            tensor<fp32, [1, 1024, 7]> encoded = cast(dtype = var_5615_cast_fp16_to_fp32_dtype_0, x = var_5615_cast_fp16)[name = tensor<string, []>("cast_4")];
+        } -> (encoded, encoded_length, cache_channel_out, cache_time_out, cache_len_out);
+}
\ No newline at end of file
diff --git a/latin/560ms/encoder.mlmodelc/weights/weight.bin b/latin/560ms/encoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..0c7ae013f7884e9ecf6c5a249b080f5919db36d1
--- /dev/null
+++ b/latin/560ms/encoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e994440ed6c936ac375d375eaab16eb445881ff2d9bc62384de991b411e234e
+size 564307200
diff --git a/latin/560ms/joint.mlmodelc/analytics/coremldata.bin b/latin/560ms/joint.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..55631c767bacfe5759097fcfa4ef6969a5223f75
--- /dev/null
+++ b/latin/560ms/joint.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:80eab40693c0a7d60ee85044af620fdc20147ce8902086c91656ad13d04bf063
+size 243
diff --git a/latin/560ms/joint.mlmodelc/coremldata.bin b/latin/560ms/joint.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..493b23c4bc88908e891d2bf67df5002ee524e21b
--- /dev/null
+++ b/latin/560ms/joint.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b638d3d0cdd52c133da7ece37df5aa815182d41328edeaa87dfde1c306c05089
+size 341
diff --git a/latin/560ms/joint.mlmodelc/model.mil b/latin/560ms/joint.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..58ce97d6dc855c13232cf61f81845dbca8e3759f
--- /dev/null
+++ b/latin/560ms/joint.mlmodelc/model.mil
@@ -0,0 +1,31 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<fp32, [1, 640, 1]> decoder, tensor<fp32, [1, 1024, 1]> encoder) {
+            tensor<int32, [3]> input_1_perm_0 = const()[name = tensor<string, []>("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> encoder_to_fp16_dtype_0 = const()[name = tensor<string, []>("encoder_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> decoder_to_fp16_dtype_0 = const()[name = tensor<string, []>("decoder_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [640, 1024]> module_enc_weight_to_fp16 = const()[name = tensor<string, []>("module_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [640]> module_enc_bias_to_fp16 = const()[name = tensor<string, []>("module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1310848)))];
+            tensor<fp16, [1, 1024, 1]> encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = tensor<string, []>("cast_2")];
+            tensor<fp16, [1, 1, 1024]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_to_fp16)[name = tensor<string, []>("transpose_1")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = module_enc_bias_to_fp16, weight = module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> module_pred_weight_to_fp16 = const()[name = tensor<string, []>("module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1312192)))];
+            tensor<fp16, [640]> module_pred_bias_to_fp16 = const()[name = tensor<string, []>("module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2131456)))];
+            tensor<fp16, [1, 640, 1]> decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = tensor<string, []>("cast_1")];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_to_fp16)[name = tensor<string, []>("transpose_0")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = module_pred_bias_to_fp16, weight = module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_23_axes_0 = const()[name = tensor<string, []>("op_23_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_23_cast_fp16")];
+            tensor<int32, [1]> var_25_axes_0 = const()[name = tensor<string, []>("op_25_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_25_cast_fp16 = expand_dims(axes = var_25_axes_0, x = linear_1_cast_fp16)[name = tensor<string, []>("op_25_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_25_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<fp16, [2829, 640]> module_joint_net_2_weight_to_fp16 = const()[name = tensor<string, []>("module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [2829, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2132800)))];
+            tensor<fp16, [2829]> module_joint_net_2_bias_to_fp16 = const()[name = tensor<string, []>("module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [2829]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5753984)))];
+            tensor<fp16, [1, 1, 1, 2829]> linear_2_cast_fp16 = linear(bias = module_joint_net_2_bias_to_fp16, weight = module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<string, []> linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("linear_2_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 1, 1, 2829]> logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = tensor<string, []>("cast_0")];
+        } -> (logits);
+}
\ No newline at end of file
diff --git a/latin/560ms/joint.mlmodelc/weights/weight.bin b/latin/560ms/joint.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..baa3ed0a821090eb9e42550135aba83fb8e89bbe
--- /dev/null
+++ b/latin/560ms/joint.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66b9d7405fad48fdd8413a7275c04d44933cb9beb80d318c158f80e52ff43fba
+size 5759706
diff --git a/latin/560ms/preprocessor.mlmodelc/analytics/coremldata.bin b/latin/560ms/preprocessor.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..0815b2c5d31e1776ba471c013d4e1d684cb49c52
--- /dev/null
+++ b/latin/560ms/preprocessor.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d628a4c3b5a7f15375b09ae18a50d608ac0cce12de0e6d0749d7df2e51101e1
+size 243
diff --git a/latin/560ms/preprocessor.mlmodelc/coremldata.bin b/latin/560ms/preprocessor.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..d0c1abe7c256ed101f990eec0a848677d014b255
--- /dev/null
+++ b/latin/560ms/preprocessor.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:47a3f47d3034f2a3232ff6f98fecb85d44a011e47ffffb78f24b45407f5a2e16
+size 371
diff --git a/latin/560ms/preprocessor.mlmodelc/model.mil b/latin/560ms/preprocessor.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..3dd53065828833a3eb30fdc575ba97353ee57d11
--- /dev/null
+++ b/latin/560ms/preprocessor.mlmodelc/model.mil
@@ -0,0 +1,122 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<fp32, [1, ?]> audio, tensor<int32, [1]> audio_length) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"audio", [1, 1]}}), ("RangeDims", {{"audio", [[1, 1], [1, 1280000]]}})))] {
+            tensor<int32, []> var_9 = const()[name = tensor<string, []>("op_9"), val = tensor<int32, []>(1)];
+            tensor<int32, []> var_10 = const()[name = tensor<string, []>("op_10"), val = tensor<int32, []>(160)];
+            tensor<int32, []> var_12 = const()[name = tensor<string, []>("op_12"), val = tensor<int32, []>(0)];
+            tensor<int32, []> var_33 = const()[name = tensor<string, []>("op_33"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> var_34 = add(x = audio_length, y = var_33)[name = tensor<string, []>("op_34")];
+            tensor<int32, []> var_35 = const()[name = tensor<string, []>("op_35"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> var_36 = sub(x = var_34, y = var_35)[name = tensor<string, []>("op_36")];
+            tensor<int32, [1]> floor_div_0 = floor_div(x = var_36, y = var_10)[name = tensor<string, []>("floor_div_0")];
+            tensor<bool, [1]> var_39 = equal(x = audio_length, y = var_12)[name = tensor<string, []>("op_39")];
+            tensor<int32, [1]> var_40 = const()[name = tensor<string, []>("op_40"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> mel_length = select(a = var_40, b = floor_div_0, cond = var_39)[name = tensor<string, []>("seq_len")];
+            tensor<string, []> audio_to_fp16_dtype_0 = const()[name = tensor<string, []>("audio_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, ?]> audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = tensor<string, []>("cast_14")];
+            tensor<int32, [2]> var_42_shape_cast_fp16 = shape(x = audio_to_fp16)[name = tensor<string, []>("op_42_shape_cast_fp16")];
+            tensor<int32, []> gather_0_axis_0 = const()[name = tensor<string, []>("gather_0_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_0_batch_dims_0 = const()[name = tensor<string, []>("gather_0_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_0_validate_indices_0 = const()[name = tensor<string, []>("gather_0_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> var_42_shape_cast_fp16_to_int16_dtype_0 = const()[name = tensor<string, []>("op_42_shape_cast_fp16_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<uint16, []> select_0_to_uint16 = const()[name = tensor<string, []>("select_0_to_uint16"), val = tensor<uint16, []>(1)];
+            tensor<int16, [2]> var_42_shape_cast_fp16_to_int16 = cast(dtype = var_42_shape_cast_fp16_to_int16_dtype_0, x = var_42_shape_cast_fp16)[name = tensor<string, []>("cast_13")];
+            tensor<int16, []> gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_42_shape_cast_fp16_to_int16)[name = tensor<string, []>("gather_0_cast_uint16")];
+            tensor<string, []> gather_0_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_0_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, []> const_0 = const()[name = tensor<string, []>("const_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> const_1 = const()[name = tensor<string, []>("const_1"), val = tensor<int32, []>(1)];
+            tensor<int32, []> gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = tensor<string, []>("cast_12")];
+            tensor<int32, [?]> var_43 = range_1d(end = gather_0_cast_uint16_to_int32, start = const_0, step = const_1)[name = tensor<string, []>("op_43")];
+            tensor<int32, [1]> var_44_axes_0 = const()[name = tensor<string, []>("op_44_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = tensor<string, []>("op_44")];
+            tensor<int32, [1]> var_45_axes_0 = const()[name = tensor<string, []>("op_45_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_45 = expand_dims(axes = var_45_axes_0, x = audio_length)[name = tensor<string, []>("op_45")];
+            tensor<bool, [1, ?]> timemask = less(x = var_44, y = var_45)[name = tensor<string, []>("timemask")];
+            tensor<int32, [2]> var_48_begin_0 = const()[name = tensor<string, []>("op_48_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_48_end_0 = const()[name = tensor<string, []>("op_48_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> var_48_end_mask_0 = const()[name = tensor<string, []>("op_48_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<bool, [2]> var_48_squeeze_mask_0 = const()[name = tensor<string, []>("op_48_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
+            tensor<fp16, [1]> var_48_cast_fp16 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, squeeze_mask = var_48_squeeze_mask_0, x = audio_to_fp16)[name = tensor<string, []>("op_48_cast_fp16")];
+            tensor<int32, [1]> var_49_axes_0 = const()[name = tensor<string, []>("op_49_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1]> var_49_cast_fp16 = expand_dims(axes = var_49_axes_0, x = var_48_cast_fp16)[name = tensor<string, []>("op_49_cast_fp16")];
+            tensor<int32, [2]> var_51_begin_0 = const()[name = tensor<string, []>("op_51_begin_0"), val = tensor<int32, [2]>([0, 1])];
+            tensor<int32, [2]> var_51_end_0 = const()[name = tensor<string, []>("op_51_end_0"), val = tensor<int32, [2]>([1, 0])];
+            tensor<bool, [2]> var_51_end_mask_0 = const()[name = tensor<string, []>("op_51_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<fp16, [1, ?]> var_51_cast_fp16 = slice_by_index(begin = var_51_begin_0, end = var_51_end_0, end_mask = var_51_end_mask_0, x = audio_to_fp16)[name = tensor<string, []>("op_51_cast_fp16")];
+            tensor<int32, [2]> var_53_begin_0 = const()[name = tensor<string, []>("op_53_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_53_end_0 = const()[name = tensor<string, []>("op_53_end_0"), val = tensor<int32, [2]>([1, -1])];
+            tensor<bool, [2]> var_53_end_mask_0 = const()[name = tensor<string, []>("op_53_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp16, [1, ?]> var_53_cast_fp16 = slice_by_index(begin = var_53_begin_0, end = var_53_end_0, end_mask = var_53_end_mask_0, x = audio_to_fp16)[name = tensor<string, []>("op_53_cast_fp16")];
+            tensor<fp16, []> var_54_to_fp16 = const()[name = tensor<string, []>("op_54_to_fp16"), val = tensor<fp16, []>(0x1.f0cp-1)];
+            tensor<fp16, [1, ?]> var_55_cast_fp16 = mul(x = var_53_cast_fp16, y = var_54_to_fp16)[name = tensor<string, []>("op_55_cast_fp16")];
+            tensor<fp16, [1, ?]> var_56_cast_fp16 = sub(x = var_51_cast_fp16, y = var_55_cast_fp16)[name = tensor<string, []>("op_56_cast_fp16")];
+            tensor<bool, []> x_3_interleave_0 = const()[name = tensor<string, []>("x_3_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, ?]> x_3_cast_fp16 = concat(axis = var_9, interleave = x_3_interleave_0, values = (var_49_cast_fp16, var_56_cast_fp16))[name = tensor<string, []>("x_3_cast_fp16")];
+            tensor<bool, [1, ?]> var_59 = logical_not(x = timemask)[name = tensor<string, []>("op_59")];
+            tensor<fp16, []> var_16_to_fp16 = const()[name = tensor<string, []>("op_16_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, ?]> input_1_cast_fp16 = select(a = var_16_to_fp16, b = x_3_cast_fp16, cond = var_59)[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<int32, [3]> concat_1x = const()[name = tensor<string, []>("concat_1x"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, ?]> input_3_cast_fp16 = reshape(shape = concat_1x, x = input_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<int32, [6]> input_5_pad_0 = const()[name = tensor<string, []>("input_5_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 256, 256])];
+            tensor<string, []> input_5_mode_0 = const()[name = tensor<string, []>("input_5_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_3_to_fp16 = const()[name = tensor<string, []>("const_3_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 1, ?]> input_5_cast_fp16 = pad(constant_val = const_3_to_fp16, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<int32, [2]> concat_2x = const()[name = tensor<string, []>("concat_2x"), val = tensor<int32, [2]>([1, -1])];
+            tensor<fp16, [1, ?]> input_cast_fp16 = reshape(shape = concat_2x, x = input_5_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [1]> expand_dims_3 = const()[name = tensor<string, []>("expand_dims_3"), val = tensor<int32, [1]>([160])];
+            tensor<int32, [1]> expand_dims_4_axes_0 = const()[name = tensor<string, []>("expand_dims_4_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, ?]> expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = input_cast_fp16)[name = tensor<string, []>("expand_dims_4_cast_fp16")];
+            tensor<string, []> conv_0_pad_type_0 = const()[name = tensor<string, []>("conv_0_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> conv_0_pad_0 = const()[name = tensor<string, []>("conv_0_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_0_dilations_0 = const()[name = tensor<string, []>("conv_0_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> conv_0_groups_0 = const()[name = tensor<string, []>("conv_0_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [257, 1, 512]> expand_dims_1_to_fp16 = const()[name = tensor<string, []>("expand_dims_1_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [1, 257, ?]> conv_0_cast_fp16 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1_to_fp16, x = expand_dims_4_cast_fp16)[name = tensor<string, []>("conv_0_cast_fp16")];
+            tensor<string, []> conv_1_pad_type_0 = const()[name = tensor<string, []>("conv_1_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> conv_1_pad_0 = const()[name = tensor<string, []>("conv_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_1_dilations_0 = const()[name = tensor<string, []>("conv_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> conv_1_groups_0 = const()[name = tensor<string, []>("conv_1_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [257, 1, 512]> expand_dims_2_to_fp16 = const()[name = tensor<string, []>("expand_dims_2_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263296)))];
+            tensor<fp16, [1, 257, ?]> conv_1_cast_fp16 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2_to_fp16, x = expand_dims_4_cast_fp16)[name = tensor<string, []>("conv_1_cast_fp16")];
+            tensor<int32, []> stack_0_axis_0 = const()[name = tensor<string, []>("stack_0_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 257, ?, 2]> stack_0_cast_fp16 = stack(axis = stack_0_axis_0, values = (conv_0_cast_fp16, conv_1_cast_fp16))[name = tensor<string, []>("stack_0_cast_fp16")];
+            tensor<fp16, []> var_19_promoted_to_fp16 = const()[name = tensor<string, []>("op_19_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 257, ?, 2]> var_74_cast_fp16 = pow(x = stack_0_cast_fp16, y = var_19_promoted_to_fp16)[name = tensor<string, []>("op_74_cast_fp16")];
+            tensor<int32, [1]> var_76_axes_0 = const()[name = tensor<string, []>("op_76_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> var_76_keep_dims_0 = const()[name = tensor<string, []>("op_76_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 257, ?]> var_76_cast_fp16 = reduce_sum(axes = var_76_axes_0, keep_dims = var_76_keep_dims_0, x = var_74_cast_fp16)[name = tensor<string, []>("op_76_cast_fp16")];
+            tensor<fp16, [1, 257, ?]> x_11_cast_fp16 = identity(x = var_76_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
+            tensor<bool, []> x_13_transpose_x_0 = const()[name = tensor<string, []>("x_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_13_transpose_y_0 = const()[name = tensor<string, []>("x_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 128, 257]> const_4_to_fp16 = const()[name = tensor<string, []>("const_4_to_fp16"), val = tensor<fp16, [1, 128, 257]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526528)))];
+            tensor<fp16, [1, 128, ?]> x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = const_4_to_fp16, y = x_11_cast_fp16)[name = tensor<string, []>("x_13_cast_fp16")];
+            tensor<fp16, []> var_83_to_fp16 = const()[name = tensor<string, []>("op_83_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [1, 128, ?]> var_84_cast_fp16 = add(x = x_13_cast_fp16, y = var_83_to_fp16)[name = tensor<string, []>("op_84_cast_fp16")];
+            tensor<fp32, []> x_epsilon_0 = const()[name = tensor<string, []>("x_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
+            tensor<fp16, [1, 128, ?]> x_cast_fp16 = log(epsilon = x_epsilon_0, x = var_84_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
+            tensor<int32, [3]> var_86_shape_cast_fp16 = shape(x = x_cast_fp16)[name = tensor<string, []>("op_86_shape_cast_fp16")];
+            tensor<int32, []> gather_5_axis_0 = const()[name = tensor<string, []>("gather_5_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_5_batch_dims_0 = const()[name = tensor<string, []>("gather_5_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_5_validate_indices_0 = const()[name = tensor<string, []>("gather_5_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> var_86_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_86_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
+            tensor<uint16, []> select_5_to_uint16 = const()[name = tensor<string, []>("select_5_to_uint16"), val = tensor<uint16, []>(2)];
+            tensor<uint16, [3]> var_86_shape_cast_fp16_to_uint16 = cast(dtype = var_86_shape_cast_fp16_to_uint16_dtype_0, x = var_86_shape_cast_fp16)[name = tensor<string, []>("cast_11")];
+            tensor<uint16, []> gather_5_cast_uint16 = gather(axis = gather_5_axis_0, batch_dims = gather_5_batch_dims_0, indices = select_5_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_86_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_5_cast_uint16")];
+            tensor<string, []> gather_5_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_5_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, []> const_5 = const()[name = tensor<string, []>("const_5"), val = tensor<int32, []>(0)];
+            tensor<int32, []> const_6 = const()[name = tensor<string, []>("const_6"), val = tensor<int32, []>(1)];
+            tensor<int32, []> gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16)[name = tensor<string, []>("cast_10")];
+            tensor<int32, [?]> mask_1 = range_1d(end = gather_5_cast_uint16_to_int32, start = const_5, step = const_6)[name = tensor<string, []>("mask_1")];
+            tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = tensor<string, []>("expand_dims_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = tensor<string, []>("expand_dims_0")];
+            tensor<int32, [1]> var_91_axes_0 = const()[name = tensor<string, []>("op_91_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_91 = expand_dims(axes = var_91_axes_0, x = mel_length)[name = tensor<string, []>("op_91")];
+            tensor<bool, [1, ?]> mask = greater_equal(x = expand_dims_0, y = var_91)[name = tensor<string, []>("mask")];
+            tensor<int32, [1]> var_93_axes_0 = const()[name = tensor<string, []>("op_93_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, ?]> var_93 = expand_dims(axes = var_93_axes_0, x = mask)[name = tensor<string, []>("op_93")];
+            tensor<fp16, [1, 128, ?]> processed_signal_cast_fp16 = select(a = var_16_to_fp16, b = x_cast_fp16, cond = var_93)[name = tensor<string, []>("processed_signal_cast_fp16")];
+            tensor<string, []> processed_signal_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("processed_signal_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 128, ?]> mel = cast(dtype = processed_signal_cast_fp16_to_fp32_dtype_0, x = processed_signal_cast_fp16)[name = tensor<string, []>("cast_9")];
+        } -> (mel, mel_length);
+}
\ No newline at end of file
diff --git a/latin/560ms/preprocessor.mlmodelc/weights/weight.bin b/latin/560ms/preprocessor.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411
--- /dev/null
+++ b/latin/560ms/preprocessor.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3
+size 592384
diff --git a/multilingual/1120ms/decoder.mlmodelc/analytics/coremldata.bin b/multilingual/1120ms/decoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..6d1da7489b64e3e6513ae9a70aa2c9cad18c649c
--- /dev/null
+++ b/multilingual/1120ms/decoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fdb14a08e42b4806a2d1505501586be71e4f04ca9256c719544fd7ed6937e509
+size 243
diff --git a/multilingual/1120ms/decoder.mlmodelc/coremldata.bin b/multilingual/1120ms/decoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..23e420e64e286eca983ee4dd26bb589fb3e2fce7
--- /dev/null
+++ b/multilingual/1120ms/decoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d13f9f434a2485bf9bac16e353f0c2ea850ec5a98e44effc33bfe17da8559d5
+size 433
diff --git a/multilingual/1120ms/decoder_joint.mlmodelc/analytics/coremldata.bin b/multilingual/1120ms/decoder_joint.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..12fefe914b844d2fe6a8f7f8bfbb549dd8b45e8c
--- /dev/null
+++ b/multilingual/1120ms/decoder_joint.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a8e98a54ed1f16c3d5125816a002b991e167d620beb8fcc557f26d9a1c092f8
+size 243
diff --git a/multilingual/1120ms/decoder_joint.mlmodelc/coremldata.bin b/multilingual/1120ms/decoder_joint.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..de7ef2da51322d665f104c9d9810c0eaf192a82a
--- /dev/null
+++ b/multilingual/1120ms/decoder_joint.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0cf3b4fe7bc448f822058c9de3e9fdc72f6277fb72f77836535c829ae3058d0a
+size 454
diff --git a/multilingual/1120ms/encoder.mlmodelc/analytics/coremldata.bin b/multilingual/1120ms/encoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..851730eed2a8b88a73a98f12c18dff5e813b2c75
--- /dev/null
+++ b/multilingual/1120ms/encoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5488ed81fa1e370dd5411218d9b0e0562d01ca33c45b9ea1935bab6c8e8dc943
+size 243
diff --git a/multilingual/1120ms/encoder.mlmodelc/coremldata.bin b/multilingual/1120ms/encoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7975a89ee6d82ebd33d5e72fd9b005a635e4a8c4
--- /dev/null
+++ b/multilingual/1120ms/encoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93d7f1b6412b965b379f3359e38f9c4330099b5b41f9d30b85bf7286b098a78d
+size 572
diff --git a/multilingual/1120ms/encoder.mlmodelc/weights/weight.bin b/multilingual/1120ms/encoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..22d4a9d6056019a89615f7756e310887fb6faff0
--- /dev/null
+++ b/multilingual/1120ms/encoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3dbd37659c7ef51d2dc691afac42fbdd1f04998fa62816ff0f2eebec5cc3d01
+size 564646848
diff --git a/multilingual/1120ms/joint.mlmodelc/analytics/coremldata.bin b/multilingual/1120ms/joint.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..85265398a9d4a252abafafb61d0364dd1682f4f7
--- /dev/null
+++ b/multilingual/1120ms/joint.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a1a90a7d5f8b86f564a42ab45b42a43eb0bbce6682176916bd94591b28cca447
+size 243
diff --git a/multilingual/1120ms/joint.mlmodelc/coremldata.bin b/multilingual/1120ms/joint.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..d3d74cc4c6138903cfe788c0a5136aafa3fa3a4c
--- /dev/null
+++ b/multilingual/1120ms/joint.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:15415f76583ac18c78be1a4192aad2023432acd9095baf4862b86e5122cf0958
+size 341
diff --git a/multilingual/1120ms/preprocessor.mlmodelc/analytics/coremldata.bin b/multilingual/1120ms/preprocessor.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..c4ce5f75c590e87f1ec0e85f34e78a1ba9e8f4c8
--- /dev/null
+++ b/multilingual/1120ms/preprocessor.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:265c27d899f10ef55d65912885b59e3420ba752d15009a925a1167db666792da
+size 243
diff --git a/multilingual/1120ms/preprocessor.mlmodelc/coremldata.bin b/multilingual/1120ms/preprocessor.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..1bca66aa9accbf6cd9a24ab661c31f001c08693a
--- /dev/null
+++ b/multilingual/1120ms/preprocessor.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d1aa8c8e7e283e4944af4b0b701db760ed99ef14919d3f989c599b9f63335a2
+size 371
diff --git a/multilingual/1120ms/preprocessor.mlmodelc/weights/weight.bin b/multilingual/1120ms/preprocessor.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411
--- /dev/null
+++ b/multilingual/1120ms/preprocessor.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3
+size 592384
diff --git a/multilingual/2240ms/decoder.mlmodelc/coremldata.bin b/multilingual/2240ms/decoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..f068f7767e3224f5a5479dc20d7cf54384f92bfb
--- /dev/null
+++ b/multilingual/2240ms/decoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a89047b6f74ee3d0a74c72f8c8e5016d76d9487d030f0661fe80220b441a6fd
+size 433
diff --git a/multilingual/2240ms/decoder_joint.mlmodelc/analytics/coremldata.bin b/multilingual/2240ms/decoder_joint.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..12fefe914b844d2fe6a8f7f8bfbb549dd8b45e8c
--- /dev/null
+++ b/multilingual/2240ms/decoder_joint.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a8e98a54ed1f16c3d5125816a002b991e167d620beb8fcc557f26d9a1c092f8
+size 243
diff --git a/multilingual/2240ms/decoder_joint.mlmodelc/coremldata.bin b/multilingual/2240ms/decoder_joint.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..237f65d18903f8d8bdab985a165663ec6f0820c9
--- /dev/null
+++ b/multilingual/2240ms/decoder_joint.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6404b542fb5d5faa79648fc96a79a0b981cbd087df0378adefd1f45ae56dd86e
+size 454
diff --git a/multilingual/2240ms/encoder.mlmodelc/coremldata.bin b/multilingual/2240ms/encoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..09fdeab19507bc02a9054ff0155e8c445379c33c
--- /dev/null
+++ b/multilingual/2240ms/encoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5471a4edf55ce02dab51e5d01c86b419a7659bcdd6aac216ba6f0b795eae8bf
+size 573
diff --git a/multilingual/2240ms/joint.mlmodelc/coremldata.bin b/multilingual/2240ms/joint.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..20901d7e565b36627b81d00b7626e8a6617789c9
--- /dev/null
+++ b/multilingual/2240ms/joint.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f750980da8ea3397d860f69e0755d3adf61ee763d7f189405d50d4e2d9f8ca0
+size 341
diff --git a/multilingual/2240ms/preprocessor.mlmodelc/analytics/coremldata.bin b/multilingual/2240ms/preprocessor.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7d40c48169474732ff5d227c86c093a242482865
--- /dev/null
+++ b/multilingual/2240ms/preprocessor.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e918fd75105ef01a971d29b5ec28f531467b42dd60978c29afb1914c4af838af
+size 243
diff --git a/multilingual/2240ms/preprocessor.mlmodelc/coremldata.bin b/multilingual/2240ms/preprocessor.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..1bca66aa9accbf6cd9a24ab661c31f001c08693a
--- /dev/null
+++ b/multilingual/2240ms/preprocessor.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d1aa8c8e7e283e4944af4b0b701db760ed99ef14919d3f989c599b9f63335a2
+size 371
diff --git a/multilingual/2240ms/preprocessor.mlmodelc/weights/weight.bin b/multilingual/2240ms/preprocessor.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411
--- /dev/null
+++ b/multilingual/2240ms/preprocessor.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3
+size 592384
diff --git a/multilingual/4480ms/decoder.mlmodelc/weights/weight.bin b/multilingual/4480ms/decoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..d337a15c3738ff2fc286d4329f29e91dac32ecee
--- /dev/null
+++ b/multilingual/4480ms/decoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dcdeccd4ccf46e2675224f9f030d46c1a89e2bda4abb316e901e1a21f1597f8f
+size 29870592
diff --git a/multilingual/4480ms/decoder_joint.mlmodelc/weights/weight.bin b/multilingual/4480ms/decoder_joint.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..5c8c062fd9461e5e579f621cdccaedaa1a57a2d9
--- /dev/null
+++ b/multilingual/4480ms/decoder_joint.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:01f21eb747fbc53bd0ed7efebea1bf0aa655ebf2816f21d0bb6554c9b7fcfc0b
+size 48782272
diff --git a/multilingual/560ms/decoder_joint.mlmodelc/weights/weight.bin b/multilingual/560ms/decoder_joint.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..5c8c062fd9461e5e579f621cdccaedaa1a57a2d9
--- /dev/null
+++ b/multilingual/560ms/decoder_joint.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:01f21eb747fbc53bd0ed7efebea1bf0aa655ebf2816f21d0bb6554c9b7fcfc0b
+size 48782272